# Changelog All notable changes to this project will be documented in this file. The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html). ## [Unreleased] ### Fixed - `@apicity/cost`: Guard OpenAI Responses `prompt_cache_options.ttl`, per-breakpoint `prompt_cache_breakpoint.ttl`, mixed-retention requests, and response `cache_write_tokens_details.ttl_12h_tokens` so 12-hour cache writes fail closed instead of pricing at the 30-minute write rate (ac-k0mwj9). ### Changed - `@apicity/cost`: Keep KIE Gemini 4 Argon (`gemini-4-argon`, `gemini-4-argon-openai`) permanently without `cacheRead: 0.04` (non-breaking) after 2026-11-22 paired cache-hit probes on both OpenAI-compat and native Gemini routes confirmed neither endpoint reports cached token counts or discounts `credits_consumed` below the `$0.80/1M` input rate (`tests/fixtures/kie-pricing-evidence/gemini-4-argon-cache-probe-2026-11-22.json`, ac-o8ihgj). ### Added - Support `omni-flash-1.2` (Gemini Omni 1.2 Flash) on `@apicity/googleflow` (`POST v1.videos`) with 3–15 s durations (3–10 s in `I2V-FL`), `360p`/`720p` resolution, up to 10 `referenceImage_*`/`character_*` slots, 15 s (360-frame) V2V trim window, and `returnSilentVideos`; `@apicity/cost` prices it across all four Google AI plans (`3`/`1` credits/s for 1–10 s and `6`/`2` credits/s for 11–15 s at `720p`/`360p`, V2V billed by edited slice length), shares slug `geminiomni1p2f`, and maps `googleflow` on the `Gemini Omni Flash 1.2` frontier `v2v` (#2), `r2v` (#4), and `t2v` (#7) rows (ac-5kldja). - Support `happyhorse-1.0-video-edit` on `@apicity/alibaba` (`api.v1.services.aigc.videoGeneration.videoSynthesis`) with `@apicity/cost` International pricing (`$0.14`/`$0.24` per s for `720P`/`1080P`, billing both input and output seconds), model slug `hh`, and frontier `v2v` `HappyHorse-1.0` mapping (ac-s5lhzh). - Support `wan3.0-video` and `wan3.0-video-prime` on `@apicity/alibaba` (`api.v1.services.aigc.videoGeneration.videoSynthesis`) with `@apicity/cost` International pricing (`$0.05`/`$0.10`/`$0.20` per s and `$0.068`/`$0.14`/`$0.28` per s for `480P`/`720P`/`1080P`), model slugs (`wan3p0`, `wan3p0p`), and frontier `t2v`/`r2v`/`v2v` `Wan 3.0` mappings. ### Added - `@apicity/google` adds the project-scoped Vertex AI global generateContent route `v1.projects.locations.global.publishers.google.models.generateContent` (`POST https://aiplatform.googleapis.com/v1/projects/{project}/locations/global/publishers/google/models/{model}:generateContent`) for `gemini-nano-banana-2.1` (and other global-only Vertex models whose express-mode `europe-west1` endpoint returns 404). Adds typed `GoogleGenerationConfigSchema` / `GoogleImageConfigSchema` / `GoogleThinkingConfigSchema` schemas (all 14 Nano Banana 2.1 aspect ratios, `1K` / `2K` / `4K` image sizes, `responseModalities`, `personGeneration`, `prominentPeople`, `imageOutputOptions`, `thinkingLevel`) and enforces the 14-image input limit pre-fetch for `gemini-nano-banana-2.1`. `@apicity/cost` adds `gemini-nano-banana-2.1` pricing (`$1.50/1M` input tokens, `$0.15/1M` cached input, `$7.50/1M` text/reasoning output, `$30/1M` image output), modality-aware `extractActual` extraction of `promptTokensDetails`/`candidatesTokensDetails` image tokens, slug `nb21` / display `"Nano Banana 2.1"`, and endpoint billing metadata; frontier t2i (#8) and i2i (#10) map Google `v1.projects.locations.global.publishers.google.models.generateContent`. Recorded a 1K `1:1` generation (`0.0393615` USD: 11 text-in, 766 thoughts, 1,120 image-out tokens) plus a bogus-model 404 read (ac-hvvd6d). - OpenAI `gpt-image-2.6-sunburst` and `gpt-image-2.6-flare` on `v1.images.generations` and `v1.images.edits`: `@apicity/openai` adds `OpenAiImageComplexitySchema` (`low`/`medium`/`high`), `frame_count`/`frame_duration_ms`, `gif` (`svg` on generations), `safety_identifier`, and `markup`/`moderation` edit forwarding with a recorded `gpt-image-2.6-flare` `low` `1024x1024` generation (`15` text-in + `824` image-out tokens, `0.024795` USD) and a bogus-model 404 read. `@apicity/cost` prices both models at `$30/1M` image output tokens via OpenAI's documented 2.6 token calculator multipliers (`x4.2` low, `x3.3` medium), shares slugs `gi2p6sb`/`gi2p6fl`, and maps OpenAI on all four `GPT Image 2.6` frontier t2i/i2i rows (ac-xbnboo). - KIE `kling-4.0/reference-to-video` and `kling-4.0/keyframe-to-video` accept `resolution: "4k"`, `dynamic_range` (`"sdr"` by default, or `"hdr"`, supported at `"1080p"` and `"4k"` only; `"hdr"` at `"720p"` is rejected, and `Kling40DynamicRangeSchema` / `Kling40DynamicRange` are exported), and aspect ratios `"4:3"` and `"3:4"`; `kling-4.0/flash-reference-to-video` and `kling-4.0/flash-keyframe-to-video` accept `"4:3"` and `"3:4"` too (matching the KIE docs read 2026-11-21). `@apicity/cost` prices `4k` (99 credits = 0.495 USD/s; 59 credits = 0.295 USD/s with video input) and the `hdr` output-seconds add-on (+8.5 credits = +0.0425 USD/s at `1080p`, +14 credits = +0.07 USD/s at `4k`; reference video input duration adds no HDR charge, and `720p` + `hdr` fails closed), reconciling all Kling 4.0 feed drift (ac-91jg63). - `@apicity/cli`, `@apicity/cost`: On a host with a pay-gate secret, the operator can mint a time-boxed provider grant (`apicity-paygate grant mint --secret-file --provider fal --ttl 4h`; `grant list` and `grant revoke` manage it). Until it expires (`1h` default, `24h` max), every endpoint of that provider that the CLI gates itself (`gate: "cli"`, 584 endpoints across 17 billed providers) runs without an OTP and appends an audit line to `~/.config/apicity/paygate-grants/uses.jsonl` (`0600`). The 27 endpoints of `PAID_ENDPOINTS` (`kie` 21, `xai` 6; `gate: "sdk"`) take no grant and still need a per-request OTP. `@apicity/cost` exports `mintGrant`, `verifyGrant`, `isGrantProviderName`, `grantFileName`, `grantDirectory`, and grant constants; `@apicity/cli` exports `GrantRefusal`, `GrantRefusalCode`, and `GRANT_USES_FILE` (ac-upjl0y). - `@apicity/wavespeed` adds GPT Image 2.5 Flare edit at `api.v3.openai.gptImage2p5Flare.edit` (`POST https://api.wavespeed.ai/api/v3/openai/gpt-image-2.5-flare/edit`), #5 on the 2026-11-19 image-editing board, as a leaf beside the `textToImage` endpoint the `gptImage2p5Flare` object already declares. WaveSpeed's catalog, read 2026-11-20, gives it the Sunburst edit's request schema, which it reuses. `@apicity/cost` prices it like the Sunburst edit, from free quotes read the same day that match it cell for cell: the GPT Image 2.5 text-to-image rate for the quality and resolution, plus 0.015 USD per input image, plus 0.015 USD once for a mask. The frontier registry adds this WaveSpeed leaf beside fal, KIE, MachGen and OpenAI, so the image-editing board stays 10/10 mapped. A recorded low-quality 1k edit of one cat photo billed 0.025 USD (ac-0dfgz3). - `@apicity/wavespeed` adds GPT Image 2.5 Sunburst edit at `api.v3.openai.gptImage2p5Sunburst.edit` (`POST https://api.wavespeed.ai/api/v3/openai/gpt-image-2.5-sunburst/edit`), #4 on the 2026-11-19 image-editing board, as a leaf beside the `textToImage` endpoint the `gptImage2p5Sunburst` object already declares. The schema follows WaveSpeed's catalog, re-read 2026-11-20 and unchanged since the survey: a prompt, 1-16 input images, an optional mask for the first image, an optional aspect ratio (it follows the input image when omitted), resolution `1k` (default), `2k` or `4k`, quality `low` to `max` (default `medium`), and png, jpeg or webp (default png). Unlike the GPT Image 2.6 edit it takes no markup image, complexity, gif or frame controls. `@apicity/cost` prices it from free quotes read the same day: the GPT Image 2.5 text-to-image rate for the quality and resolution, plus 0.015 USD per input image, plus 0.015 USD once for a mask. The frontier registry adds this WaveSpeed leaf beside fal, KIE, MachGen and OpenAI, so the image-editing board stays 10/10 mapped. A recorded low-quality 1k edit of one cat photo billed 0.025 USD (ac-nhmyvr). - `@apicity/wavespeed` adds Kling Image 4.0 text-to-image at `api.v3.kwaivgi.klingImageV4p0.textToImage` (`POST https://api.wavespeed.ai/api/v3/kwaivgi/kling-image-v4.0/text-to-image`), #7 on the 2026-11-19 text-to-image board, as a leaf on the `klingImageV4p0` object the edit endpoint already declares. The schema follows WaveSpeed's catalog and model page, both re-read 2026-11-20 and unchanged since the survey: a prompt, an optional aspect ratio (default `1:1`), resolution `2k` (default), `4k` or `8k`, 1-9 output images (default 1), and png, jpeg or webp (default png). There is no images field. `@apicity/cost` prices it from the page and from free quotes read the same day: 0.063 / 0.094 / 0.157 USD per image at 2k / 4k / 8k, with no input surcharge. The frontier registry adds this WaveSpeed leaf beside fal and KIE, so the text-to-image board stays 9/10 mapped. A recorded prompt-only 2k image billed 0.063 USD (ac-v4amf8). - `@apicity/wavespeed` adds Kling Image 4.0 edit at `api.v3.kwaivgi.klingImageV4p0.edit` (`POST https://api.wavespeed.ai/api/v3/kwaivgi/kling-image-v4.0/edit`), #6 on the 2026-11-19 image-editing board, as a `klingImageV4p0` object under the existing `kwaivgi` namespace. The schema follows WaveSpeed's catalog, re-read 2026-11-20 and unchanged since the survey: a prompt, 1-10 reference images, an optional aspect ratio, resolution `2k` (default), `4k` or `8k`, 1-9 output images (default 1), and png, jpeg or webp (default png). `@apicity/cost` prices it from free quotes read the same day: 0.063 / 0.094 / 0.157 USD per output image at 2k / 4k / 8k, plus 0.003 USD per input image. The frontier registry adds this WaveSpeed leaf beside fal and KIE, so the image-editing board stays 10/10 mapped. A recorded 2k edit of one cat photo billed 0.066 USD (ac-bfkzn9). - `@apicity/fal` adds Seedream 5.5 edit at `run.bytedance.seedream.v5p5.edit` (`POST https://fal.run/bytedance/seedream/v5.5/edit`), #8 on the 2026-11-19 image-editing board, as a `v5p5` object under the existing `bytedance.seedream` namespace. The schema follows fal's queue OpenAPI, re-read 2026-11-20 and unchanged since the survey: a prompt, one or more reference images (the last 10 are used), `image_size` presets or custom dimensions (default `auto_2K`), 1-6 outputs, JPEG or PNG, watermark, sync mode, safety checker, and background `auto`/`opaque`/`transparent`. `@apicity/cost` prices it from the card: $0.03 per output at 1K or 1.5K, $0.0405 at 2K, plus $0.012 per input image per output. A recorded `auto_1K` edit of the owned cat photo returned one 864x1152 PNG and billed 1.4 units (0.042 USD), matching the card. The frontier registry adds this fal leaf beside the existing KIE, WaveSpeed, and MachGen rows, so the image-editing board stays 10/10 mapped. A free queue read of an unknown request id on `bytedance/seedream` is recorded too (ac-dghypo). - `@apicity/fal` adds Kling Image 4.0 text-to-image at `run.kling.v4.textToImage` (`POST https://fal.run/kling/v4/text-to-image`), #7 on the 2026-11-19 text-to-image board, as a leaf on the existing `kling.v4` object. The schema follows fal's queue OpenAPI, re-read 2026-11-20 and unchanged since the survey: a prompt, single or series output, 2K/4K/8K, and an aspect ratio. `@apicity/cost` leaves it dynamic. The pricing API quotes $0.063 per unit and does not define the unit, and a recorded default 2K image billed 1 unit (0.063 USD) for a 2048x2048 PNG. The frontier registry adds this fal leaf beside the existing KIE row, so the text-to-image board stays mapped. A free queue read of an unknown request id on `kling/v4` is recorded too (ac-5xayh8). - `@apicity/fal` adds Kling Image 4.0 image-to-image at `run.kling.v4.imageToImage` (`POST https://fal.run/kling/v4/image-to-image`), #6 on the 2026-11-19 image-editing board, as a leaf on the existing `kling.v4` object. The schema follows fal's queue OpenAPI, re-read 2026-11-20 and unchanged since the survey: a prompt, 1-10 reference images, up to 7 elements, single or series output, 2K/4K/8K, and an aspect ratio. `@apicity/cost` leaves it dynamic. The pricing API quotes $0.063 per unit and does not define the unit, and a recorded default 2K edit of the owned cat photo billed 1.0476 units (0.0659988 USD). The frontier registry adds this fal leaf beside the existing KIE row, so the image-editing board stays 10/10 mapped. A free queue read of an unknown request id on `kling/v4` is recorded too (ac-xjxelx). - `@apicity/wavespeed` adds SkyReels V4 image-to-video at `api.v3.skyworkAi.skyreelsV4.imageToVideo` (`POST https://api.wavespeed.ai/api/v3/skywork-ai/skyreels-v4/image-to-video`), #5 on the 2026-11-19 image-to-video board, as a leaf beside text-to-video: a prompt and a required first-frame `image`, an optional `last_image` and up to six middle `images`, 3-15 s, 480p, 720p or 1080p, optional sound, and std or fast mode (fast needs sound off). `@apicity/cost` prices it per second from free quotes read 2026-11-20, the same table as text-to-video: 0.11 / 0.14 / 0.35 USD/s in std and 0.08 / 0.11 / 0.275 USD/s in fast mode at 480p / 720p / 1080p; sound and the extra frames add nothing. The frontier registry maps SkyReels V4 for image-to-video (9/10). A recorded 3 s, 480p, fast clip of the owned cat photo billed 0.24 USD, the quote (ac-3kldvi). - `@apicity/fal` adds MAI-Image-2.7 Flash editing at `run.microsoft.maiImage2p7.flash.edit` (`POST https://fal.run/microsoft/mai-image-2.7/flash/edit`), #7 on the 2026-11-19 image-editing board, as a `flash` object under `microsoft.maiImage2p7`. Its request matches the 2.7 edit field for field. `@apicity/cost` prices it dynamically from fal's Flash token card ($1.75 / $2.50 / $10 per 1M text-in / image-in / image-out tokens), and the frontier registry points MAI-Image-2.7-Flash at it, so every image-editing board row now has an apicity endpoint (10/10). A recorded edit of the owned cat photo with no optional input billed 0.0122 USD (ac-c3yi63). - `@apicity/fal` adds MAI-Image-2.7 text-to-image at `run.microsoft.maiImage2p7.textToImage` (`POST https://fal.run/microsoft/mai-image-2.7/text-to-image`), #4 on the 2026-11-19 text-to-image board, as a leaf on the `microsoft.maiImage2p7` object: a prompt, preset or custom `image_size` (default `square_hd`), `resolution` 1K or 2K, a transparent `background`, and PNG, JPEG or WebP output. `@apicity/cost` prices it dynamically from fal's token card ($5 / $34 per 1M text-in / image-out tokens; a 1024x1024 output is $0.034816 before input tokens), and the frontier registry points MAI-Image-2.7 at it for text-to-image (9/10 mapped). A recorded default-size call billed 0.035 USD (ac-lcuwhi). - `@apicity/fal` adds MAI-Image-2.7 editing at `run.microsoft.maiImage2p7.edit` (`POST https://fal.run/microsoft/mai-image-2.7/edit`), #1 on the 2026-11-19 image-editing board: up to five input images, preset or custom `image_size`, `resolution` 1K or 2K, a transparent `background`, and PNG, JPEG or WebP output. It declares the `microsoft.maiImage2p7` object, which the 2.7 text-to-image and Flash edit leaves will join. `@apicity/cost` prices it dynamically from fal's token card ($5 / $8 / $34 per 1M text-in / image-in / image-out tokens), and the frontier registry points MAI-Image-2.7 at it for image editing (9/10 mapped). A recorded edit of the owned cat photo with no optional input billed 0.0409 USD (ac-uco3qf). - KIE `ltx/2-5-pro-image-to-video`, LTX 2.5 Pro image-to-video by Lightricks, on `api.v1.jobs.createTask`, from its docs read on 2026-11-19 and re-read on 2026-11-20: a required `image_url` (the first frame; it sets the aspect ratio, so there is no `aspect_ratio` field), a 1-5000 character `prompt`, `resolution` `720p` or `1080p` (default), `duration` 6, 8 or 10 seconds (default 10) at any frame rate, `fps` and an optional `seed`. `@apicity/cost` prices it at the Pro rates, 14.4 / 20.4 credits (0.072 / 0.102 USD) per second at 720p / 1080p, from the feed rows and the models API pricingDesc, which agree; the image adds nothing. With it, `check:kie-feed-drift` compares all thirteen LTX 2.5 rows. A recorded 6 s clip at 720p from `cat1.jpg` billed 86.4 credits (0.432 USD) (ac-kevvtj). - KIE `ltx/2-5-pro-text-to-video`, LTX 2.5 Pro text-to-video by Lightricks, on `api.v1.jobs.createTask`, from its docs read on 2026-11-19 and re-read on 2026-11-20: the Fast model's `prompt`, `fps`, `aspect_ratio` and `seed`, with `resolution` `720p` or `1080p` (default) and `duration` 6, 8 or 10 seconds (default 10) at any frame rate. `@apicity/cost` prices it per second at 14.4 / 20.4 credits (0.072 / 0.102 USD) at 720p / 1080p, from the feed rows and the models API pricingDesc, which agree. `check:kie-feed-drift` compares its two rows. A recorded 6 s clip at 720p billed 86.4 credits (0.432 USD) (ac-dwmekr). - KIE `ltx/2-5-fast-audio-to-video`, LTX 2.5 Fast audio-to-video by Lightricks, on `api.v1.jobs.createTask`, from its docs re-read on 2026-11-20: a required `audio_url` (speech, singing or music, at least 2 seconds, kept as the soundtrack), an optional first-frame `image_url` (without it the output is 16:9), a 1-5000 character `prompt`, `resolution` and `fps` as on the other Fast models, an optional `seed`, and an optional `duration` of 2-20 seconds with decimals, cut from the start of the audio. It allows at most 10 seconds at 1440p or at 48 or 50 fps; omitted, the video matches the audio, cut to that limit. `@apicity/cost` prices it at the Fast rates, 10.8 / 15.6 / 22.8 credits (0.054 / 0.078 / 0.114 USD) per second at 720p / 1080p / 1440p, on the output length: `duration` when set, else `costHints.durationSeconds` (the audio's length, cut to the limit), failing closed without either. `check:kie-feed-drift` compares its three rows. A recorded 2 s clip at 720p from `dialog.mp3` billed 21.6 credits (0.108 USD), charged at success (ac-ium3f4). - KIE `ltx/2-5-fast-image-to-video`, LTX 2.5 Fast image-to-video by Lightricks, on `api.v1.jobs.createTask`, from its docs read on 2026-11-19 and re-read on 2026-11-20: a required `image_url` (the first frame, a public JPG, PNG or WEBP URL up to 10 MB, which also sets the aspect ratio, so there is no `aspect_ratio` field), a 1-5000 character `prompt`, and the `resolution`, `fps`, `duration` and `seed` of the text-to-video model, with the same 6, 8 or 10 second limit at 1440p or at 48 or 50 fps. `@apicity/cost` prices it at the Fast rates, 10.8 / 15.6 / 22.8 credits (0.054 / 0.078 / 0.114 USD) per second at 720p / 1080p / 1440p, from the feed rows and the models API pricingDesc, which agree; the image adds nothing. `check:kie-feed-drift` compares its three rows. A recorded 6 s clip at 720p from `cat1.jpg` billed 64.8 credits (0.324 USD) (ac-ey3b9p). - KIE `ltx/2-5-fast-text-to-video`, LTX 2.5 Fast text-to-video by Lightricks, on `api.v1.jobs.createTask`, from its docs read on 2026-11-19 and re-read on 2026-11-20: a 1-5000 character `prompt`, `resolution` `720p`, `1080p` (default) or `1440p`, `fps` 24 (default), 25, 48 or 50, `duration` 6-20 seconds in steps of 2 (default 10), `aspect_ratio` and an optional `seed`. Only 6, 8 or 10 seconds are allowed at 1440p or at 48 or 50 fps, and the schema refuses longer requests there. The video has native audio. `@apicity/cost` prices it per second by resolution at 10.8 / 15.6 / 22.8 credits (0.054 / 0.078 / 0.114 USD), from the feed rows and the models API pricingDesc, which agree; fps does not change the price. `check:kie-feed-drift` compares its three rows, 1440p included. A recorded 6 s clip at 720p billed 64.8 credits (0.324 USD) (ac-ujmlmh). - `@apicity/machgen` `api.v0.generate` accepts `Seedream-5.5` for text-to-image and image-to-image. The id is on this account's model list. An edit takes up to 10 source images, and output is PNG or JPEG. `@apicity/cost` prices the docs page re-read 2026-11-20: 0.008 USD per input image plus 0.020 USD per output through 2,601,123 pixels, or 0.035 USD through 4,624,220. The public pricebook the same day lists one flat 0.020 USD per image and no input surcharge. The frontier image-to-image row maps the existing call. A recorded 1024 edit of one source image billed 0.028 USD (task 0d1e6d3e) (ac-3pjy9k). - The frontier registry maps Hyperflow H3 (text-to-video, image-to-video, and video-to-video) and Pika H3 Flex (image-to-video and video-to-video) through the existing MachGen `api.v0.generate` call, as `MiniMax-H3-Turbo` adapters `hyperflow` and `pika_h3_flex`. Video-to-video is reference-to-video with a video reference. MachGen runs those open weights at 480p and 768p. The board rated Pika's own 1080p service, which apicity does not wrap. The public pricebook, re-read 2026-11-19, still prices `MiniMax-H3-Turbo` at 0.015 USD/s through 2027-01-01T07:59Z, with no adapter surcharge. A recorded 480p 1 s Hyperflow text-to-video billed 0.015 USD (task ad7a09cf) (ac-deyq9x). - `@apicity/machgen` `api.v0.generate` accepts `Gemini-Omni-1.2-Flash` for text-to-video, image-to-video, reference-to-video (reference, edit and extend) and upscale. The id is on this account's model list and absent from the public docs (re-read 2026-11-19). `@apicity/cost` prices the generation rows at 0.023172 / 0.069504 / 0.104256 / 0.208512 USD per second for 360 / 720 / 1080 / 2160. The public pricebook's Generation Credit on those rows is 0. A recorded 360p 3 s text-to-video billed 0.073664 USD against the 0.069516 quote (ac-ssmj0i). ### Removed - **Breaking:** `@apicity/machgen` removes `MiniMax-H3-Omni` from `MACHGEN_VIDEO_MODEL_TASK_TYPES` and `MACHGEN_VIDEO_GRIDS`, and `@apicity/cost` removes `PRICING.machgen["MiniMax-H3-Omni"]`, `MODEL_SLUGS.machgen["MiniMax-H3-Omni"]` (`mmh3o`), and `MODEL_DISPLAY.machgen["MiniMax-H3-Omni"]` (`MiniMax H3 Omni`). MachGen has kept `MiniMax-H3-Omni` private since 2026-11-11 (verified 2026-11-21: absent from both public and keyed pricebooks and `GET /api/v0/models`, HTTP 404 model page, and HTTP 403 `not available for your account` on `POST /api/v0/generate` while an unknown id answers HTTP 400); `modelSlug`, `modelDisplay`, and `MachGenVideoModel` fail on `"MiniMax-H3-Omni"` and cost estimation fails closed as an unpriced model. `MiniMax-H3-Lite` stays absent (also private) (ac-2igfqa). ### Fixed - `@apicity/cli`: Armed OTPs are now single-use across separate CLI runs via an on-disk consumed-token ledger (`$APICITY_PAYGATE_LEDGER_DIR`, else `~/.config/apicity/paygate-consumed/` at mode `0700`, named by a SHA-256 digest of `jti`, pruned strictly past each token's `exp`). Passed to both `gate: "sdk"` (`kie`/`xai` factories via `instantiateProvider`) and `gate: "cli"` (`dispatchThroughGate` / `dispatchGated`); exclusive file creation (`O_EXCL`) rejects concurrent races with `otp-replayed`, and an unwritable ledger fails closed (`paygate-ledger-unwritable`, naming the directory) before dispatch. Unarmed hosts, un-gated calls, bad/mismatched OTPs, and provider grants never touch the ledger; the ledger is per OS user and per host and as tamper-proof as the secret file (ac-ykn4bs). - KIE `grok-imagine-image-2-0/segment-edit` accepts mask index 0. `mask_indexs` takes the `index` values of the segment-map result, which count from 0, and index 0 is usually the main subject. Its docs, read on 2026-11-19, give `minimum: 0` ("0-based"), where the 2026-08-28 page gave 1. A recorded edit that sent `[0]` recolored segment 0, the "white cat". `GrokImagineImage2SegmentEditRequestSchema`, its descriptor and the README now count the indexes from 0, as does the Breaking migration note below (ac-4asuok). - The frontier registry lists Google Flow Nano Banana 2.1 for text-to-image and image-to-image, and lists WaveSpeed Gemini Omni Flash 1.2 and MachGen Seedance 2.5 for video editing (ac-757evu). - `pnpm run audit` passes again. GitHub withdrew advisory GHSA-fc5q-39xc-vjpf (picomatch <=4.0.7) on 2026-11-18, and npm's bulk advisory feed stopped reporting it, so the gate failed on its exception in `scripts/audit-exceptions.json` as matching no current advisory. The entry is removed and the file holds no exceptions; dependency versions and overrides are unchanged (ac-5s46f5). - `@apicity/har` and the test harness remove the temporary `X-Amz-Security-Token` from presigned URLs before a recording is written, and now know the other storage families' signed-URL credentials too. The scrubber removed `X-Amz-Credential` and `X-Amz-Signature` but kept the token. It now removes, in any letter case and compared decoded: AWS `X-Amz-Credential`, `X-Amz-Signature`, `X-Amz-Security-Token`; Alibaba Cloud OSS V1 `OSSAccessKeyId`, `Signature`, `security-token`; OSS V4 `x-oss-credential`, `x-oss-signature`, `x-oss-security-token`; Volcengine TOS `X-Tos-Credential`, `X-Tos-Signature`, `X-Tos-Security-Token`; and Google Cloud Storage `X-Goog-Credential`, `X-Goog-Signature`. It also reaches a URL percent-encoded inside another URL, a URL whose `&` the JSON body escapes as `\u0026`, and the request's `queryString` list, which kept a signed parameter after the URL lost it. The Telegram preview uses the same list. `tests/unit/har-request-secrets.test.ts`, a fast-gate guard, now fails any committed HAR or text fixture that holds a live value for one of them; the ElevenLabs `audio-isolation` recording held two GCS signed URLs (expired 2026-06-29) and is re-scrubbed (ac-rm9wz4). - `@apicity/cost` prices fal MiniMax H3 Max text-to-video, image-to-video and camera-controls at the card's standard 0.05/0.08/0.16 USD/s, and H3 Max Turbo text-to-video and image-to-video at 0.025/0.04/0.08 USD/s, read 2026-11-16 after the November promotion ended. `fal-ai/vidu/q4` keeps its list rates through the promotion ending 2026-11-30 (ac-cukhhs). - `@apicity/machgen` `api.v0.generate` takes 1-3 s on `MiniMax-H3-Turbo` T2V, I2V and R2V. MachGen now accepts 1-15 s there (its `GET /api/v0/models` list and three free 400 probes on 2026-11-11, `tests/fixtures/machgen-evidence/machgen-turbo-durations-2026-11-11.json`), where its 2026-10-25 400s allowed only 4-15 s, so `.schema.safeParse` and `apicity describe` refused a valid short clip. A recorded 1 s 480p T2V billed 15000 micro-USD, the estimate at the 25 % promotion rate (task 14f37c18) (ac-t71px5). - `@apicity/cost` quotes fal MiniMax H3 Max extend, insert, and reference-to-video, including turbo reference-to-video, from the requested seconds and resolution. 3d-to-video and lip-sync quote from `costHints.durationSeconds` plus resolution, and stay dynamic without that hint. The quote is a lower bound: reference tokens past the included 4,096, output overrun, generated references, and a longer lip-sync bill are named in a warning and not added. A payload that does not fix the length and the resolution stays on fal's pricing API. `resolveFalDynamicEstimate` returns that floor instead of fal's historical per-call average, and still calls the pricing API for a positive `unitQuantity`. `unitQuantity: 0` is 0 USD with a `zeroReason` and no pricing-API call (ac-vo9l4a). - `@apicity/cost` WaveSpeed Seedance 2 and 2.5 text-to-video and reference-to-video estimates with reference videos and no `duration` now price the output at the catalog's default 5 s, as WaveSpeed's free quotes do. They followed the references (at least 4 s), the published formula's fallback, which never fires, so a 3 s reference read 1 s short and a 7 s reference 2 s long (ac-hgtlgj). - `@apicity/cost` follows what fal and KIE charge today, from cost learning pass 1 (`docs/cost-learning-pass-1.md`). fal `minimax/h3/t2v/trainer` floors at the card's 100 steps (0.50 USD), so a 1-step run no longer reads 0.005. `minimax/h3-max` and `minimax/h3-max-turbo` move to their November promotion (0.035 and 0.0175 USD/s at 480P, through 2026-11-15). `heygen/heygen-video` moves to 0.02 and 0.06 USD/s now its 50% promotion has ended. `mureka/api/generate/lyrics` is 0.009 USD and its song grid drops by a third. `fal-ai/nano-banana` is 0.0398 USD per image. KIE `pixverse-v6/reference-to-video` takes the shared 0.02-0.092 USD/s ladder. Two hand-written KIE recordings, `kling-v3-turbo-text-to-video` and its 1080p twin, no longer state a charge that was never billed (ac-dcshsu). - `@apicity/har` scrubs credentials from form-encoded bodies. A `@apicity/dropbox` or `@apicity/x` `oauth2.token` call kept its refresh token, and for Dropbox its client secret, in plain text in the capture and in anything `scrubEntry` or `scrubHar` wrote. A field whose name is a known credential now reads `***`, in a form-encoded request or response, in a HAR's `params`, and in a form held as a string in JSON or in call input. `apicity har scrub --check` now fails on a capture that holds one until `apicity har scrub` has rewritten it. No committed recording or real capture held one (ac-to4jdu). - `@apicity/har` no longer stops capturing after one dropped connection or timeout abort. A response or request body that failed partway disabled the capture for the rest of the process and lost the call. It is now saved with `bodyTruncated`, plus `responseCut` for a response, and the call's outcome is `network_error`. A stream keeps its whole lines. Any other body keeps its text only if it parses as a whole JSON document and otherwise becomes a placeholder, because the scrub rules read whole documents. A malformed multipart body no longer disables the capture, and a content-type parameter such as a boundary spelling `xml` no longer makes a body read as text. A caller's abort now still reaches a body after a garbage collection: the capture's copies of the `Request` could be collected once headers arrived, and the abort then went nowhere. - `@apicity/har` `log._apicity.input` no longer holds raw media. Each data URL or base64 field over 1024 characters, and each byte array, becomes a placeholder with its hash, decoded size and media type, which took the 59 largest real captures from 97.3 MB to 1.8 MB. `har promote` refuses such a capture as `media-summary`, and `har report` reads the placeholder as a media value. - `@apicity/har` drops every earlier value of a duplicate JSON key, so `{"api_key":"…","api_key":null}` no longer keeps a credential, for the named provider rules as well as the generic one. Long unpunctuated prose and lists with one word, path or ID to a line are no longer stored as ``: base64 may wrap only on line ends at one width of 60 characters or more. - `@apicity/cost` `extractActual` forwards OpenAI Realtime cached audio and image token counts, from `cached_tokens_details` on either `input_token_details` or `input_tokens_details`, and reads singular `output_token_details` as well as the plural names. A `cached_tokens` total of 0 with no split is both modality counts at 0. A positive cached total without the split stays unpriced, and the derived figure matches `computeEstimate` given the same hints (ac-iikgzi). - `@apicity/cost` `extractActual` reads a response whose body was cut before it closed as unsettled unless its headers stated a figure, so a poll cut that way no longer ends its call on a stale figure (ac-9ce7pu). - `@apicity/machgen` `api.v0.generate` takes up to 16 `src_image_urls` on an I2I edit with `GPT-Image-2`, `GPT-Image-2.5-Sunburst` and `GPT-Image-2.5-Flare`, and up to 14 with `Nano-Banana-2` and `Nano-Banana-2.1`: each model's `maxInputImages` in MachGen's capability contract, read 2026-11-01 (`tests/fixtures/machgen-evidence/machgen-image-caps-2026-11-01.json`). The schema capped `src_image_urls` at 9 for every task, so `apicity describe` and `.schema.safeParse` refused the tenth to sixteenth edit source that MachGen accepts. An edit over its model's cap is refused with a message that names the model and the cap, and the new `MACHGEN_I2I_MAX_IMAGES` export holds the caps. T2I and image ids outside the table are not capped. Video tasks keep their limits: at most 9 images, the reference limit in MachGen's REST docs, 2 on I2V, and each R2V grid's own cap. The schema's refinement now checks the 9, so the JSON Schema that `apicity describe` prints no longer gives `src_image_urls` a `maxItems` (ac-b9b7kw). - `@apicity/xai` `v1.videos.generations` accepts the `21:9` and `5:2` aspect ratios (as does the 1.5 `imageToVideo` leaf), which xAI's REST reference and video overview document for all three video models. A free 422 probe on 2026-11-01 had the request parser list exactly `1:1`, `16:9`, `9:16`, `4:3`, `3:4`, `3:2`, `2:3`, `21:9` and `5:2`. The request also takes the documented `generate_audio`, `output` and `user` fields (ac-zjpet9). - `@apicity/xai` `v1.videos.generations` follows xAI's current reference-to-video doc (read 2026-11-01) on grok-imagine-video-1.5. It takes `last_frame`, its `last_frame_file_id` alias and up to 4 `keyframes`, accepts `image` beside references as the pinned first frame, allows 14 reference images (7 stay on the grok-imagine-video family), and makes `prompt` optional whenever a frame is pinned, on the `imageToVideo` leaf too. grok-imagine-video and other models outside the 1.5 family still refuse an image beside references and refuse the new fields, and a source video still excludes every other mode. `@apicity/cost` prices each pinned frame as one input image, 0.01 USD on the 1.5 models, which xAI's cookbook bills corroborate. Zero spend: no probe (ac-dlb6li). - `@apicity/xai` `v1.videos.generations` refuses `reference_images` and `reference_image_file_ids` on `grok-imagine-video-1.5-lite` and its snapshots, in the schema and before fetch. The lite model page lists text and image to video only, the guide gives lite 1080p for those two modes only, and fal exposes lite as text and image to video only; a lite request that parses can bill, so no probe was sent (ac-p5g552). - `@apicity/xai` `v1.videos.generations` accepts `resolution: "1080p"`, which xAI documents for grok-imagine-video-1.5 text and image to video and for 1.5-lite. The schema listed only `480p` and `720p`, so `apicity describe` and `.schema.safeParse` refused valid 1080p requests. A free 422 probe on 2026-10-31 listed exactly `480p`, `720p` and `1080p`, and `@apicity/cost` already prices 1080p for 1.5 and lite (ac-nw0cw3). - `@apicity/cost` prices every xAI Grok Imagine video resolution and the video input images from recorded bills. `grok-imagine-video` bills 0.05 and 0.07 USD/s at 480p and 720p. `grok-imagine-video-1.5` bills 0.08, 0.14 and 0.25 at 480p, 720p and 1080p, and `grok-imagine-video-1.5-lite` bills 0.02, 0.03 and 0.14. Each input image (`image`, or each reference image) adds 0.002 USD on `grok-imagine-video` and 0.01 on the 1.5 models. Before, 720p on `grok-imagine-video` and lite priced at the 480p rate with a floor warning, and 1080p failed closed; 1080p on `grok-imagine-video`, which renders up to 720p, still fails closed. Eight more recorded 1 s calls (0.762 USD) pin each bill to its estimate (ac-9jsawg). - `@apicity/xai` `v1.videos.generations` refuses `reference_audios` on `grok-imagine-video-1.5-lite`, in the schema and before fetch. The 1.5-family check matched lite by prefix, but xAI documents voice references on grok-imagine-video-1.5 only, and lite takes text and image to video (ac-kfekq8). - `@apicity/openai` `v1.images.edits` accepts the GPT Image 2 and 2.5 options: `quality` `xhigh` and `max` (gpt-image-2.5-sunburst and -flare), and any `WIDTHxHEIGHT` size such as `2048x2048` or `3840x2160` (gpt-image-2 and 2.5). The schema refused both, so `apicity describe` and `.schema.safeParse` rejected documented edit requests. The known sizes stay listed; OpenAI applies the per-model size rules (ac-r8riqc). - `@apicity/fal` `xai.grokImagineImage` and `qwenImage` document the URLs they post to, `https://fal.run/xai/grok-imagine-image` and `https://fal.run/fal-ai/qwen-image`, in place of `https://api.fal.ai/v1/...`, and link each model's API page on fal.ai. The fal README and the CLI catalog named the wrong URLs (ac-nacco3). - `@apicity/cost` prices xAI Grok Imagine as xAI bills it. An image edit also bills each input image, once per request: 0.002 USD on `grok-imagine-image` and 0.01 USD on `grok-imagine-image-2.0`. Before, a one-input `grok-imagine-image` edit that billed 0.022 was estimated at 0.02. `grok-imagine-image-2.0` bills each image at the quality it serves and at its resolution, from the grid that `GET /v1/image-generation-models` lists: low 0.04, 0.05 and 0.06 USD and medium 0.06, 0.07 and 0.08 USD at 1k, 1.5k and 2k. `auto` serves low for a generation and medium for an edit, so the floor warning for medium is gone. Video bills by resolution too. The published per-second figure is the 480p rate, and `grok-imagine-video-1.5` bills 0.14 USD/s at 720p. No 720p rate is known for `grok-imagine-video` or `grok-imagine-video-1.5-lite`, so their 720p estimates keep the 480p rate and warn that it is a floor. A video edit or extension, which keeps its source's resolution, gets the same warning, and 1080p fails closed. Nine recorded calls, 0.62 USD in total, pin each bill to its estimate (ac-k7qv4n). - `@apicity/xai` image requests accept grok-imagine-image-2.0's documented options: `quality` (`low`, `medium` or `auto`), `resolution: "1.5k"`, the `21:9` and `5:2` aspect ratios, and up to five edit sources instead of three. Before, the request schemas refused these values, so `apicity describe` and `.schema.safeParse` rejected valid 2.0 requests (ac-k9mqsp). - The text-to-image and image-to-image frontier boards list OpenAI's own images calls on every GPT Image row, beside fal, KIE and MachGen: `v1.images.generations` for text-to-image and `v1.images.edits` for image-to-image, with the model ids `gpt-image-2.5-sunburst`, `gpt-image-2.5-flare`, `gpt-image-2` and (text-to-image only) `gpt-image-1.5`. They show no price until `@apicity/cost` has an OpenAI image rate (ac-57h4kb). - The `@apicity/fal` Seedance 2.5 US text-, image- and reference-to-video leaves take `draft`, fal's documented flag for a 480p preview that returns the `draft_id` `seedance2p5.us.draft.complete` finishes at 1080p. Before, no typed request could obtain that id. `@apicity/cost` prices a US text- or reference-to-video draft at 480p, as it already did outside the US (ac-8wmej3). - `@apicity/fal` `gptImage1p5` documents the URL it posts to, `https://fal.run/fal-ai/gpt-image-1.5`, in place of `https://api.fal.ai/v1/fal-ai/gpt-image-1.5`, and links the model's API pages on fal.ai for it and its `edit` child. The text-to-image frontier board now maps GPT Image 1.5 (high), rank 9, to that fal leaf as well as the KIE `gpt-image/1.5-text-to-image` id (ac-annrf9). - The text-to-image, image-to-image and video-to-video frontier boards now list the endpoints shipped since their 2026-10-22 capture. Krea 3 (Ultra), image-to-image #5, maps to fal `krea.v3.edit` instead of "not yet in apicity", so `frontier("i2i")` maps 8 of 10 models. GPT Image 2 (high) adds fal `openai.gptImage2` and `openai.gptImage2.edit`, and Kling 3.0 Omni 1080p (Pro) adds fal `klingVideo.o3.pro.videoToVideo.edit` and `klingVideo.o3.pro.videoToVideo.reference`. MachGen `api.v0.generate` joins the GPT Image 2.5 Sunburst, GPT Image 2.5 Flare, GPT Image 2 and Nano Banana 2.1 rows of both image boards and the Nano Banana 2 row of text-to-image. Text-to-image also lists xAI `v1.images.generations` with `grok-imagine-image-2.0` and Google `generateContent` with `gemini-3.1-flash-image`; both are `unpriced`, because `@apicity/cost` has no rate for either id (ac-0hjfmt). - The image-to-video frontier now names the model on its xAI row for Grok Imagine Video 1.5. `frontier("i2v")` lists it as the shared `v1.videos.generations` call with `modelId` `grok-imagine-video-1.5`, the row reference-to-video already has, instead of the `v1.videos.generations.imageToVideo` leaf, which carried no model id. An id built from the provider and model id was `xai:v1.videos.generations.imageToVideo` and matched no model; it is now `xai:grok-imagine-video-1.5`. xAI lists the leaf's `grok-imagine-video-1.5-preview` as an alias of that id, and the price is unchanged at 0.08 USD/s (ac-dp27py). - An unknown flag on an endpoint that already takes parameter flags now tells you to pass any other request field in `--data`. The agent skill's simplefunctions candles example passes the required `--ticker` and puts `limit` in `--data`, instead of a `marketId` that endpoint does not take (ac-7xgdzh). - `@apicity/free-media-upload`'s README no longer asks for an API key: its Quick Start and Middleware samples passed `process.env.FREE-MEDIA-UPLOAD_API_KEY`, which the keyless provider does not take, because doc-gen's settings for it sat under its old name, `free`. The Middleware sample now wraps `uguu.upload` instead of a `get.v1.models` the provider does not have (ac-37sz5l). - Generated READMEs no longer name a hyphenated provider's client with an invalid variable: `@apicity/free-media-upload`'s samples declared `const free-media-upload = …` and called `free-media-upload.catbox.upload(…)`; they now use `freeMediaUpload` (ac-qiac9z). - The CLI no longer saves web pages as media. A JSON result's `url` (or other media key) that answers with `text/html`, such as an OpenAI web-search citation or a docs link in an ElevenLabs answer, is left as sent with no `_savedTo` sibling, and its body is not read; it used to be saved as a `.bin` file. Downloads from one result also no longer overwrite each other: they run at once and were named to the millisecond, so a result with 40 URLs left 36 files and four `_savedTo` siblings that named another URL's file. When a result yields several files, each name now ends in its number (`-1`, `-2`, …); a single download keeps its old name (ac-ht9p12). - `apicity har report` no longer lists every successful fal `v1.queue.submit` under schema too strict. The submit sends the request's `input` member as its whole body, and that body was validated against the whole request schema (`endpoint_id`, `input`, `priority`, …), so `endpoint_id` and `input` always looked missing. The schema sections now check the annotated call input, factory defaults compare the body with `input.input`, and a legacy submit, which has no input, reports `no annotated input`. Error samples still show the wire body (ac-dvsxwf). - `apicity har report` now matches fal queue exchanges whose endpoint id spans path segments (`POST https://queue.fal.run/fal-ai/flux/dev`) to their catalog rows. A `v1.queue.submit` capture with more than one exchange, such as a retried submit, used to leave every exchange auxiliary, because only the sole-entry fallback could attribute a submit; it now resolves through the row's URL, and queue submit, status and result exchanges count as `strict` in `matchedBy`. `@apicity/har` `findMatchingRow` gains a third pass, after the provider-scoped path match, in which fal's `{endpoint_id}` covers one or more whole segments (`matchHarEntrySpanning`), so a fal model row keeps a queue submit to its own path. Other providers match as before (ac-bynay5). - Six fal endpoints written as method shorthand were invisible to the endpoint walker, so their URL comments went unchecked and the catalog and README left them out. `v1.serverless.files.list`, `v1.serverless.apps.queue`, `v1.serverless.metrics` and `v1.workflows.get` now have catalog rows, with call shapes for the `{owner}`/`{name}` and `{username}`/`{workflow_name}` path fields, and `v1.models.usage` and `v1.models.analytics` now appear in the fal README. Call behavior is unchanged (ac-3ba9xb). - fal queue polls are now in the endpoint catalog: `v1.queue.status` (`GET https://queue.fal.run/{endpoint_id}/requests/{request_id}/status`) and `v1.queue.result` (`GET https://queue.fal.run/{endpoint_id}/requests/{request_id}`) have rows and call shapes, so `apicity har report` resolves polling captures instead of listing them as not-in-catalog, and a failed poll counts as a provider error. The `v1.queue.submit` row now names the URL the factory sends, `POST https://queue.fal.run/{endpoint_id}`, instead of `https://api.fal.ai/v1/POST` (ac-8zeblu). - Two provider rules seen only in recorded rejections now apply locally. KIE `pixverse-v6/image-to-video` createTask rejects two `image_urls` without a `template_id` before the request leaves (KIE's 422 "template_id cannot be empty when multiple image_urls are provided"), and the xAI chat completions schema rejects an empty `messages` list (xAI's 400 "Messages cannot be empty") (ac-tsul8r). - `har:promote` accepts repeated and valueless query keys, reserved or escaped characters, and fragments now that capture identities match Polly; malformed percent or UTF-8 query escapes remain refused (ac-eulnmf). - Fixed 31 unreachable CLI catalog rows: 29 KIE Suno routes now resolve through an unambiguous sub-provider fallback, ambiguous roots are named and refused, and fal logs-stream and simplefunctions panel-run rows name their real leaves. HAR promotion follows the same Suno accessor (ac-chttrw). - `@apicity/har` `pollyEntryId` now normalizes a request URL the way Polly does, through a dependency-free port of the qs parse and RFC 3986 stringify behind Polly's url-parse URL: escaped reserved characters and spaces, repeated keys, valueless keys such as S3's `?cors`, bracket arrays and fragments now hash to Polly's id. It reproduces 2,759 of 3,098 committed entry `_id`s with default options (was 2,720); the rest differ by request body (ac-9e2ha7). - xAI video edit and extension examples now use a freely verified, reachable 5-second MP4 sample instead of the old URL that returns 404; examples are explicitly static, and original recordings remain unchanged (ac-fevz1c). - `gen:examples` no longer ships payloads whose own recorded exchange failed: removed 45 examples (dolthub 5, elevenlabs 16, fireworks 2, google 1, kie 20 including five GET status reads, quo 1), replaced ten KIE Gemini and `/api/v1/responses` invalid-key examples with static minimal payloads, and re-sourced fal `bytedance.seedance2p5.draft.complete` and kimicoding `coding.v1.messages` from successful recordings. The report lists demoted sources; a corpus guard checks exact-payload provenance and static schemas (ac-ygpjmh). - HAR capture now scrubs generic credential body fields such as `api_key`, `apiKey`, access tokens and client secrets in requests, responses and call input, including local failures (ac-5ljiib). ### Added - KIE `grok-imagine/upscale` takes `resolution`, the target of the upscale: `720p` (default) or `1080p`, as its docs list it on 2026-11-19. `@apicity/cost` prices it per upscale from the source-to-target pair, which the feed and the models API price alike: 10 credits (0.05 USD) for 480p to 720p, 30 (0.15 USD) for 480p to 1080p and 20 (0.10 USD) for 720p to 1080p. The request names the source task, not its resolution, so the estimate reads `costHints.inputResolution` (`480p` or `720p`) and fails closed without it, and on a 720p source at the 720p target, which KIE does not price. It used to fail closed on every request. `check:kie-feed-drift` now compares the three upscale rows, which it reported NOT COMPARED. No upscale was billed; a not-found `jobs.recordInfo` read is recorded (ac-xnzcl4). - `@apicity/nunchux` `v2.openai.images.edits` edits images with GPT Image 2.6 Sunburst and Flare (`POST /v2/openai/images/edits`). The request takes 1–16 input images and an optional mask or markup. `@apicity/cost` prices the edit from the same declared token rates as generations, and the input images count as image input. Nunchux does not offer `xhigh` or `max`, so the frontier (max) setting cannot be reproduced. A recorded Flare edit of `cat1.jpg` at `low` and `1024x1024` billed 0.030974 USD (ac-die87i). - `@apicity/nunchux` `v2.openai.images.generations` generates with GPT Image 2.6 Sunburst and Flare. `@apicity/cost` prices both from declared token usage at $5 / $8 / $30 per 1M text-input / image-input / image-output tokens. Nunchux does not offer `xhigh` or `max`, so the frontier (max) setting cannot be reproduced. A recorded Flare image at `low` and `1024x1024` billed 0.02477 USD (ac-qggnrr). - KIE Seedream 5.0 Lite (`seedream/5-lite-text-to-image`, `seedream/5-lite-image-to-image`) takes `quality` `ultra` (4K) beside `basic` (2K) and `high` (3K), and an optional `output_format` (`png` or `jpeg`), from the docs read on 2026-11-19. `quality` stays required, and an omitted `output_format` is not filled in. The descriptors now say `high` is 3K, where they said 4K, and give the schemas' `16:9` aspect-ratio default. KIE bills a flat 5.5 credits per image at every tier, so the price is unchanged (ac-bxpbil). - KIE Seedream 5 Pro (`seedream/5-pro-text-to-image`, `seedream/5-pro-image-to-image`) accepts the aspect ratio `21:9`, which both docs pages list as read on 2026-11-19, so the two tasks take the eight documented ratios. The default stays `1:1`. Prices are unchanged at 7 and 14 credits per image at 1K and 2K (ac-51w0vl). - Added KIE Kling 4.0 Long Image `kling-4-0/long-image` on `kie.post.api.v1.jobs.createTask`, from the docs read on 2026-11-19. It renders one tall image as an `ecommerce` product detail page, from 1-3 product images with a short, medium or long `length`, or as a `webpage` design with optional style and subject images, at `1k` or `2k`. `@apicity/cost` prices each page by skill, length and resolution: ecommerce 42.5-155 credits and webpage 99 / 107.5, plus 0.5 credits per input image. This follows the feed and the models API; the docs' sentence that only resolution sets the price disagrees with them. `check:kie-feed-drift` now compares the nine long-image rows (661 cells agree). A recorded ecommerce Short 1k page from one product image billed 43 credits (0.215 USD) at submission, and a not-found `jobs.recordInfo` read is recorded too (ac-zakqnf). - Added KIE Gemini Omni Flash 1.2 Extend `google/gemini-omni-flash-1-2-extend` on `kie.post.api.v1.jobs.createTask`, from the docs read on 2026-11-19. It adds 3-10 seconds (`duration`) to the end of a completed Gemini Omni Flash 1.2 task or extend task (`task_id`; the source is 360p or 720p and at most 15 s), with a `prompt` and up to 3 reference `image_urls`. `@apicity/cost` prices only the added seconds at the source's resolution, which the caller declares as `costHints.inputResolution`: 4 credits (0.02 USD) per second at 360p and 9 (0.045 USD) at 720p, per the feed and the models API. `check:kie-feed-drift` now compares the two extend rows with the extend key, where before they read as base 1.2 cells. The recording extends a fresh 3 s 360p source by 3 s: source and extend billed 12 credits each (0.12 USD in all). A not-found `jobs.recordInfo` read is recorded too (ac-20k2yx). - Added KIE Kling 4.0 Image `kling-4-0/image` on `kie.post.api.v1.jobs.createTask`, from the docs read on 2026-11-19. One id covers text-to-image and image-to-image: up to 10 reference `image_urls` and `elements` together, `resolution` `2k`/`4k`/`8k`, nine aspect ratios, and `mode` `default` (`count` 1-9) or `series` (`count` 2-10, required). The dashed version keeps the id enum-only. `@apicity/cost` prices each output image at 11.3 / 17 / 28.3 credits (0.0565 / 0.085 / 0.1415 USD) at 2k / 4k / 8k, plus 0.5 credits per reference image or element, from the feed and the models API, which agree. `check:kie-feed-drift` compares the four rows (652 cells agree). The frontier boards map Kling Image 4.0 (text-to-image #7, image-to-image #6) to it. A recorded 2k image billed 11.3 credits (0.0565 USD) at success, and a not-found `jobs.recordInfo` read is recorded too (ac-i4vebc). - KIE Seedance 2.0 Mini (`bytedance/seedance-2-mini`) takes `first_frame_url` and `last_frame_url` for image-to-video, from the docs read on 2026-11-19. Each is an image URL or `asset://{assetId}`. The docs list first frame, first and last frames, and multimodal reference as mutually exclusive scenarios, so a frame with any non-empty `reference_*_urls` is refused. Unlike Seedance 2.5, the Mini page requires no `adaptive` aspect ratio for a frame. A frame is an image, so it bills in the existing no-video column (3.8 / 8.2 credits per second at 480p / 720p) (ac-7cp4o7). - KIE GPT Image 2 (`gpt-image-2-text-to-image`, `gpt-image-2-image-to-image`) takes an optional `background` (`transparent`, `opaque` or `auto`), from the docs read on 2026-11-19. The docs support it at 1K only, so a 2K or 4K request with `background` is refused. Image-to-image gains the aspect ratios `2:1`, `1:2`, `3:1`, `1:3` and `9:21`, matching text-to-image. The docs also say 2K does not support 5:4, 4:5, 3:1, 1:3 or 9:21, and 4K does not support 3:1, 1:3 or 9:21. Those combinations are described on the fields but not refused, because text-to-image has always accepted them. Prices are unchanged at 6, 10 and 16 credits per image at 1K, 2K and 4K (ac-ftdouc). - KIE `grok-imagine-image-2-0/segment-map` takes a direct http(s) `image_url` in place of `task_id` on `kie.post.api.v1.jobs.createTask`, from the docs read on 2026-11-19, which make `input` oneOf `{ task_id }` or `{ image_url }`; exactly one is required. It stays free: the feed and the models API price segment-map at 0, and a recorded run on an uploaded `cat1.jpg` billed 0 credits for 36 named segments. A not-found `jobs.recordInfo` read is recorded too (ac-mtdgxe). - KIE Seedance 2.5 (`bytedance/seedance-2-5`) offers its 4K tier: `resolution: "4k"` (3840x2160 at 16:9, H.265 (HEVC) 10-bit, per the docs read on 2026-11-19). It also accepts `omni_reference_task_type` (`auto`, `reference`, `edit`, `extend`) under the rules Flash has: reference materials only, `edit` and `extend` need a reference video and the `adaptive` aspect ratio, and `edit` needs `duration: -1`; both models share `Seedance25OmniReferenceTaskTypeSchema`. `@apicity/cost` prices 4K per second from the 2026-11-19 feed and models API, which agree: 251 credits (1.255 USD) without video input and 155 credits (0.775 USD) with it, on input + output seconds. The two cells are pricing-only reconciliation exceptions backed by a dated evidence file, the `seedance-2-5-4k-unpriced` rule is gone, and `check:kie-feed-drift` compares the two 4K rows with these cells, which agree. No recording: a 4 s 4K clip would bill about 5 USD (ac-vggwxn). - Added KIE Seedance 2.5 Flash `bytedance/seedance-2-5-flash` on `kie.post.api.v1.jobs.createTask`, from the docs read on 2026-11-19. It takes the Seedance 2.5 fields at the Flash limits: a 20480-character prompt (required for text-to-video), up to 10 reference images, 5 videos and 5 audio clips, `480p` or `720p` (default 720p), no draft mode, and `duration` 4-30 or -1 with no default. `omni_reference_task_type` (`auto`, `reference`, `edit`, `extend`) is accepted with reference materials only; `edit` and `extend` need a reference video and the `adaptive` aspect ratio, and `edit` needs `duration: -1`. The first-frame, first-and-last-frame and reference scenarios stay mutually exclusive, through a rule now shared with Seedance 2.5. `@apicity/cost` prices it per second from the feed and the models API: 18.5 / 41 credits (0.0925 / 0.205 USD) at 480p / 720p without video input, and 11.5 / 25.5 credits (0.0575 / 0.1275 USD) with it, on input + output seconds. An omitted or -1 duration needs `costHints.durationSeconds`. `check:kie-feed-drift` now compares the four Flash rows with these cells, and they agree (646 cells); the `seedance-2-5-flash-separate-model` rule is gone, and the Seedance 2.5 branch of the reconciliation no longer claims suffixed variants. The numeric-input audit gains its `duration`. A recorded 4 s 480p clip billed 74 credits (0.37 USD), charged at success, and a not-found `jobs.recordInfo` read is recorded too (ac-zrziuf). - MachGen image requests support automatic sizing, model-specific rendering controls, GIF frames, SVG author settings, edit masks and markup, continued edits, and a safety identifier. `MACHGEN_IMAGE_SURFACES` lists supported settings for each registered model and task (ac-znkz5o). - MachGen supports GPT Image 2.6 Flare text-to-image and image-to-image on `api.v0.generate`, at the same output-token price as GPT Image 2.6 Sunburst. The recorded low-quality 1280×720 generation billed $0.009345 and the 720×1280 edit of one source image $0.156849, matching the estimates (ac-hgm6m1). - MachGen supports GPT Image 2.6 Sunburst text-to-image and image-to-image on `api.v0.generate`, with quality selection through `image_output.quality`. - WaveSpeed supports HeyGen Video 1.0 image-to-video with a first-frame image, prompt and 5–15 second duration. - `@apicity/wavespeed` adds Gemini Omni 1.2 Flash reference-to-video at `api.v3.google.geminiOmni1p2Flash.referenceToVideo` (`POST https://api.wavespeed.ai/api/v3/google/gemini-omni-1.2-flash/reference-to-video`), the leaf beside text-to-video and video-edit: up to 10 reference images and 3 reference videos (each at most 3 s), 3-15 s at 360p, 720p, 1080p or 4k. WaveSpeed bills the text-to-video rates, 0.023 to 0.21 USD/s list, 0.56 USD by default (8 s, 720p); references add nothing. Recorded at 3 s, 360p from the cat photo, which billed 0.069 USD in 31 s (ac-p3eht2). - WaveSpeed supports Gemini Omni 1.2 Flash text-to-video with 3–15 second clips through `api.v3.google.geminiOmni1p2Flash.textToVideo`. - Nunchux Wan 3.0 text-to-video is available through `v1.alibaba.services.aigc.videoGeneration.videoSynthesis`. A two-second 480P clip costs $0.10. - Nunchux Wan 3.0 reference-to-video runs on the same `v1.alibaba.services.aigc.videoGeneration.videoSynthesis` call with `reference_image` media; reference images add nothing, so a two-second 480P clip costs $0.10. It is the Nunchux entry for Wan 3.0 on the reference-to-video frontier. - Nunchux Wan 2.7 video edit (`wan2.7-videoedit`) restyles a 2–10 second source clip on the same `v1.alibaba.services.aigc.videoGeneration.videoSynthesis` call with one `video` media item. It is the Nunchux entry for Wan 2.7 on the video-editing frontier. - Nunchux HappyHorse 1.0 video edit (`happyhorse-1.0-video-edit`) restyles a source clip on the same `v1.alibaba.services.aigc.videoGeneration.videoSynthesis` call with one `video` media item. Nunchux's price feed lists it although the HappyHorse page documents only t2v, i2v and r2v; a free probe on 2026-11-18 showed it served. It is the Nunchux entry for HappyHorse-1.0 on the video-editing frontier. - `@apicity/nunchux` adds the free Wan/HappyHorse task poll `v1.alibaba.tasks(taskId)` (0 USD). - `@apicity/nunchux` adds the free Seedance task poll `v1.bytedance.contents.generations.tasks.retrieve(taskId)` (`GET https://api.nunchux.com/v1/bytedance/contents/generations/tasks/{task_id}`, 0 USD). - Nunchux Seedance 2.5 text-to-video is available through `v1.bytedance.contents.generations.tasks`. The list rate is $10.70 per million video tokens, or $6.40 with reference video. - Nunchux Seedance 2.5 image-to-video now appears on the frontier. The four-second 480p first-frame recording used 39,820 tokens, or $0.426074 at list. - Nunchux Seedance 2.5 reference-to-video now appears on the frontier. A four-second 480p clip guided by a reference image used 38,830 output tokens, or $0.415481 at list. - Nunchux Seedance 2.0 image-to-video is available on the frontier at $7.00 per million tokens for 480p and 720p, or $7.70 for 1080p. - **nunchux**: new `@apicity/nunchux` provider with `v1.credits`, a free account balance and plan-limit read ($0). - Nunchux Nano Banana 2.1 now generates images through `v1.google.v1beta.interactions`. A 1K image lists at $0.0336. - Nunchux Nano Banana 2.1 now edits images through `v1.google.v1beta.interactions`: send the text prompt first, then up to 14 base64 image parts. A 1K edit lists at $0.0336, the same as a 1K generation, and is the Nunchux entry for Nano Banana 2.1 on the image-editing frontier. - `@apicity/nunchux` adds the free MiniMax task poll `v1.minimax.v2.query.videoGeneration(taskId)` (`GET https://api.nunchux.com/v1/minimax/v2/query/video_generation/{task_id}`, 0 USD). - `@apicity/nunchux` adds MiniMax H3 video generation at `v1.minimax.v2.videoGeneration` (`POST https://api.nunchux.com/v1/minimax/v2/video_generation`), polled with `v1.minimax.v2.query.videoGeneration`. One `MiniMax-H3` model serves text-to-video, image-to-video and reference-to-video, chosen by the `content` items; 768P or 2K, 4-15 s. A four-second 768P clip lists at $0.32. It is the Nunchux entry for MiniMax H3 (768p) on the text-to-video frontier. - Nunchux MiniMax H3 image-to-video runs on the same `v1.minimax.v2.videoGeneration` call with a `first_frame` image and no `ratio`, so the output follows the image. A four-second 768P clip lists at $0.32, and it is the Nunchux entry for MiniMax H3 (768p) on the image-to-video frontier. - Nunchux MiniMax H3 reference-to-video runs on the same `v1.minimax.v2.videoGeneration` call with `reference_image` items and no `ratio`, so the output follows the reference set. A four-second 768P clip from one reference image lists at $0.32, and it is the Nunchux entry for MiniMax H3 (768p) on the reference-to-video frontier. - Added KIE Video Upscale `video/upscale`, KIE's own video upscaler and the first `video/` model id, on `kie.post.api.v1.jobs.createTask`, from the docs as re-read on 2026-11-12 (unchanged since the survey). It takes one public http(s) `video_url` (MP4, MOV or MKV, at most 50 MB and 60 s, checked after download) and a `resolution` output box of `"1080P"`, `"2K"` or `"4K"` (uppercase, default `"2K"`); the source is fitted inside the box with its aspect ratio kept. The id stays enum-only: one id is not a grammar. `@apicity/cost` prices each second of input video by the box, 0.0075 / 0.015 / 0.03 USD (1.5 / 3 / 6 credits) at 1080P / 2K / 4K, on the clip length the caller declares as `costHints.durationSeconds`, as for Topaz video. `check:kie-feed-drift` compares the three new rows, and they agree (645 cells). A recorded 1080P upscale of the 3 s `cat1-kling-o3-3s.mp4` (720 x 960 to 1080 x 1440) billed 4.5 credits (0.0225 USD), charged at success (ac-ggy4a4). - Added KIE Image Upscale `kie/image-upscale`, KIE's own upscaler and the first `kie/` model id, on `kie.post.api.v1.jobs.createTask`, from the docs as re-read on 2026-11-12 (unchanged since the survey). It takes one public http(s) `image_url` (JPG, PNG, WEBP or GIF, at most 20 MB) and an `upscale_factor` of `"1"`, `"2"`, `"4"` or `"8"` (default `"2"`) that multiplies the width and the height; `"1"` is a same-size enhancement. The page says numbers are accepted too, but declares and sends strings, so the schema keeps the strings. The id stays enum-only: one id is not a grammar. `@apicity/cost` prices each task by its output pixels, the input size times the factor squared: 0.0075 USD (1.5 credits) up to 4.19 MP (2K), 0.02 up to 16.78 MP (4K) and 0.06 up to 67.11 MP (8K). The request does not carry the input size, so the new cost hints `inputWidth` and `inputHeight` declare it; without them the estimate quotes the 8K tier as an upper bound and warns, and a size KIE rejects with 422 (over 67,108,864 output pixels or 16,384 px per side) fails closed. `check:kie-feed-drift` compares the three new rows, and they agree (642 cells). The numeric-input audit gains its `upscale_factor`. A recorded 2x upscale of `cat1.jpg` (768 x 1024 to 1536 x 2048) billed 1.5 credits (0.0075 USD), charged at submission (ac-s2rue4). - Added KIE Kling 4.0 Flash reference-to-video `kling-4.0/flash-reference-to-video` on `kie.post.api.v1.jobs.createTask`, from the docs as re-read on 2026-11-12 (unchanged since the survey). It takes Kling 4.0 reference-to-video's materials (`image_urls` up to 10, `video_urls` up to 5, `keep_original_sound`, and up to 7 `elements`, at most 3 of them video subjects, 15 materials in all) with the Flash limits (`duration` 3-20, default 5; 720p only) and an `aspect_ratio` that accepts auto. Flash has no standalone `voices`, and each element takes a `type` (image, the default, with 2-4 `image_urls`; video with one `video_url`); the strict input and the request rules check both locally. It reuses the Flash schemas from keyframe-to-video and exports `Kling40FlashElementSchema`. `@apicity/cost` prices it at 0.085 USD per output second without video input and 0.0565 USD per second with it, on output + max(output, input) seconds from `costHints.inputDurationSeconds`, as Kling 4.0 reference does; the docs callback examples are non-Flash copies. `check:kie-feed-drift` compares both new rows, and they agree (639 cells). The numeric-input audit gains its `duration`. A recorded 3 s clip from one `cat1.jpg` reference image billed 51 credits (0.255 USD), charged at submission (ac-xfuy6c). - Added KIE Kling 4.0 Flash keyframe-to-video `kling-4.0/flash-keyframe-to-video` on `kie.post.api.v1.jobs.createTask`, from the docs as re-read on 2026-11-12 (unchanged since the survey). Flash is a cheaper 720p-only tier: it takes a 1-8,000 character prompt, an integer `duration` from 3 to 20 seconds (default 5), `resolution` 720p (the only value), an optional `first_frame_url` mentioned as `@first_frame`, and, without a frame, an `aspect_ratio` of 16:9, 9:16, 1:1 or 21:9 (no auto). Flash takes no middle frames and no last frame, so the input is strict and rejects `keyframe_urls`, `last_frame_url` and any other undocumented field locally. The shared `Kling40FlashDurationSchema` and `Kling40FlashResolutionSchema` are exported for the Flash reference model; the Kling alias already read the id, which is now listed. `@apicity/cost` prices each output second at 0.085 USD (17 credits); the docs page's callback examples are non-Flash copies and are not used. `check:kie-feed-drift` compares the new feed row, and it agrees (637 cells). The numeric-input audit gains its `duration`. A recorded 3 s clip from `cat1.jpg` billed 51 credits (0.255 USD), charged at submission (ac-ye4q12). - Added KIE MiniMax H3 Image edit `minimax-h3/image-edit` on `kie.post.api.v1.jobs.createTask`, from the docs as re-read on 2026-11-12 (unchanged since the survey). It takes text-to-image's fields (the 1-5,000 character prompt, which names the references ``, `` and so on, the `resolution` tier and `aspect_ratio` or a custom `width` and `height`, `seed`, `expand_prompt` and `output_format`) plus 1-16 required `image_urls`; `auto` follows the first reference image. It reuses the MiniMax H3 Image schemas text-to-image added, and the fields both models describe alike now share one definition. `@apicity/cost` prices the output image as text-to-image does, plus 0.0025 USD for every reference image, the first included. `check:kie-feed-drift` compares the reference-image row as that per-image extra (a unit test covers the mapping, and a mutation check showed it fails without the change), and all three new rows agree (636 cells). The numeric-input audit gains its `seed`, `width` and `height`. A recorded 1K edit of `cat1.jpg` billed 2.3 credits (0.0115 USD), charged at submission (ac-urxhp1). - Added KIE MiniMax H3 Image text-to-image `minimax-h3/text-to-image` on `kie.post.api.v1.jobs.createTask`, from the docs as re-read on 2026-11-12 (unchanged since the survey). It takes a 1-5,000 character prompt and one of two sizings: a `resolution` tier (1K, 2K or 4K, documented default 2K) with one of eleven `aspect_ratio` values (documented default auto), or a custom `width` and `height` (1-8,192 each, 655,360-20,000,000 pixels in total, between 1:3 and 3:1). The schema rejects a custom size sent with `resolution` or `aspect_ratio`, as KIE does, and a lone dimension; it leaves the tier and ratio defaults to KIE rather than injecting them. `seed` (0-2,147,483,647), `expand_prompt` (default true) and `output_format` (png or jpeg, default png) complete the contract, and `MiniMaxH3ImageTaskResultJsonSchema` reads the result, which also returns the seed. The model owns the shared MiniMax H3 Image schemas that image-edit will reuse, and the H3 family stays enum-only. `@apicity/cost` prices each image by its total pixels: 0.009 USD up to 4.5 MP (1K and 2K) and 0.0275 up to 20 MP (4K), and a custom size takes its tier from its pixel count. `check:kie-feed-drift` compares the two new feed rows, and both agree (633 cells). The numeric-input audit gains `seed`, `width` and `height`. A recorded 1K square jpeg billed 1.8 credits (0.009 USD), charged at submission (ac-hqj1h2). - Added KIE Seedream 5.5 layer decomposition `seedream/5-5-layer-decomposition` on `kie.post.api.v1.jobs.createTask`, from the docs as re-read on 2026-11-12 (unchanged since the survey). It takes one `image_url` and an optional `prompt` of up to 2,000 characters, with the documented defaults: `size` auto (or 1K, 1.5K or 2K), png for the base image, `max_images` 30 (an integer from 2 to 100, the base image included), and `text_recognition` and `nsfw_checker` false. `@apicity/cost` prices each returned image by size, 0.018 USD at 1K and 1.5K and 0.0315 at 2K and at auto, plus 0.0072 for the input image. That matches the docs callback receipt of 20.34 credits for three images at auto. The estimate reads the returned count from `costHints.outputImages`, which is checked against `max_images`. Without it, the estimate quotes `max_images` as an upper bound and says so in a warning. `check:kie-feed-drift` compares the three new feed rows, and all agree. The numeric-input audit gains `max_images`. A recorded 1K decomposition of `cat1.jpg` with `max_images` 2 returned two images and billed 8.64 credits (0.0432 USD), charged at success (ac-hp1ess). - Added KIE Seedream 5.5 text-to-image `seedream/5-5-text-to-image` on `kie.post.api.v1.jobs.createTask`, from the docs as re-read on 2026-11-12 (unchanged since the survey). It takes a 5-5,000 character prompt and a required `aspect_ratio` (1:1, 4:3, 3:4, 16:9, 9:16, 2:3, 3:2 or 21:9; there is no auto), with the documented defaults: `size` 2K (1K, 1.5K or 2K), png output, an auto background, and `nsfw_checker` false. A transparent background needs png. It reuses the Seedream 5.5 schemas that image-to-image added. `@apicity/cost` prices each output image by size, 0.018 USD at 1K and 1.5K and 0.0315 at 2K, with no input-image extra. That matches the docs callback receipt of 6.3 credits at the 2K default. `check:kie-feed-drift` compares the three new feed rows, and all agree. The model joins the t2i frontier map under Seedream 5.5, the first provider on that row. A recorded 1K square jpeg billed 3.6 credits (0.018 USD), charged at submission (ac-uaqxxs). - Added KIE Seedream 5.5 image-to-image `seedream/5-5-image-to-image` on `kie.post.api.v1.jobs.createTask`. It takes a 5-5,000 character prompt and 1-10 `image_urls`, with the documented defaults: `aspect_ratio` auto, `size` 2K (1K, 1.5K or 2K), png output, an auto background, and `nsfw_checker` false. A transparent background needs png. The Seedream model alias now also reads kie's dashed minor version (`seedream/5-5-*`) beside the dotted one (`seedream/4.5-*`). `@apicity/cost` prices each output image by size, 0.018 USD at 1K and 1.5K and 0.0315 at 2K, plus 0.0072 for every input image. That matches the docs callback receipt of 7.74 credits for a 2K edit of one image. `check:kie-feed-drift` compares the input-image row as that per-image extra. The model joins the i2i frontier map under Seedream 5.5, the first provider on that row. A recorded 1K edit of `cat1.jpg` billed 5.04 credits (0.0252 USD), charged at submission (ac-7z7wn9). - Added KIE Gemini Omni Flash 1.2 `google/gemini-omni-flash-1-2` on `kie.post.api.v1.jobs.createTask`, from the docs as re-read on 2026-11-12, after KIE added `character_ids` to the page. One id covers text with reference images and characters, a first frame, a first and last frame, and a video edit (`video_list`). The schema takes an integer `duration` of 3-15 s (3-10 with both frames) that is required except in a video edit, where it must not be sent; `aspect_ratio` (16:9, 9:16 or auto, default 16:9); `resolution` (default 720p); `seeds`; and the page's exclusions and 7-slot quota. It stays enum-only, like 1.1, because 1.2 rewrote 1.1's request contract. `@apicity/cost` prices it per output second by resolution from the feed: 0.02, 0.045, 0.05 and 0.11 USD at 360p, 720p, 1080p and 4k. A video edit bills the input video's length, declared as `costHints.durationSeconds`. The model joins the t2v, v2v and r2v frontier maps. A recorded 3 s 360p text-to-video billed 12 credits (0.06 USD), charged at submission (ac-mnn08e). - Added KIE GPT Image 2.6 Flare `gpt-image-2-6/flare-text-to-image` on `kie.post.api.v1.jobs.createTask`. KIE's Flare text-to-image page is its Sunburst text-to-image page under another name, so the schema takes the same fields, defaults and format rules, including the GIF `size`. `@apicity/cost` prices it from Flare's own feed rows, which carry Sunburst's five token rates, from declared usage; `check:kie-feed-drift` now compares those rows. A recorded 1K low-quality square jpeg billed 4.46 credits (0.0223 USD), charged at success, the same as Sunburst (ac-o746yh). - Added KIE GPT Image 2.6 Flare image-to-image `gpt-image-2-6/flare-image-to-image` on `kie.post.api.v1.jobs.createTask`. KIE's Flare page is its Sunburst image-to-image page under another name, so the schema takes the same fields and rules: 1-16 `input_urls`, an inpaint `mask_url` or a `markup` overlay, never both, with no mask on gif or svg output, and the shared format rules. `@apicity/cost` prices it from Flare's own feed rows, which carry Sunburst's five token rates, from declared usage. A recorded 1K low-quality edit of `cat1.jpg` billed 5.58 credits (0.0279 USD), charged at success, the same as the Sunburst edit (ac-k5awqh). - Added KIE GPT Image 2.6 Sunburst image-to-image `gpt-image-2-6/sunburst-image-to-image` on `kie.post.api.v1.jobs.createTask`. It shares the text-to-image fields and format rules, requires 1-16 `input_urls`, and takes an inpaint `mask_url` or a `markup` overlay, never both, with no mask on gif or svg output. `@apicity/cost` prices it at the same five token rates from declared usage, the reference images counting as image input. A recorded 1K low-quality edit of `cat1.jpg` billed 5.58 credits (0.0279 USD), charged at success (ac-fahshl). - Added KIE GPT Image 2.6 Sunburst `gpt-image-2-6/sunburst-text-to-image` on `kie.post.api.v1.jobs.createTask`. The request schema follows the docs: a 1-32,000 character prompt, 16 aspect ratios, 1K/2K/4K, six quality levels, planning `complexity`, `moderation`, and the format rules (`output_compression` with jpeg or webp, `size`, `frame_count` and `frame_duration_ms` with gif, the `svg` object with svg, a transparent background with png, webp or gif). The GPT Image model alias accepts the 2.6 spelling, a `/` between the version and the tier. `@apicity/cost` prices it from kie's five token rates (text 4.50, image input 7.20, image output 27.00 USD per 1M, cached input at a quarter), from declared usage only, as OpenAI's per-modality image rows are; KIE token rows with per-modality rates now go through `evaluateDeclaredUsage`. `check:kie-feed-drift` compares a feed row whose last clause names Image with the image rate and reads "Cached Text Input" and "Cached Image Input" as cached input. A recorded 1K low-quality jpeg billed 4.46 credits (0.0223 USD), charged at success (ac-ag1t5r). - `@apicity/fal` adds Gemini Omni Flash 1.2 upscale at `run.geminiOmniFlash.v1p2.upscale` (`POST https://fal.run/google/gemini-omni-flash/v1.2/upscale`): upscales a source video of up to 15 seconds to 720p, 1080p or 4k, completing the `geminiOmniFlash.v1p2` object. `@apicity/cost` prices it dynamically, as fal's other video upscalers are, because the rate and the billed length follow the source video. A recorded 360p-to-720p upscale of an owned 3 s clip billed 0.12403 USD (ac-rmm7j3). - `@apicity/fal` adds Gemini Omni Flash 1.2 extend at `run.geminiOmniFlash.v1p2.extend` (`POST https://fal.run/google/gemini-omni-flash/v1.2/extend`): continues a source video by 3 to 10 seconds at 360p to 4k and returns the whole extended video. `@apicity/cost` prices it dynamically, as the other 1.2 endpoints are. A recorded 3 s, 360p extension of fal's 5 s example clip billed 0.135325 USD, most of the excess over the new seconds being input tokens for the source (ac-bxz9xk). - `@apicity/fal` adds Gemini Omni Flash 1.2 image-to-video at `run.geminiOmniFlash.v1p2.imageToVideo` (`POST https://fal.run/google/gemini-omni-flash/v1.2/image-to-video`): a required first frame and an optional end frame, 3 to 15 seconds at 360p to 4k. `@apicity/cost` prices it dynamically, as the other 1.2 endpoints are. A recorded 3 s, 360p call from the owned cat photo billed 0.089475 USD (ac-4xoxb6). - `@apicity/fal` adds Seedance 2.0 Mini text-to-video at `run.bytedance.seedance2p0.mini.textToVideo` (`POST https://fal.run/bytedance/seedance-2.0/mini/text-to-video`), completing the Mini object. `@apicity/cost` prices it statically at 0.007 USD per 1000 output tokens, as Mini image-to-video is priced; a recorded 4 s, 480p, square call billed 0.2716 USD, exactly the estimate (ac-e1ujc0). - `@apicity/fal` adds Seedance 2.0 Mini reference-to-video at `run.bytedance.seedance2p0.mini.referenceToVideo` (`POST https://fal.run/bytedance/seedance-2.0/mini/reference-to-video`): up to 9 images, 3 videos and 3 audio files as references. `@apicity/cost` prices image references alone statically at 0.007 USD per 1000 output tokens (a recorded call billed 0.2716 USD, exactly the estimate) and fails closed when video or audio references bring in the card's unrecorded reduced rate (ac-5qsajb). - `@apicity/fal` adds Gemini Omni Flash 1.2 text-to-video at `run.geminiOmniFlash.v1p2.textToVideo` (`POST https://fal.run/google/gemini-omni-flash/v1.2/text-to-video`), #6 on the 2026-11-11 text-to-video board: 3 to 15 seconds at 360p to 4k. `@apicity/cost` prices it dynamically, as the other 1.2 endpoints are, and the frontier registry points Gemini Omni Flash 1.2 at it for text-to-video. A recorded 3 s, 360p call billed 0.088792 USD (ac-c9b83v). - `@apicity/fal` adds Gemini Omni Flash 1.2 reference-to-video at `run.geminiOmniFlash.v1p2.referenceToVideo` (`POST https://fal.run/google/gemini-omni-flash/v1.2/reference-to-video`): up to ten reference images and three reference videos, 3 to 15 seconds of output. `@apicity/cost` prices it dynamically, as the 1.2 edit is, and the model now enters the derived reference-to-video frontier board at #4 (t2v Elo 1154). A recorded 3 s, 360p call with the owned cat photo billed 0.092852 USD (ac-7tfowj). - `@apicity/fal` adds Seedance 2.0 Mini image-to-video at `run.bytedance.seedance2p0.mini.imageToVideo` (`POST https://fal.run/bytedance/seedance-2.0/mini/image-to-video`), under a new `bytedance.seedance2p0.mini` object for the other Mini endpoints. `@apicity/cost` prices it statically by output tokens at 0.007 USD per 1000, as 2.0 Fast is priced; a recorded 4 s, 480p, square call from the owned cat photo billed 0.2716 USD, exactly the estimate (ac-1b5mco). - `@apicity/fal` adds Gemini Omni Flash 1.2 video editing at `run.geminiOmniFlash.v1p2.edit` (`POST https://fal.run/google/gemini-omni-flash/v1.2/edit`), #2 on the 2026-11-11 video-editing board, under a new `geminiOmniFlash.v1p2` object for the other 1.2 endpoints. `@apicity/cost` prices it dynamically, as v1.1 is priced: the card bills output seconds by resolution plus input tokens, which no request field fixes. A recorded 360p edit of fal's 5 s example clip billed 7.137917 seconds, 0.194865 USD, and the frontier registry points Gemini Omni Flash 1.2 at it for video editing (ac-0k5eqv). - `@apicity/fal` adds GPT Image 2.6 Flare text-to-image at `run.openai.gptImage2p6.flare.textToImage` (`POST https://fal.run/openai/gpt-image-2.6/flare/text-to-image`), #3 on the 2026-11-11 text-to-image board, with the same request as Sunburst text-to-image: `reasoning_effort`, GIF and SVG output, and their frame and `svg_model` options. `@apicity/cost` prices it dynamically from fal's token card, and the frontier registry points GPT Image 2.6 Flare (max) at it for text-to-image. A recorded call at quality low, reasoning_effort low, one square JPEG billed 0.0217 USD (ac-na2744). - `@apicity/fal` adds GPT Image 2.6 Flare editing at `run.openai.gptImage2p6.flare.edit` (`POST https://fal.run/openai/gpt-image-2.6/flare/edit`), #2 on the 2026-11-11 image-editing board, with the same request as Sunburst editing: up to 16 reference images, a mask, a markup image, `reasoning_effort`, and GIF output. `@apicity/cost` prices it dynamically from fal's token card, and the frontier registry points GPT Image 2.6 Flare (max) at it for image editing. A recorded low-quality edit of the owned cat photo billed 0.0232 USD (ac-0fkzgx). - `@apicity/fal` adds GPT Image 2.6 Sunburst editing at `run.openai.gptImage2p6.sunburst.edit` (`POST https://fal.run/openai/gpt-image-2.6/sunburst/edit`), #1 on the 2026-11-11 image-editing board: up to 16 reference images, a mask, a markup image, `reasoning_effort`, and GIF output. `@apicity/cost` prices it dynamically from fal's token card, and the frontier registry points GPT Image 2.6 Sunburst (max) at it for image editing. A recorded low-quality edit of the owned cat photo billed 0.0232 USD (ac-qd2l6z). - `@apicity/fal` adds GPT Image 2.6 Sunburst text-to-image at `run.openai.gptImage2p6.sunburst.textToImage` (`POST https://fal.run/openai/gpt-image-2.6/sunburst/text-to-image`), #1 on the 2026-11-11 text-to-image board. It adds `reasoning_effort`, GIF and SVG output, and their frame and `svg_model` options to the 2.5 request. `@apicity/cost` prices it dynamically from fal's token card, as 2.5 is priced, and the frontier registry now points GPT Image 2.6 Sunburst (max) at it. A recorded call at quality low, reasoning_effort low, one square JPEG billed 0.0217 USD (ac-pns6xz). - `@apicity/machgen` adds `api.v0.adapters.del(adapterId)` (`DELETE https://api.machgen.ai/api/v0/adapters/{adapterId}`), which deletes a custom LoRA. It is free and safe to repeat: both calls answered `{ adapter, status: "deleted" }`. A deleted LoRA then reads as a 404 and leaves the list. The recording deletes the synthetic LoRA the other adapter recordings made, so the account's list is empty again, and the wallet was unchanged (ac-fj3vil). - `@apicity/machgen` adds `api.v0.adapters.list()` (`GET https://api.machgen.ai/api/v0/adapters`), a free list of the account's custom LoRAs and their storage use: `adapters`, `used_bytes`, `max_bytes` (15 GB), `used_count` and `max_count` (100), typed as `MachGenAdapterList`. The recording lists the ready synthetic LoRA, and the wallet was unchanged (ac-k99bh1). - `@apicity/machgen` adds `api.v0.adapters.retrieve(adapterId)` (`GET https://api.machgen.ai/api/v0/adapters/{adapterId}`), a free read of one custom LoRA. A ready LoRA answers `adapter`, `model`, `status`, `size_bytes` and `created_at`. An unknown, deleted or expired id is a 404 ("Adapter upload not found"). The recording reads the committed synthetic LoRA and an unknown id, and the wallet was unchanged (ac-qo9yxp). - `@apicity/machgen` adds the custom-LoRA commit at `api.v0.adapters.commit(adapterId)` (`POST https://api.machgen.ai/api/v0/adapters/{adapterId}/commit`), the last of MachGen's three upload calls. It checks the uploaded file's size, MD5 and safetensors header and answers the upload with `status: "ready"`, without `expires_at`, `chunk_size` or `upload_url`. It is free and safe to repeat. The file itself goes to the reserve's `upload_url` by `PUT`, a Google Cloud Storage URL with no apicity method. The recording reserves the synthetic 408-byte LoRA again (same id, fresh URL), sends it with a `PUT` outside Polly, so the live URL is never recorded, and commits it: `ready`, and the wallet was unchanged (ac-3xv6pw). - `@apicity/machgen` adds the custom-LoRA reserve at `api.v0.adapters.create` (`POST https://api.machgen.ai/api/v0/adapters?model={model}`), the first of MachGen's three upload calls: reserve, `PUT` the `.safetensors` file to the returned `upload_url`, commit. `MiniMax-H3` is the only model with custom LoRAs (read 2026-11-11). `model` goes in the query and the file's size, MD5 and two configs in the body; the schema pins MachGen's checks from 20 free 400 probes (`tests/fixtures/machgen-evidence/machgen-adapter-reserve-2026-11-11.json`), including the undocumented `renoise` sampler and the 5 GB size limit. Reserving bills nothing: a recorded reserve of a synthetic 408-byte LoRA left the wallet unchanged. The response's one-time `upload_url` lets anyone holding it write the file for 7 days, so the recording keeps only `https://storage.googleapis.com/***`, and a test fails on any live one in a MachGen recording (ac-516jzc). - `@apicity/wavespeed` adds P-Video-2-Pro image-to-video at `api.v3.prunaAi.pVideo2Pro.imageToVideo` (`POST https://api.wavespeed.ai/api/v3/pruna-ai/p-video-2-pro/image-to-video`), #8 on the 2026-11-16 image-to-video frontier board and the first P-Video endpoint in apicity: a first-frame image and an optional last frame, 5-15 s at 480p or 768p, in `speed`, `quality` or `cost` mode, with optional prompt expansion and seed. WaveSpeed bills by mode and resolution, from 0.01 USD/s (`cost`, 480p) to 0.075 USD/s (`quality`, 768p) list, 0.175 USD by default (5 s, 768p, `speed`); the last frame, prompt expansion and seed add nothing. Recorded at 5 s, 480p in `cost` mode from the cat photo, which billed 0.05 USD in 21 s (ac-cdx7ad). - `@apicity/wavespeed` adds Seedream 5.5 edit at `api.v3.bytedance.seedreamV5p5.edit` (`POST https://api.wavespeed.ai/api/v3/bytedance/seedream-v5.5/edit`): edits up to 10 input images from a text instruction at 1k, 1.5k or 2k, as png or jpeg, with an optional transparent or opaque background. WaveSpeed bills 0.02 USD at 1k and 1.5k or 0.035 USD at 2k, plus 0.008 USD per input image, list. The recorded 1k edit of one image is quoted at 0.028 USD (ac-5ttwpq). - `@apicity/wavespeed` adds Grok Imagine Image 2.0 text-to-image at `api.v3.xAi.grokImagineImageV2p0.textToImage` (`POST https://api.wavespeed.ai/api/v3/x-ai/grok-imagine-image-v2.0/text-to-image`): one image per request at 1k or 2k in two quality tiers and 13 aspect ratios. WaveSpeed bills the image by quality and resolution, from 0.04 USD (low, 1k) to 0.08 USD (medium, 2k) list; the default is the 0.08 USD tier. The recorded low-quality 1k image is quoted at 0.04 USD (ac-tkei3h). - `@apicity/wavespeed` adds GPT Image 2 text-to-image at `api.v3.openai.gptImage2.textToImage` (`POST https://api.wavespeed.ai/api/v3/openai/gpt-image-2/text-to-image`): one png, jpeg or webp image per request at 1k, 2k or 4k in three quality tiers. WaveSpeed bills the image by quality and resolution, from 0.01 USD (low, 1k) to 0.72 USD (high, 4k) list, 0.06 USD by default. The recorded low-quality 1k image is quoted at 0.01 USD (ac-w6g9vb). - `@apicity/wavespeed` adds SkyReels V4 text-to-video at `api.v3.skyworkAi.skyreelsV4.textToVideo` (`POST https://api.wavespeed.ai/api/v3/skywork-ai/skyreels-v4/text-to-video`), #10 on the 2026-11-16 text-to-video frontier board and the first SkyReels endpoint in apicity: 3-15 s at 480p, 720p or 1080p, in `std` or `fast` mode, with optional sound (`fast` needs sound off). WaveSpeed bills by mode and resolution, from 0.08 USD/s (`fast`, 480p) to 0.35 USD/s (`std`, 1080p) list, 1.75 USD by default (5 s, 1080p, `std`); sound and aspect ratio add nothing. Recorded at 3 s, 480p in `fast` mode, which billed 0.24 USD in 16 s (ac-9poz7k). - `@apicity/wavespeed` adds GPT Image 2.5 Flare text-to-image at `api.v3.openai.gptImage2p5Flare.textToImage` (`POST https://api.wavespeed.ai/api/v3/openai/gpt-image-2.5-flare/text-to-image`), with the Sunburst request schema. WaveSpeed quotes it at the Sunburst rates, from 0.01 USD (low, 1k) to 1 USD (max, 4k) list, 0.024 USD by default. The recorded low-quality 1k image is quoted at 0.01 USD (ac-cg5q9o). - `@apicity/wavespeed` adds GPT Image 2.5 Sunburst text-to-image at `api.v3.openai.gptImage2p5Sunburst.textToImage` (`POST https://api.wavespeed.ai/api/v3/openai/gpt-image-2.5-sunburst/text-to-image`): one png, jpeg or webp image per request at 1k, 2k or 4k in five quality tiers. WaveSpeed bills the image by quality and resolution, from 0.01 USD (low, 1k) to 1 USD (max, 4k) list, 0.024 USD by default. The recorded low-quality 1k image is quoted at 0.01 USD (ac-3zf6m7). - `@apicity/wavespeed` adds Nano Banana 2.1 edit at `api.v3.google.nanoBanana2p1.edit` (`POST https://api.wavespeed.ai/api/v3/google/nano-banana-2.1/edit`): edits 1-14 input images from a text instruction at 1k, 2k or 4k, with optional web and image search. WaveSpeed bills the text-to-image rate for the resolution and searches, plus 0.002 USD per input image. The recorded 1k edit of one image is quoted at 0.042 USD list (ac-huxf0i). - `@apicity/wavespeed` adds Nano Banana 2.1 text-to-image at `api.v3.google.nanoBanana2p1.textToImage` (`POST https://api.wavespeed.ai/api/v3/google/nano-banana-2.1/text-to-image`): one png or jpeg image per request at 1k, 2k or 4k, with optional web and image search. WaveSpeed bills 0.04, 0.06 and 0.135 USD list by resolution, plus 0.014 USD for each enabled search. The recorded 1k image is quoted at 0.04 USD (ac-huxf0i). - `@apicity/wavespeed` adds GPT Image 2.6 Flare edit at `api.v3.openai.gptImage2p6Flare.edit` (`POST https://api.wavespeed.ai/api/v3/openai/gpt-image-2.6-flare/edit`), with the Sunburst edit's request schema: 1-16 input images and an optional mask or markup image. WaveSpeed quotes it at the Sunburst edit rates: the text-to-image rate for the quality and resolution, plus 0.015 USD per input image, plus 0.015 USD for a mask or markup image. The recorded low-quality 1k edit of one image is quoted at 0.045 USD list (ac-y3hr3z). - `@apicity/wavespeed` adds GPT Image 2.6 Flare text-to-image at `api.v3.openai.gptImage2p6Flare.textToImage` (`POST https://api.wavespeed.ai/api/v3/openai/gpt-image-2.6-flare/text-to-image`), with the Sunburst model's request schema. WaveSpeed quotes it at the Sunburst rates: 0.03 USD (low, 1k) to 0.85 USD (max, 4k) list, 0.05 USD by default; svg adds 0.02 USD. The recorded low-quality 1k image is quoted at 0.03 USD (ac-y3hr3z). - `@apicity/wavespeed` adds GPT Image 2.6 Sunburst edit at `api.v3.openai.gptImage2p6Sunburst.edit` (`POST https://api.wavespeed.ai/api/v3/openai/gpt-image-2.6-sunburst/edit`): edits 1-16 input images from a text instruction, with an optional mask or markup image, at 1k, 2k or 4k in five quality tiers. WaveSpeed bills the text-to-image rate for the quality and resolution, plus 0.015 USD per input image, plus 0.015 USD for a mask or markup image. The recorded low-quality 1k edit of one image is quoted at 0.045 USD list (ac-798tx1). - `@apicity/wavespeed` adds GPT Image 2.6 Sunburst text-to-image at `api.v3.openai.gptImage2p6Sunburst.textToImage` (`POST https://api.wavespeed.ai/api/v3/openai/gpt-image-2.6-sunburst/text-to-image`): one image per request at 1k, 2k or 4k in five quality tiers, as png, jpeg, webp, gif or svg. WaveSpeed bills the image by quality and resolution, from 0.03 USD (low, 1k) to 0.85 USD (max, 4k) list, 0.05 USD by default; svg adds 0.02 USD. The recorded low-quality 1k image is quoted at 0.03 USD (ac-2eh7af). - `@apicity/wavespeed` adds Happy Horse 1.1 reference-to-video at `api.v3.alibaba.happyhorse1p1.referenceToVideo` (`POST https://api.wavespeed.ai/api/v3/alibaba/happyhorse-1.1/reference-to-video`): 1-9 reference images guide a 3-15 s video at 480p, 720p or 1080p. WaveSpeed bills the output's seconds (5 by default) at the list 0.07, 0.14 and 0.18 USD/s; references add nothing. A recorded 3 s 480p render billed 0.21 USD (ac-98rlld). - `@apicity/wavespeed` adds Kling o3 Pro video edit at `api.v3.kwaivgi.klingVideoO3Pro.videoEdit` (`POST https://api.wavespeed.ai/api/v3/kwaivgi/kling-video-o3-pro/video-edit`): edits a 3-15 s video from a text instruction, with up to 4 reference images and 3 elements. WaveSpeed bills only the source clip's length, rounded up, from 3 to 15 s, at the list 0.168 USD/s; the estimate needs `costHints.inputDurationSeconds` and fails closed without it. A recorded edit of a 3 s clip billed 0.504 USD (ac-124he8). - `@apicity/wavespeed` adds Grok Imagine Video 1.5 reference-to-video at `api.v3.xAi.grokImagineVideoV1p5.referenceToVideo` (`POST https://api.wavespeed.ai/api/v3/x-ai/grok-imagine-video-v1.5/reference-to-video`): up to 7 reference images guide a 1-15 s video at 480p or 720p. WaveSpeed bills the seconds (6 by default) at the list 0.08 and 0.14 USD/s, plus 0.01 USD per reference image. A recorded 1 s 480p render billed 0.09 USD (ac-7885m6). - `@apicity/wavespeed` adds Runway Aleph 2 video edit at `api.v3.runwayml.aleph2` (`POST https://api.wavespeed.ai/api/v3/runwayml/aleph-2`), apicity's first endpoint for the video-to-video board's Aleph 2.0, which leaves that board fully mapped: edits a 2-30 s video from a text instruction, with up to 5 keyframes. WaveSpeed bills only the source clip's length, rounded up, from 2 to 30 s, at the list 0.37 USD/s; the estimate needs `costHints.inputDurationSeconds` and fails closed without it. A recorded edit of a 3 s clip billed 1.11 USD (ac-jogzs9). - `@apicity/wavespeed` adds Gemini Omni 1.1 Flash reference-to-video at `api.v3.google.geminiOmni1p1Flash.referenceToVideo` (`POST https://api.wavespeed.ai/api/v3/google/gemini-omni-1.1-flash/reference-to-video`): up to 10 reference images and 3 reference videos guide a 3-10 s video at 360p, 720p, 1080p or 4k. WaveSpeed bills the output's seconds (8 by default) at the list 0.03, 0.10, 0.15 and 0.30 USD/s; references add nothing. A recorded 3 s 360p render billed 0.09 USD (ac-rt31ij). - `@apicity/wavespeed` adds Wan 2.7 video edit at `api.v3.alibaba.wan2p7.videoEdit` (`POST https://api.wavespeed.ai/api/v3/alibaba/wan-2.7/video-edit`): edits a video from a text instruction, with up to 3 reference images and a negative prompt, at 720p or 1080p for 2-10 s or the input's length. WaveSpeed bills the output alone, 2-10 s, at the list 0.20 USD/s (720p) and 0.30 USD/s (1080p). A set `duration` needs no hint; with `duration` 0 or omitted the output follows the source, so the estimate needs `costHints.inputDurationSeconds` and fails closed without it. A recorded 2 s 720p edit billed 0.40 USD (ac-szxfdn). - `@apicity/wavespeed` adds Seedance 2 reference-to-video at `api.v3.bytedance.seedance2p0.referenceToVideo` (`POST https://api.wavespeed.ai/api/v3/bytedance/seedance-2.0/reference-to-video`): up to 9 reference images, 3 reference videos and 3 reference audio clips guide a 4-15 s video at 480p, 720p, 1080p or 4k. WaveSpeed bills it as Seedance 2 text-to-video: the list 0.12, 0.24, 0.60 and 1.20 USD/s, or 0.075, 0.15, 0.375 and 0.75 USD/s for the output plus the reference videos' length, which the estimate takes from `costHints.inputDurationSeconds` and fails closed without. A recorded 4 s 480p render billed 0.432 USD against the 0.48 list under a 90% promotion (ac-hgtlgj). - `@apicity/wavespeed` adds Seedance 2 video edit at `api.v3.bytedance.seedance2p0.videoEdit` (`POST https://api.wavespeed.ai/api/v3/bytedance/seedance-2.0/video-edit`), apicity's first endpoint for the video-to-video board's Dreamina Seedance 2.0: up to 9 reference images and 3 reference audio clips, six aspect ratios, 480p to 4k and a 4-15 s output. WaveSpeed bills the source clip's length (2-15 s, rounded up) plus the output's, which follows the source (at least 4 s) unless `duration` is set, at the list 0.075, 0.15, 0.375 and 0.75 USD/s; the estimate needs `costHints.inputDurationSeconds` and fails closed without it. A recorded 4 s 480p edit of a 3 s clip billed 0.4725 USD against the 0.525 list under a 90% promotion (ac-batizj). - `@apicity/wavespeed` adds Happy Horse 1.0 video edit at `api.v3.alibaba.happyhorse1p0.videoEdit` (`POST https://api.wavespeed.ai/api/v3/alibaba/happyhorse-1.0/video-edit`): edits a video from a text instruction, with up to 5 reference images, at 720p or 1080p. WaveSpeed bills only the source clip's length, rounded up, from 3 to 60 s, at the list 0.28 USD/s (720p) and 0.48 USD/s (1080p); the estimate needs `costHints.inputDurationSeconds` and fails closed without it. A recorded 720p edit of a 3 s clip billed 0.84 USD (ac-zm0fih). - `@apicity/wavespeed` adds Seedance 2.5 reference-to-video at `api.v3.bytedance.seedance2p5.referenceToVideo` (`POST https://api.wavespeed.ai/api/v3/bytedance/seedance-2.5/reference-to-video`): up to 30 reference images, 10 reference videos and 10 reference audio clips guide a 4-30 s video at 480p, 720p, 1080p or 4k. WaveSpeed bills it as Seedance 2.5 text-to-video: the list 0.18, 0.36, 0.90 and 1.80 USD/s, or 0.11, 0.22, 0.55 and 1.10 USD/s for the output plus the reference videos' length when there are any. `WaveSpeedBytedanceSeedance2p5ReferenceToVideoRequestSchema` follows WaveSpeed's model catalog. Recorded with the cat photo as the one reference at 4 s, 480p and no audio, which billed 0.648 USD (list 0.72, under a 90% promotion) in 370 s (ac-ubt29p). - `@apicity/wavespeed` adds MiniMax H3 video edit at `api.v3.wavespeedAi.minimaxH3.videoEdit` (`POST https://api.wavespeed.ai/api/v3/wavespeed-ai/minimax-h3/video-edit`), apicity's first endpoint for the video-to-video board's MiniMax H3 (768p) entry. It is WaveSpeed's open-weights MiniMax H3, because MiniMax's own `minimax/h3` has no video edit on WaveSpeed. It edits a clip, with optional reference images and audio, into a 3-15 s video at 480p, 540p, 768p or 1080p. WaveSpeed bills the source's seconds (at most 15) and the output's at the list 0.05, 0.075, 0.10 and 0.20 USD/s, plus 0.02 USD per reference. `WaveSpeedWavespeedAiMinimaxH3VideoEditRequestSchema` follows WaveSpeed's model catalog. Recorded as a 3 s 480p edit of the 3 s cat clip, which billed 0.15 USD (list 0.30, under a 50% promotion) in 44 s (ac-yn8cbg). - `@apicity/wavespeed` adds MiniMax H3 reference-to-video at `api.v3.minimax.h3.referenceToVideo` (`POST https://api.wavespeed.ai/api/v3/minimax/h3/reference-to-video`): up to 9 reference images, 3 reference videos and 3 reference audio clips guide a 4-15 s video at 768p or 2k. WaveSpeed bills the output's seconds plus the reference videos' length (2-15 s) at the list 0.10 and 0.14 USD/s, and 0.05 USD for each reference image past the fifth. `WaveSpeedMinimaxH3ReferenceToVideoRequestSchema` follows WaveSpeed's model catalog. Recorded with the cat photo as the one reference at 4 s and 768p, which billed 0.40 USD in 172 s (ac-ifjwtg). - `@apicity/wavespeed` adds Gemini Omni 1.1 Flash video edit at `api.v3.google.geminiOmni1p1Flash.videoEdit` (`POST https://api.wavespeed.ai/api/v3/google/gemini-omni-1.1-flash/video-edit`): edits a video from a text instruction at 360p, 720p, 1080p or 4k. WaveSpeed bills 0.02 USD/s to read the source (at most 10 s) plus the output, as long as the source and at least 3 s, at the list 0.04, 0.12, 0.18 and 0.36 USD/s. `WaveSpeedGoogleGeminiOmni1p1FlashVideoEditRequestSchema` follows WaveSpeed's model catalog. Recorded as a 360p edit of the 3 s cat clip, which billed 0.18 USD in 43 s (ac-51eoy3). - `@apicity/wavespeed` adds HeyGen Video reference-to-video at `api.v3.heygen.video1p0.referenceToVideo` (`POST https://api.wavespeed.ai/api/v3/heygen/video-1.0/reference-to-video`): up to 9 reference images, 3 reference videos and 3 reference audio clips guide a 5-15 s video at 480p, 768p, 1080p or 2k. WaveSpeed bills the output's seconds plus the reference videos' full length at the list 0.085, 0.12 and 0.22 USD/s, and 0.05 USD for each reference image past the second. `WaveSpeedHeygenVideo1p0ReferenceToVideoRequestSchema` follows WaveSpeed's model catalog. Recorded with the cat photo as the one reference at 5 s and 480p, which billed 0.425 USD in about 16 s (ac-5rikyh). - `@apicity/wavespeed` adds Seedance 2.5 video edit at `api.v3.bytedance.seedance2p5.videoEdit` (`POST https://api.wavespeed.ai/api/v3/bytedance/seedance-2.5/video-edit`): edits a clip of up to 30 s, with optional reference images and audio, at 480p, 720p, 1080p or 4k. WaveSpeed bills the source's seconds twice (source and an equal-length output, with a 4 s floor) at the list 0.11, 0.22, 0.55 and 1.10 USD/s. `WaveSpeedBytedanceSeedance2p5VideoEditRequestSchema` follows WaveSpeed's model catalog. Recorded as a 480p edit of the 3 s cat clip, which billed 0.792 USD (list 0.88, under a 90% promotion) in 157 s (ac-r7jkle). - `@apicity/wavespeed` adds Wan 3.0 reference-to-video at `api.v3.alibaba.wan3p0.referenceToVideo` (`POST https://api.wavespeed.ai/api/v3/alibaba/wan-3.0/reference-to-video`): up to 10 reference images, 5 reference videos and 5 reference audio clips guide a 2-30 s video at 480p, 720p or 1080p, at WaveSpeed's list 0.05, 0.10 and 0.20 USD/s, with the reference videos' seconds billed on top. `WaveSpeedAlibabaWan3p0ReferenceToVideoRequestSchema` follows WaveSpeed's model catalog. Recorded with the cat photo as the one reference at 2 s, 480p and no audio, which billed 0.095 USD (list 0.10, under a 95% promotion) in 84 s (ac-b7gw8b). - `@apicity/wavespeed` adds Wan 3.0 video edit at `api.v3.alibaba.wan3p0.videoEdit` (`POST https://api.wavespeed.ai/api/v3/alibaba/wan-3.0/video-edit`), apicity's first endpoint for the video-to-video board's #1 entry: edits a clip of up to 15 s, with optional reference images and audio, into a 2-15 s video at 480p, 720p or 1080p. WaveSpeed bills the input's seconds and the output's at the list 0.05, 0.10 and 0.20 USD/s. `WaveSpeedAlibabaWan3p0VideoEditRequestSchema` follows WaveSpeed's model catalog. Recorded as a 2 s 480p edit of the 3 s cat clip, which billed 0.25 USD in 91 s (ac-qvgsyd). - `@apicity/wavespeed` adds HappyHorse 1.1 image-to-video at `api.v3.alibaba.happyhorse1p1.imageToVideo` (`POST https://api.wavespeed.ai/api/v3/alibaba/happyhorse-1.1/image-to-video`): 3-15 s at 480p, 720p or 1080p, at WaveSpeed's list 0.07, 0.14 and 0.18 USD/s. `WaveSpeedAlibabaHappyhorse1p1ImageToVideoRequestSchema` follows WaveSpeed's model catalog. Recorded from the cat photo at 3 s and 480p, which billed 0.21 USD in 79 s (ac-h7lqna). - `@apicity/wavespeed` adds Grok Imagine Video 1.5 image-to-video at `api.v3.xAi.grokImagineVideoV1p5.imageToVideo` (`POST https://api.wavespeed.ai/api/v3/x-ai/grok-imagine-video-v1.5/image-to-video`): 1-15 s at 480p, 720p or 1080p, at WaveSpeed's list 0.08, 0.14 and 0.25 USD/s plus 0.01 USD for the input image. `WaveSpeedXAiGrokImagineVideoV1p5ImageToVideoRequestSchema` follows WaveSpeed's model catalog. Recorded from the cat photo at 1 s and 480p, which billed 0.09 USD in 80 s (ac-rttjg9). - `@apicity/wavespeed` adds FLUX 3 image-to-video at `api.v3.blackForestLabs.flux3.imageToVideo` (`POST https://api.wavespeed.ai/api/v3/black-forest-labs/flux-3/image-to-video`): a start image, 5-20 s at 720p or 1080p, at WaveSpeed's list 0.17 and 0.29 USD/s. `WaveSpeedBlackForestLabsFlux3ImageToVideoRequestSchema` follows WaveSpeed's model catalog. Recorded from the cat photo at 5 s, 720p and no audio, which billed 0.85 USD in 95 s (ac-8x2ghe). - `@apicity/wavespeed` adds Seedance 2.0 image-to-video at `api.v3.bytedance.seedance2p0.imageToVideo` (`POST https://api.wavespeed.ai/api/v3/bytedance/seedance-2.0/image-to-video`): a start frame and optional last frame, 4-15 s at 480p, 720p, 1080p or 4k, at WaveSpeed's list 0.12, 0.24, 0.60 and 1.20 USD/s. `WaveSpeedBytedanceSeedance2p0ImageToVideoRequestSchema` follows WaveSpeed's model catalog. Recorded from the cat photo at 4 s, 480p and no audio, which billed 0.432 USD (list 0.48, under a 90% promotion) in 248 s (ac-7t3w7s). - `@apicity/wavespeed` adds Wan 3.0 image-to-video at `api.v3.alibaba.wan3p0.imageToVideo` (`POST https://api.wavespeed.ai/api/v3/alibaba/wan-3.0/image-to-video`): a first frame and optional last frame, 2-30 s at 480p, 720p or 1080p, at WaveSpeed's list 0.05, 0.10 and 0.20 USD/s. `WaveSpeedAlibabaWan3p0ImageToVideoRequestSchema` follows WaveSpeed's model catalog. Recorded from the cat photo at 2 s, 480p and no audio, which billed 0.095 USD (list 0.10, under a 95% promotion) in 203 s (ac-up83qi). - `@apicity/wavespeed` adds Gemini Omni 1.1 Flash image-to-video at `api.v3.google.geminiOmni1p1Flash.imageToVideo` (`POST https://api.wavespeed.ai/api/v3/google/gemini-omni-1.1-flash/image-to-video`): an image to animate and an optional end frame, 3-10 s at 360p, 720p, 1080p or 4k, at WaveSpeed's list 0.03, 0.10, 0.15 and 0.30 USD/s. `WaveSpeedGoogleGeminiOmni1p1FlashImageToVideoRequestSchema` follows WaveSpeed's model catalog. Recorded from the cat photo at 3 s and 360p, which billed 0.09 USD in 23 s (ac-11jkwt). - `@apicity/wavespeed` adds Seedance 2.5 image-to-video at `api.v3.bytedance.seedance2p5.imageToVideo` (`POST https://api.wavespeed.ai/api/v3/bytedance/seedance-2.5/image-to-video`): a start frame and optional last frame, 4-30 s at 480p, 720p, 1080p or 4k, at WaveSpeed's list 0.18, 0.36, 0.90 and 1.80 USD/s. `WaveSpeedBytedanceSeedance2p5ImageToVideoRequestSchema` follows WaveSpeed's model catalog. Recorded from the cat photo at 4 s, 480p and no audio, which billed 0.648 USD (list 0.72, under a 90% promotion) in 142 s (ac-4ku5th). - `@apicity/wavespeed` adds MiniMax H3 image-to-video at `api.v3.minimax.h3.imageToVideo` (`POST https://api.wavespeed.ai/api/v3/minimax/h3/image-to-video`): a first frame (`image`) and optional last frame (`last_image`), 4-15 s at 768p or 2k, at WaveSpeed's list 0.10 and 0.14 USD/s. `WaveSpeedMinimaxH3ImageToVideoRequestSchema` follows WaveSpeed's model catalog. The MiniMax H3 and HeyGen Video text-to-video schemas now also carry the catalog's prompt limits (7000 characters; 1-32000). Recorded from the cat photo at 4 s and 768p, which billed 0.40 USD in 95 s (ac-l31zdp). - `@apicity/wavespeed` adds Gemini Omni 1.1 Flash text-to-video at `api.v3.google.geminiOmni1p1Flash.textToVideo` (`POST https://api.wavespeed.ai/api/v3/google/gemini-omni-1.1-flash/text-to-video`): 3-10 s at 360p, 720p, 1080p or 4k in 16:9 or 9:16, at WaveSpeed's list 0.03, 0.10, 0.15 and 0.30 USD/s. `WaveSpeedGoogleGeminiOmni1p1FlashTextToVideoRequestSchema` follows WaveSpeed's model catalog. Recorded at 3 s and 360p, which billed 0.09 USD in 22 s (ac-hvra26). - `@apicity/wavespeed` adds Seedance 2.0 text-to-video at `api.v3.bytedance.seedance2p0.textToVideo` (`POST https://api.wavespeed.ai/api/v3/bytedance/seedance-2.0/text-to-video`): 4-15 s at 480p, 720p, 1080p or 4k, with optional reference images, videos and audio and web search. List price is 0.12, 0.24, 0.60 or 1.20 USD/s by resolution. With reference videos the rate drops to 0.075, 0.15, 0.375 or 0.75 USD/s, and their length (up to 15 s) bills on top of the output. `WaveSpeedBytedanceSeedance2p0TextToVideoRequestSchema` follows WaveSpeed's model catalog. Recorded at 4 s, 480p and no audio, which billed 0.432 USD (list 0.48, under a 90% promotion) in 172 s (ac-okm0to). - `@apicity/wavespeed` adds FLUX 3 text-to-video at `api.v3.blackForestLabs.flux3.textToVideo` (`POST https://api.wavespeed.ai/api/v3/black-forest-labs/flux-3/text-to-video`): 5-20 s at 720p or 1080p, at WaveSpeed's list 0.17 and 0.29 USD/s. `WaveSpeedBlackForestLabsFlux3TextToVideoRequestSchema` follows WaveSpeed's model catalog. Recorded at 5 s, 720p and no audio, which billed 0.85 USD in 69 s (ac-9wenpk). - `@apicity/wavespeed` adds Seedance 2.5 text-to-video at `api.v3.bytedance.seedance2p5.textToVideo` (`POST https://api.wavespeed.ai/api/v3/bytedance/seedance-2.5/text-to-video`): 4-30 s at 480p, 720p, 1080p or 4k, with optional reference images, videos and audio. List price is 0.18, 0.36, 0.90 or 1.80 USD/s by resolution. With reference videos the rate drops to 0.11, 0.22, 0.55 or 1.10 USD/s, and their length bills on top of the output. `WaveSpeedBytedanceSeedance2p5TextToVideoRequestSchema` follows WaveSpeed's model catalog. Recorded at 4 s, 480p and no audio, which billed 0.648 USD (list 0.72, under a 90% promotion) in 157 s (ac-75ufw1). - `@apicity/wavespeed` adds MiniMax H3 text-to-video at `api.v3.minimax.h3.textToVideo` (`POST https://api.wavespeed.ai/api/v3/minimax/h3/text-to-video`), MiniMax's own H3: 4-15 s at 768p or 2k, at WaveSpeed's list 0.10 and 0.14 USD/s. `WaveSpeedMinimaxH3TextToVideoRequestSchema` follows WaveSpeed's model catalog. Recorded at 4 s and 768p, which billed 0.40 USD in 85 s (ac-afqaub). - `@apicity/wavespeed` adds HeyGen Video text-to-video at `api.v3.heygen.video1p0.textToVideo` (`POST https://api.wavespeed.ai/api/v3/heygen/video-1.0/text-to-video`), WaveSpeed's HeyGen video model for the board's HeyGen Video (ultraquality) entry: 5-15 s at 480p or 768p, at WaveSpeed's list 0.025 and 0.045 USD/s. `WaveSpeedHeygenVideo1p0TextToVideoRequestSchema` follows WaveSpeed's model catalog; its `prompt_enhancement` rewrites the prompt and is not a quality switch. Recorded at 5 s and 480p, which billed 0.125 USD in 79 s (ac-os8g84). - `@apicity/wavespeed` adds Wan 3.0 text-to-video at `api.v3.alibaba.wan3p0.textToVideo` (`POST https://api.wavespeed.ai/api/v3/alibaba/wan-3.0/text-to-video`): 2-30 s at 480p, 720p or 1080p, at WaveSpeed's list 0.05, 0.10 and 0.20 USD/s. `WaveSpeedAlibabaWan3p0TextToVideoRequestSchema` follows WaveSpeed's model catalog. Recorded at 2 s, 480p and no audio, which billed 0.095 USD (list 0.10, under a 95% promotion) after about 16 minutes in WaveSpeed's queue (ac-adiuc8). - Added the WaveSpeed provider (`@apicity/wavespeed`) with `api.v3.predictions.result` (`GET https://api.wavespeed.ai/api/v3/predictions/{id}/result`), the shared poll for the frontier video endpoints that follow: WaveSpeed runs one endpoint per model id, and each answers with a prediction id to read until `data.status` is `completed` (results in `data.outputs`) or `failed`, `cancelled`, `timeout` or `deleted`. `createWaveSpeed({ apiKey?, baseURL?, timeout?, fetch? })` sends `WAVESPEED_API_KEY` as `Authorization: Bearer`. API errors reject with `WaveSpeedError`, whose `code` is WaveSpeed's envelope code; the key is redacted from any error text. Replays an unknown prediction (404) and an invalid key (401), both free (ac-fzxklt). - `@apicity/machgen` `api.v0.generate` takes `MiniMax-H3-Omni`, MachGen's hosted reference-to-video model. MachGen's models table omits it, but its model page and pricebook list it. The schema restates MachGen's own 400s from 18 free probes on 2026-11-01: R2V only, 768p, 1-15 s, the `adaptive` and wide aspect ratios, at most 9 reference images, 3 videos and 3 audio clips (12 files in all), audio only beside an image or video, no `reference_video_start_secs`, and the `reference` operation alone (`tests/fixtures/machgen-evidence/machgen-omni-2026-11-01.json`). `@apicity/cost` prices it at the pricebook's 0.02 USD/s, on input plus output seconds with a reference video (slug `mmh3o`). Two recorded R2V clips billed their estimates: 1 s from an image, 20000 micros (task 4c1a75bf), and 1 s from the 4 s reference video, 100000 micros (task bbd1046b) (ac-1aklsl). - `@apicity/cost` prices OpenAI's GPT Image models, so the frontier OpenAI rows show a price. `gpt-image-2`, `gpt-image-2.5-sunburst` and `gpt-image-2.5-flare` bill 30 USD per 1M image output tokens, and the tokens follow from the quality and the `WIDTHxHEIGHT` size by the formula of the calculator in OpenAI's image generation guide. That formula reproduces OpenAI's published `gpt-image-2` price table and the 1756 output tokens of every committed `gpt-image-2` recording. `gpt-image-1.5` is priced per image from its model page: low 0.009 to 0.013, medium 0.034 to 0.05 and high 0.133 to 0.20 USD. `auto` quality or size, the default, depends on the generated image, so it fails closed. Text and image input tokens are not in the request, so every estimate warns that it is a floor. The t2i and i2i frontier boards drop the `"unpriced"` key from their seven OpenAI rows (ac-b12rs0). - `@apicity/cost` prices xAI's `grok-imagine-image-2.0` at 0.04 USD per image and `grok-imagine-video-1.5-lite` at 0.02 USD per second, the rates on docs.x.ai/developers/pricing. The 0.04 is the published low-quality rate, so a request served at medium quality (`quality: "medium"`, or an edit at the default `auto`) keeps it and warns that the estimate is a floor: xAI publishes no medium rate. The text-to-image frontier board now prices the xAI row of Grok Imagine Image 2.0 at 0.04 USD/generation instead of `unpriced` (ac-bf3zo7). - `@apicity/machgen` `api.v0.generate` takes `task_type: "SHARPEN"`, MachGen's video sharpen for MiniMax-H3. The schema restates MachGen's own 400s: MiniMax-H3 only, `video_config.height` 480 or 768 (`MACHGEN_VIDEO_SHARPEN`; MachGen refuses the 1440 its web contract lists), one owned source (a task id in `src_task_ids`, or an `@input/` upload in `src_video_urls`; a public URL is refused), and `duration_secs` rather than `num_frames`. `@apicity/cost` prices it at the pricebook's 0.015 USD/s at 480p and 0.066 USD/s at 768p on the source's seconds, passed as `costHints.inputDurationSeconds`; MachGen's quote bills the source's height and length whatever the request says. The recorded 1 s 480p sharpen billed 15000 micro-USD, the estimate. `video_config.sharpen` on a generation still has no published surcharge, and its warning now names the SHARPEN rates. MachGen's MiniMax-H3 frontier price range now starts at the SHARPEN rate, 0.015 USD/s (ac-xug9gr). - Added `feed.explore.categories` to `@apicity/omkar` (`GET https://tiktok-scraper.omkar.cloud/feed/explore/categories`): the Explore tabs `feed.explore` accepts, each an `id` (the `category` value) and a display `name`. The recorded list is the same 21 tabs as the `category` enum, which the replay test checks (ac-4ev2i7). - Added `feed.explore` to `@apicity/omkar` (`GET https://tiktok-scraper.omkar.cloud/feed/explore`): one tab of TikTok's Explore page for a country, up to 30 videos and a fresh batch on every call, in the same video shape as `feed.trending`. `category` is one of Omkar's 21 Explore tabs (`all`, `singing-dancing`, `comedy`, `sports`, ...; Omkar defaults to `all`) and `country` an ISO 3166-1 alpha-2 code (default `US`). Recorded for Thailand's `all` tab (ac-4ev2i7). - Added the Omkar Cloud TikTok Scraper provider (`@apicity/omkar`) with `feed.trending` (`GET https://tiktok-scraper.omkar.cloud/feed/trending`): the For You feed TikTok shows in a country right now, up to 30 videos with stats, author, sound, media qualities and subtitles, and a fresh batch on every call. `country` takes an ISO 3166-1 alpha-2 code (Omkar defaults to `US`). The key, `OMKAR_API_KEY`, is sent as the `API-Key` header; the free plan allows 1,000 successful requests a month. Gateway errors reject with `OmkarError`, whose `code` carries Omkar's stable error code (`invalid_api_key`, `quota_exceeded`, `rate_limited`, ...). Recorded for Thailand (ac-4ev2i7). - Added FLUX 3 Klein edit-image at `run.blackforestlabs.flux3.klein.editImage`, now callable with its LoRA variant still at `.lora`; its input is the LoRA endpoint's without `loras`. Pricing is dynamic: an edit of our cat photo at 256x256 billed 0.8889 megapixels at $0.009 while the card prices images per resolution tier (ac-eahf2q). - Added FLUX 3 Klein text-to-image at `run.blackforestlabs.flux3.klein.textToImage`, now callable with its LoRA variant still at `.lora`; its input is the LoRA endpoint's without `loras`. Pricing is dynamic: a 256x256 image billed 0.8889 megapixels at $0.009 while the card prices images per resolution tier (ac-eahf2q). - Added TikTok Creative Center `ads.creativeOne.report.getTopContentsOverview` (`GET https://ads.tiktok.com/CreativeOne/Report/GetTopContentsOverview`): the Unix seconds that end the latest daily, weekly and monthly top-video lists, the values `creativeCenterGetTopContentsList` takes as `periodEndTimestamp`. Unofficial and keyless like the list. Recorded on 2026-10-28 (ac-qiac9z). - Added the TikTok Creative Center provider (`@apicity/tiktok-creative-center`), keyless and free, with `ads.creativeOne.report.creativeCenterGetTopContentsList` (`GET https://ads.tiktok.com/CreativeOne/Report/CreativeCenterGetTopContentsList`): a country's top trending videos for a week or month, with views, engagement and author data. These are unofficial, undocumented endpoints behind Creative Center's Trending videos page that TikTok can change; without a TikTok One session each list is the top 4 rows whatever `page` and `limit` ask, and session cookies are not supported. `itemID`, `authorID` and `creatorID` come back as strings because TikTok sends them as JSON numbers above 2^53, and a nonzero `BaseResp.StatusCode` (which Creative Center sends with HTTP 200) rejects with `TiktokCreativeCenterError`. Recorded for Thailand's weekly list, with and without sponsored posts (ac-qiac9z). - Added FLUX 3 Klein edit-image LoRA at `run.blackforestlabs.flux3.klein.editImage.lora`, reusing the shared edit schema and recording our trained edit LoRA with the owned cat photo (ac-6dfs70). - Added FLUX 3 Klein base edit-image LoRA at `run.blackforestlabs.flux3.klein.base.editImage.lora`, with the shared edit schema and a recording using our trained edit LoRA and owned cat photo (ac-1rn8vm). - Added Endless `social.trends.public` (`GET https://api.endless.app/social/trends/public/{region}`): what is trending on X in a region, with each trend's search link and tweet volume (`null` when X reports none), the last 24 hours of snapshots in `timeline`, and the longest-trending names. `region` is one of 63 slugs, `worldwide` and 62 countries, checked by the request schema. The CLI calls it as `apicity endless social.trends.public --region worldwide`. Recorded for worldwide (ac-cle0z4). - Added the Endless provider (`@apicity/endless`), keyless and free, with `social.tiktokTrends.public` (`GET https://api.endless.app/social/tiktok-trends/public/{region}`): a country's most played trending TikTok videos of the last 24 hours, with the longest-trending videos, sounds and hashtags. `region` is one of 20 country slugs, checked by the request schema; calls send `Endless-Api-Version: 1` unless `apiVersion` says otherwise, and failures raise `EndlessError` with Endless's reason in `code`. The CLI calls it as `apicity endless social.tiktokTrends.public --region thailand`, its cost tier is `cheap`, and `@apicity/har` maps `api.endless.app` to it. Recorded for thailand (ac-cle0z4). - Added FLUX 3 Klein text-to-image LoRA at `run.blackforestlabs.flux3.klein.textToImage.lora`, reusing the shared LoRA schema and our trained cat LoRA (ac-1a7zs9). - Added FLUX 3 Klein base text-to-image LoRA at `run.blackforestlabs.flux3.klein.base.textToImage.lora`, with shared LoRA schemas and a recorded image using our trained cat LoRA (ac-uhklr9). - Added FLUX 3 Klein base edit trainer at `run.blackforestlabs.flux3.klein.base.trainer.edit`, preserving the callable base trainer. Pricing is dynamic because the reference-image multiplier depends on the dataset ZIP contents (ac-9uj0pi). - Added FLUX 3 Klein base trainer at `run.blackforestlabs.flux3.klein.base.trainer`, with its shared trainer schema and a recorded 100-step training run on the owned cat photo (ac-tra7ie). - Added shared FLUX 3 Klein trainer and LoRA request schemas, response types and callable types (ac-v8q1o7). The six endpoints follow separately in ac-tra7ie, ac-9uj0pi, ac-uhklr9, ac-1a7zs9, ac-1rn8vm and ac-6dfs70. - Added MachGen `api.v0.upscale.quote` (`POST https://api.machgen.ai/api/v0/upscale/quote`), MachGen's free upscale price check. It takes an UPSCALE task input, creates no task and answers `{ cost_micros }`, the micro-USD the upscale would bill. MachGen's REST docs do not list it (read 2026-10-26), so MachGen may change it without notice: its catalog row has no docs link, and the README, the schema description and the factory comment mark it undocumented. MachGen's web app calls it, and the MCP server's `quote_video_upscale` tool is its documented form (https://www.machgen.ai/docs/mcp#tools). Its schema, `MachGenUpscaleQuoteRequestSchema`, is the generate schema with `task_type` fixed to `UPSCALE` and the five upscale models listed, so it keeps generate's UPSCALE checks. A quote recorded on 2026-10-26 priced a 2160p Topaz Video Generative Fast upscale of a 4 s 480p task at 596000 micros, and the Topaz upscale replay now reads its quote through the endpoint (ac-23mccw). - `@apicity/cost` adds `resolveMachGenUpscaleEstimate(payload, quote, { hints })`, which prices the MachGen upscales its table cannot, such as the four Topaz upscalers, with MachGen's quote: pass `machgen.api.v0.upscale.quote`, and the estimate is its `cost_micros` in USD. Every other request, MiniMax-H3's upscale included, falls through to `computeEstimate` with no network call, and a failed or malformed quote degrades to the local `usd: 0` with a warning. `isMachGenQuotePriced` tells which requests ask the quote. The recorded Topaz upscale billed exactly its quote (ac-23mccw). - MachGen `api.v0.generate` now checks the T2V aspect ratio of the four T2V-only models, which went unchecked: Veo-3.1 and Veo-3.1-Fast take 16:9 and 9:16, Grok-Imagine-Video takes 16:9, 9:16, 1:1 and 4:3, and Kling-o3 takes 16:9, 9:16 and 1:1. The lists come from MachGen's own 400s to four free probes on 2026-10-25, now in `tests/fixtures/machgen-evidence/machgen-grids-2026-10-25.json` (ac-5qfga3). - `@apicity/cost` prices MachGen's T2V-only models from the 2026-10-25 pricebook, per output second by height: Veo 3.1 at 0.471 USD at 720p and 1080p and 0.706 at 2160p (its audio is always on), Veo 3.1 Fast at 0.118, 0.142 and 0.353, and Grok Imagine Video at 0.059 at 480p and 0.083 at 720p. Every MachGen video model is now priced for each task it takes except the Topaz upscalers. Three recorded T2V clips billed exactly their estimates, 2.415 USD gross: Veo 3.1 at 720p for 4 s (1.884), Veo 3.1 Fast at 720p for 4 s (0.472) and Grok Imagine Video at 480p for 1 s (0.059) (ac-5qfga3). - MachGen `api.v0.generate` now validates video-to-video requests. `task_type: "UPSCALE"` joins the schema, for MiniMax-H3 and the four Topaz models in MachGen's models table: Topaz Video Precision, Precision Animation, Generative and Generative Fast. A new `MACHGEN_VIDEO_UPSCALE` export holds each upscale model's target heights and the Topaz engines MachGen checks in `upscale_config.engine`; other `upscale_config` knobs pass through. `MACHGEN_VIDEO_GRIDS` gains the R2V grid of all 16 R2V models: heights, `duration_secs` values and aspect ratios, reference caps (images, videos, audio clips and their total), whether audio needs an image or video beside it, whether a reference video may start at an offset, and the `reference_video_operation` values the model takes. The schema rejects locally, with MachGen's own wording, what MachGen answers with a 400 before accepting the task: an off-grid R2V height or duration, references over a cap, audio alone where a visual reference is required, an unsupported or misaligned `reference_video_start_secs`, an edit or extend the model lacks, a `src_video_urls` entry that is not a public http(s) URL, an UPSCALE without a source, and an off-tier target height or engine. An edit renders at its source's length, so its `duration_secs` is not checked. Two documented rules apply too: exactly one of `duration_secs` and `num_frames`, and each subject name in only one `subject_to_*_ids` map, pointing at references that exist. `reference_order` gains its `{ type, index }` item shape. The grids and rules come from free 400 probes on 2026-10-25, the UPSCALE ones through MachGen's upscale price check; 15 of the 400s are in `tests/failures/machgen` (ac-ks4no2). - `@apicity/cost` prices MachGen video-to-video. R2V is priced on all 16 models that take it, 11 of them new: Vidu Q3 Turbo, Vidu Q3 Pro, MiniMax H3 Turbo, Seedance 2.0 Fast and Mini, Vidu Q4 Preview, PixVerse V6 and C1, HappyHorse 1.0, Kling 3.0 Omni (also T2V, with audio on and off priced apart) and Vidu Q3. A reference video now bills the seconds MachGen reads from it on top of the output on every model that takes one, as MachGen's docs say, not only on Seedance 2.5 and 2.0 and MiniMax H3; receipts confirm it on Seedance 2.0 Fast and Mini, MiniMax H3 Turbo and Wan 3.0. Pass `costHints.inputDurationSeconds` with those seconds, after any start offset. A Seedance 2.5 edit renders, and bills, its source's length. UPSCALE is priced on MiniMax H3, which bills the source's seconds in and out at 0.05 USD/s and takes the source's length from `costHints.inputDurationSeconds`. The Topaz upscalers bill vendor credits that depend on the source, so they fail closed, as does Vidu Upscale Pro: it is on the pricebook, but MachGen answered HTTP 500 to every request for it. 22 recorded tasks billed 4.662 USD gross, each exactly its estimate, or for the Topaz upscale exactly MachGen's quote; the Wan 3.0 one, a 4 s reference with 2 s out at 480p (294000 micros), showed that Wan's single rate bills the reference too. Three probes MachGen accepted by mistake billed another 0.60 USD (ac-ks4no2). - MachGen `api.v0.generate` now validates every video model in MachGen's models table (24 models; 19 take I2V), up from six. A per-model grid (`MACHGEN_VIDEO_GRIDS`) holds the heights, `duration_secs` values and aspect ratios each model accepts for T2V and I2V, whether its I2V takes an end frame (`src_image_urls` entry 1, on 11 models) and whether it renders toward a lone end frame (`keyframe_indices: [-1]`, on MiniMax-H3 and MiniMax-H3-Turbo). The schema rejects locally, with MachGen's own wording, what MachGen answers with a 400 before accepting a task: an unsupported task type, an off-grid height, duration or aspect ratio, an aspect ratio on a model that takes it from the image, a second image or a lone end frame where the model has none, misaligned `keyframe_indices`, and an I2V request with no image or more than two. Model ids outside the table pass unchecked. The grids come from 227 rejected (unbilled) probes on 2026-10-25; ten of the 400s are in `tests/failures/machgen`. Kling's `negative_prompt`, multi-shot and element fields are now in the video config (ac-fs70lj). - Added MachGen `api.v0.tasks.billing` (`GET https://api.machgen.ai/api/v0/tasks/{taskId}/billing`), which reads a finished task's `generation_cost_micros`, `generation_credit_micros` and `net_cost_micros`, so an actual charge can be reconciled with the estimate. `api.v0.tasks` stays callable. The MachGen replays now read their receipts through it (ac-fs70lj). - `@apicity/cost` prices every MachGen image-to-video model for T2V and I2V, adding 13 models: Wan 2.2 A14B, LTX 2.3 Pro, Vidu Q3 Turbo, Vidu Q3 Pro, MiniMax H3 Turbo, Seedance 2.0 Fast and Mini, Kling 3.0, Vidu Q4 Preview, Vidu Q3 Pro Fast, PixVerse V6 and C1, and HappyHorse 1.0. Rates are per second by height, read from the pricebook on 2026-10-25. Vidu Q3 Turbo bills its 1-5, 6-10 and 11-16 s bands one band at a time; Kling 3.0 and PixVerse price audio on and off apart; MiniMax H3 Turbo is at its promotional 0.01 USD/s. The MachGen endpoints now classify as `expensive` instead of `prohibitive`. Each newly priced model has a recorded I2V generation (Wan 2.2 also T2V), as do Kling 3.0 and PixVerse with audio on, a 6 s Vidu Q3 Turbo clip, and MiniMax H3 and H3 Turbo with an end frame and with a lone end frame: all 22 tasks billed exactly their estimate, 2.793 USD in total. One of those receipts corrects an earlier MachGen rate: an H3 I2V request with an end frame bills the Image to Video rate (0.035 USD/s at 480p), not the Frame to Frame rate (0.05), which is now priced for R2V only (ac-fs70lj). - KIE `grok-imagine-video-1-5-preview` accepts `resolution: "1080p"`, which KIE's docs and live model schema now list. At 1080p it takes exactly one `image_urls` entry, as the docs say; 480p and 720p still accept more, because a two-image request succeeded. `@apicity/cost` prices 1080p at $0.04/s (8 credits). The pricing feed has no 1080p row for this model yet, and the product page prints the tier as a second "720p" entry at 8 credits/s, so the rate is a pricing-only reconciliation exception. A recorded 1 s 1080p generation billed 8 credits and returned a 1920x1088 video (ac-0039hv). - KIE `grok-imagine-video-1-5-preview` now accepts the frame controls KIE added on 2026-10-22: `first_frame_url`, `last_frame_url`, `loop_url` (one image as both the first and the last frame, which cannot be combined with a different first or last frame), up to four `keyframes` of `{ url, timestamp_s }`, and `enable_audio`, which defaults to `true`. The createTask guard checks the documented keyframe timing before the request leaves: each `timestamp_s` greater than 0 and less than `duration`, keyframes at least 1/3 second apart. KIE's OpenAPI types `timestamp_s` as an integer, but the API accepted and stored `1.5`, so the schema takes any number in range. Two recorded 480p generations cover the new fields: a 3 s clip pinned at its first frame, a keyframe at 1.5 s and its last frame (7.2 credits), and a 2 s loop with `enable_audio: false`, which returned a video with no audio track (4.8 credits) (ac-2knv18). - Added the KIE OmniHuman 2.0 `omnihuman-2-0` createTask model. Its request schema follows the docs: exactly one source image in `image_urls`, a required `audio_url` (WAV or MP3, 4-300 s), an optional `prompt` of up to 300 characters and `enable_watermark`, which defaults to `true`. It also comes with `Omnihuman20TaskResultJsonSchema`, a `modelInputSchemas` descriptor and pricing of $0.07 per output second. The request has no duration, so the estimate reads the audio length from `costHints.durationSeconds` and rounds it up to a whole second, as the render does. OmniHuman 1.5's human-identification and subject-detection sub-tasks have no 2.0 counterpart. The model has a recorded generation: a 6.008 s clip rendered 7 s and billed 98 credits (ac-bdsciv). - Added KIE PixVerse V6 Max `pixverse-v6-max/text-to-video` and `pixverse-v6-max/image-to-video` createTask models. Their request schemas follow the docs: a 1-5000 character prompt, the documented quality and aspect-ratio enums, an integer 1-15 s duration, and one first-frame image for image-to-video, which rejects `aspect_ratio` because the output follows the image. They also come with `PixverseV6MaxTaskResultJsonSchema`, `modelInputSchemas` descriptors and per-second pricing by quality and audio, from $0.024/s (360p without audio) to $0.112/s (1080p with audio). The KIE PixVerse model alias now accepts a tier after the version, such as `pixverse-v6-max/...`. Each model has a recorded 1 s 360p generation (ac-hp0emr). - Added KIE Kling 4.0 `kling-4.0/reference-to-video` and `kling-4.0/keyframe-to-video` createTask models. Their request schemas check the documented limits: material quotas, element and voice names, and the keyframe frame cap. They also come with `Kling40TaskResultJsonSchema`, `modelInputSchemas` descriptors and per-second pricing: $0.1425/s at 720p and $0.2125/s at 1080p. With video input, reference-to-video bills $0.085/s or $0.1275/s on output + max(output, input) seconds, the input read from `costHints.inputDurationSeconds`. Each model has a recorded 3 s generation (ac-0t56dd). - `@apicity/har` capture calls accept their own `context` (`HarCallMeta.context`), merged shallowly over the capture's `context` for that call only, so one capture instance can attribute concurrent books, runs or attempts. The construction-time context stays the default; recording identity and scrubbing are unchanged (ac-g0ani3). - Documented the first offline video/image learning pass in `docs/har-learning-pass-1.md`: triaged committed HAR findings and videobook overrides, retained newer provider constraints over historical successes, and filed six provider-family follow-ups for unproven constraints and product-dependent defaults. No schemas, defaults or recordings changed (ac-07gwrq). - Added the failing-payload regression suite: `tests/failures//.har` with `.expect.json` sidecars and `tests/failures/passthrough.json`, held by `tests/integration/captured-failures.test.ts` (a cross-cutting guard) and `lint:recordings`. Every provider rejection there is either caught locally with its rule phrase or listed with a reason. It is seeded from the triage of the committed corpus's 139 failed exchanges (130 are not payload failures) and from two plan-evidence captures: 6 caught, 5 pass-throughs (ac-tsul8r). - Added offline `har:promote` to turn successful captures into checked Polly recordings and integration tests, replay-only `POLLY_RECORDINGS_DIR`, and orphan-check `--recordings`/`--tests` roots (ac-3uda2b). - `apicity har report` analyzes capture directories and legacy recordings offline: grouped provider errors, schema acceptance gaps, defaults, response drift, and local failures, with deterministic Markdown and versioned JSON. Shared exchange-success and endpoint-matching helpers now ship in `@apicity/har` (ac-45uanr). - Added CLI `--capture-har`, `APICITY_HAR_CAPTURE` and `APICITY_HAR_DIR` (default `$HOME/.config/apicity/har`), capture paths in envelopes, `har list|show|scrub` commands and a doctor capture row (ac-7m8lw4). - Added opt-in `createHarCapture` with call scopes, failures mode, scrubbed HAR files and an index; Alibaba OSS and fal storage uploads now honor the factory’s injected fetch (ac-s3b2gq). - Added `@apicity/har` with HAR serialization, secret scrubbing and Polly identity helpers. - Added fal HeyGen Video Agent 3 (`fal-ai/heygen/v3/video-agent`) as `run.heygen.v3.videoAgent`, with `"auto"` plus the stock avatar and voice catalogs; dynamically priced, since the agent chooses the billed output length (ac-egreqg). - Added fal HeyGen Video Agent 2 (`fal-ai/heygen/v2/video-agent`) as `run.heygen.v2.videoAgent`, reusing the stock avatar catalog; dynamically priced, since the agent chooses the billed output length and `config.duration` is only approximate (ac-8eqafo). - Added fal Kling 4 video edit (`kling/v4/video-to-video/edit`) as `run.kling.v4.videoToVideo.edit`, owning the new `videoToVideo` object; dynamically priced, since fal bills the source video's length plus the output's ($0.094/s at 720p, $0.141/s at 1080p, counted in $0.157 units) (ac-j3wid6). - Added fal Kling 4 reference-to-video (`kling/v4/reference-to-video`) as `run.kling.v4.referenceToVideo`, with its element rule restated from fal's validator; text and image references are priced per requested second by resolution ($0.157/s at 720p, $0.235/s at 1080p), while reference videos and video elements, which the card bills on input plus output length, warn instead of pricing (ac-jn033c). - Added fal Kling 4 image-to-video (`kling/v4/image-to-video`) as `run.kling.v4.imageToVideo`, priced per requested second by resolution like text-to-video ($0.157/s at 720p, $0.235/s at 1080p), with the shared Kling 4 media URL rule restated from fal's validator (ac-7e9ecr). - Added fal Kling 4 text-to-video (`kling/v4/text-to-video`) as `run.kling.v4.textToVideo`, the first member of a new `run.kling` namespace (separate from `run.klingVideo`), priced per requested second by resolution ($0.157/s at 720p, $0.235/s at 1080p; a multi-shot video bills the sum of its shot durations) (ac-dhxn43). - Added fal HeyGen Filler Word Removal (`fal-ai/heygen/v3/filler-word-removal`) as `run.heygen.v3.fillerWordRemoval`; dynamically priced, since the input video's length and whether any cut is made set the charge (ac-zx7lj9). - Added fal HeyGen Translate Precision (`fal-ai/heygen/v2/translate/precision`) as `run.heygen.v2.translate.precision`, sharing the speed tier's input schema; dynamically priced, billing two $0.05 units per whole output second (the card's $0.10 a second) (ac-2dfgvj). - Added fal HeyGen Translate Speed (`fal-ai/heygen/v2/translate/speed`) as `run.heygen.v2.translate.speed`, with its 178 output languages; dynamically priced, since the translation sets the billed output seconds (ac-hzil38). - Added fal HeyGen Lipsync Precision (`fal-ai/heygen/v3/lipsync/precision`) as `run.heygen.v3.lipsync.precision`, sharing the speed tier's input schema; dynamically priced (ac-k8hxp1). - Added fal HeyGen Lipsync Speed (`fal-ai/heygen/v3/lipsync/speed`) as `run.heygen.v3.lipsync.speed`; dynamically priced, since the source media set the billed output seconds (ac-cjtbhw). - Added fal HeyGen Avatar 5 digital twin (`fal-ai/heygen/avatar5/digital-twin`) as `run.heygen.avatar5.digitalTwin`, with its background, watermark and transparent webm options; dynamically priced (whole output seconds set by the script) (ac-tsxf3o). - Added fal HeyGen Avatar 4 image-to-video (`fal-ai/heygen/avatar4/image-to-video`) as `run.heygen.avatar4.imageToVideo`, restating fal's rule that a prompt without `audio_url` needs a `voice`; dynamically priced (ac-fh66p2). - Added fal HeyGen Avatar 4 digital twin (`fal-ai/heygen/avatar4/digital-twin`) as `run.heygen.avatar4.digitalTwin`, reusing the stock catalogs; dynamically priced (whole output seconds set by the script) (ac-0y2ni6). - Added fal HeyGen Avatar 3 digital twin (`fal-ai/heygen/avatar3/digital-twin`) as `run.heygen.avatar3.digitalTwin`, with HeyGen's published stock avatar (1286) and voice (102) catalogs; dynamically priced, since the spoken script sets the billed output seconds (ac-5jmpds). - Added fal HeyGen Video reference-to-video (`heygen/heygen-video/reference-to-video`) as `run.heygen.heygenVideo.referenceToVideo`, priced per requested second by resolution, and added HeyGen Video (ultraquality) to the derived reference-to-video frontier list at rank 2 (ac-habde4). - Added fal HeyGen Video image-to-video (`heygen/heygen-video/image-to-video`) as `run.heygen.heygenVideo.imageToVideo`, priced per requested second by resolution like text-to-video (ac-3bshuk). - Added fal HeyGen Video text-to-video (`heygen/heygen-video/text-to-video`) as `run.heygen.heygenVideo.textToVideo`, priced per requested second by resolution ($0.02/s at 480p and 768p, $0.06/s at 1080p and 2K; fal billed the recorded call at half the card rate), and mapped the frontier text-to-video board's HeyGen Video (ultraquality) to it (ac-djhegp). - Added fal Ideogram Vector text-to-vector (`ideogram/vector/text-to-vector`) as `run.ideogram.vector.textToVector`, priced per generated SVG by quality ($0.13 low, $0.16 medium, $0.20 high, $0.32 high with reference images) (ac-h61hu4). - Added fal Ideogram Vector image-to-vector (`ideogram/vector/image-to-vector`) as `run.ideogram.vector.imageToVector`, priced at $0.10 per generated SVG (ac-hsbaox). - Added fal HeyGen Voice Turbo text-to-speech (`heygen/heygen-voice/turbo`) as `run.heygen.heygenVoice.turbo`, priced per input character from its card; fal billed the recorded call at half the card rate (ac-52hhep). - Added fal HeyGen Voice text-to-speech (`heygen/heygen-voice`), priced per input character from its card; fal billed the recorded call at half the card rate (ac-7wb7u7). - Added a `price` to each `frontier()` row and to `apicity frontier `, taken from the bundled `@apicity/cost` rates (`0.04-0.16 USD/s`, `0.08 USD/image`, `dynamic`, or `-`). `docs/frontier.md` has the same column. A missing rate stays `unpriced`; no rate table changed. - Added MachGen provider (`@apicity/machgen`) for MiniMax H3, Seedance 2.5/2.0, Wan 3.0, HappyHorse 1.1 and Grok Imagine Video 1.5, with task status, authenticated asset download, per-second pricing and recorded replays. At recording time (2026-10-22) MachGen rejected Wan 3.0 requests that omitted its documented `video_config` (`aspect_ratio: "adaptive"`, `fps: 30`, `audio: true`), at zero charge; the Wan replays use the documented config. - Added `apicity frontier ` and the `frontier()` registry in `@apicity/cost`, keyed by task type, with the ranking moved to `docs/frontier.md`. - Added frontier text-to-image (`t2i`) and image-to-image (`i2i`) snapshots. - Added frontier text-to-video (`t2v`) and video-to-video (`v2v`) snapshots, and a reference-to-video (`r2v`) list derived from the `t2v` and `i2v` top 10. - Added MachGen to the frontier registry for `t2v`, `i2v` and `r2v`, through `api.v0.generate` with the MachGen model id. - Added KIE `gpt-6-1-sol` on `POST /codex/v1/responses`, with per-token pricing and a recorded completion. - Added KIE `claude-haiku-5-5` on `POST /claude/v1/messages`, with per-token pricing and a recorded completion. - Added KIE Seedance 2.5 draft generation and 1080p rendering with recorded replay and current pricing. - Added MiniMax H3 Max reframe-video with queue replay and dynamic billed-unit pricing guidance. - Added Bria Fibo 2 edit with up to four reference images and structured VGL. - Added Bria Fibo 2 generate with structured VGL prompts and image pricing. - First-class `costHints.durationSeconds` metadata for pricing media whose wire request does not contain its billable duration. - Static pricing coverage for Fal video and edit/image endpoints, Alibaba Wan 2.7 Image, and KIE Nano Banana. - Fal request input/output aliases and an endpoint-to-input type map. - `@apicity/cli`: `apicity setup 1password` saves the 1Password vault and service-account token given by `--op-vault` and `--op-token` to the CLI's env file (`APICITY_OP_VAULT` and `APICITY_OP_SERVICE_TOKEN`, the token as given, so `env:VAR` keeps it off disk) and checks the pair once with `op item list`. `--no-verify` skips the check, and `--remove` deletes the two lines. A new directory is created `0700` and a new file `0600`. - `@apicity/cli`: `op://vault/item/field` references in the env file or the environment resolve for the provider a call addresses, all of them in one `op inject`, with the configured token or `op`'s own sign-in. ### Changed - **Breaking:** `@apicity/cli`, `@apicity/cost`: On a host with a pay-gate secret, `apicity` now needs an OTP for every endpoint it marks paid (611 of 1,950 today: 27 `sdk`-gated registry rows plus 584 `cli`-gated billed rows across `metered`, `accrual`, and `unpriced`, LLM calls and S3 requests included), not only the 27 in `PAID_ENDPOINTS`. Nothing changes on a host without a secret, except passing `--otp` on any paid endpoint outside the registry now exits 4 with `paygate-not-configured` instead of being ignored. `gated` is now equal to `paid` on every row; `select(.gate == "sdk")` lists the 27 it used to mark (`CatalogEntry` gains required `gate: "sdk" | "cli" | "none"`, and `apicity describe` adds a `gate:` line). Provider package signatures and direct SDK behavior are unchanged (`dispatchGated` enforces the gate in the CLI for `gate: "cli"` rows, hashing `{ path, body }` when `pathParams` are non-empty or the merged request object otherwise). Pay-gate refusal hints now emit `--provider`, `--method`, `--dot-path`, and `--request-hash` instead of `--payload-file` and include `meta.requestHash`; `@apicity/cost` exports `dispatchGated`, `mintOtp` adds `requestHash`, and `apicity-paygate otp mint` accepts `--request-hash` alongside `--payload-file` (one of the two, never both) (ac-cbun15). To keep an armed host working, mint a grant for each provider (see Added). - KIE input limits follow the docs as re-read on 2026-11-19, where KIE now allows more: `google/nano-banana` and `google/nano-banana-edit` take a `prompt` of up to 20000 characters (was 5000), `kling-2.6/text-to-video` and `kling-2.6/image-to-video` up to 2500 (was 1000), `kling/v2-5-turbo-image-to-video-pro` a `negative_prompt` of up to 2496 (was 500), `qwen2/text-to-image` a `prompt` of up to 5000 (was 800), `seedream/5-pro-text-to-image` and `seedream/5-pro-image-to-image` up to 5000 (was 3000; the minimum stays 3, below the docs' 4), and `grok-imagine/image-to-image` 1-5 `image_urls` (was exactly 1), each referenced in the prompt as `@image1` ... `@image5`. The Grok price does not depend on the number of input images (4 credits per generation in the feed and the models API). The limits follow the docs; no call probed them, and nothing was tightened (ac-r04q6v). - KIE `bytedance/seedance-2-mini` takes `duration: -1`, automatic length, next to 4-15 seconds (default still 5), as its docs define it on 2026-11-20: "Video duration in 4-15 seconds or -1." KIE then bills the duration it generates. `@apicity/cost` prices -1 the way it prices Seedance 2.5: from `costHints.durationSeconds`, failing closed without it. With reference videos, only the output side comes from that hint, and the clips' total length still comes from `costHints.inputDurationSeconds`. A raw -1 estimate used to read -0.041 USD (-1 second at 720p) with no warning. The change follows the docs; no call probed it (ac-6n1pli). - **Breaking:** KIE `grok-imagine-image-2-0/image-edit` now takes the contract KIE documents for it since 2026-11-19: a `prompt` (1-8,000 characters), a required `aspect_ratio` (`1:1`, `2:3`, `3:2`, `16:9`, `9:16` or `auto`) and 1-5 required `image_urls`. `GrokImagineImage2ImageEditRequestSchema` and its descriptor no longer take `task_id` or `mask_indexs`, so a request in the old form fails `.safeParse()` and the createTask guard. KIE moved that form to `grok-imagine-image-2-0/segment-edit`, and a free probe of the old body on image-edit answered 422 "image_urls is required". Migration: send `{ prompt, task_id, mask_indexs }` with `model: "grok-imagine-image-2-0/segment-edit"`, passing the segment-map `index` values unchanged (they count from 0), or send image URLs to image-edit. `GrokImagineImage2ImageEditAspectRatioSchema` is exported. The price is unchanged at 4 credits (0.02 USD) per image, and a recorded 1:1 edit of an uploaded `cat1.jpg` billed exactly that. The Grok Image 2.0 lifecycle recording drops its retired image-edit exchange (ac-sb6m5h). - The frontier registry follows the 2026-11-19 Artificial Analysis boards (was 2026-11-16). Text-to-image: MAI-Image-2.7 (#4) and Kling Image 4.0, out of preview (#7), enter; Kling Image 4.0 (preview) and Canva Craft leave. Image-to-image: MAI-Image-2.7 (#1), Kling Image 4.0 (#6) and MAI-Image-2.7-Flash (#7) enter; MAI-Image-2.7-Preview, Kling Image 4.0 (preview) and MAI-Image-2.6 leave. Text-to-video: Hyperflow H3 (#5) enters and SkyReels V4 leaves, so the board maps 8 of 10 models. The new models have no apicity endpoint yet. Image-to-video, video-to-video and the derived reference-to-video list are unchanged apart from the capture date (ac-tjkysm). - **Breaking:** MachGen request types now allow automatic image sizing. To read a numeric `height` from a request, first narrow its config with `if (config && config.size_mode !== "auto")`. The exported `MachGenImageConfig` explicit-size type still requires `height`. The generate request schema also validates image settings by model and task. For example, version 0.28.1 accepted `xhigh` and `max` quality for `GPT-Image-2` in `.safeParse()`. Use `auto`, `low`, `medium` or `high` for that model (ac-znkz5o). - `apicity` discovery: `paid` now follows each endpoint's reviewed billing record instead of the pay-gate registry, so it says whether a call bills. On this base 595 of 1,933 rows read paid (27 before): metered, accrual and unpriced billing reads paid; free, followup and subscription do not. Two new fields: `billing` (the record's kind) and `gated` (in the pay-gate registry; the same 27 rows). `select(.gated)` gives what `select(.paid)` gave before. `CatalogEntry` gains the two required fields. Invocation is unchanged: only `gated` binds the `--otp` approval slot, so a billed endpoint outside the registry is still called with the body alone (ac-e65yyk). - The frontier registry follows the 2026-11-11 Artificial Analysis boards on all six types: GPT Image 2.6, Seedream 5.5, MAI-Image-2.7-Preview, Gemini Omni Flash 1.2, SkyReels V4 and four MiniMax H3-based models enter unmapped; image-to-video moves to AA-Video-I2V v2.1, and the derived reference-to-video list is now 6 models (ac-pehukx). - `@apicity/cost` prices xAI `grok-imagine-image-quality` as `grok-imagine-image-2.0` at low quality: 0.04, 0.05 and 0.06 USD per image at 1k, 1.5k and 2k, plus 0.01 USD per edit input image, whatever quality the request names. xAI retired the slug on 2026-11-02; it still resolves, and xAI's migration note bills it at the 2.0 low rate. Before, it priced a flat 0.05 USD per image (ac-2ppeko). - `@apicity/cost` prices MachGen MiniMax H3 Turbo at 0.015 USD/s, re-read from the public pricebook on 2026-11-01: the 50% promotion (0.01 USD/s) ended at 06:59Z, and a 25% one off the 0.02 USD/s list rate runs until 2027-01-01T07:59Z. The recorded Turbo receipts stay the 2026-10-25 promotional bills, checked against the dated discount in each pricebook snapshot (ac-lyovv1). - The `api-key` request header no longer enters a Polly recording's match identity, like the other credential headers, in the test harness and in `@apicity/har` `pollyEntryId`. The header was already scrubbed to `***` on disk, but a recording made with the real key matched only that key, so it could not replay without it. No committed recording sent the header before `@apicity/omkar` (ac-4ev2i7). - **Breaking:** MachGen `reference_video_start_secs` is now a list (`number[]`) with one start second per `src_video_urls` entry, as MachGen requires. The schema typed it as a single number, which MachGen rejects with HTTP 400 "Input should be a valid list", so no request that set it could succeed (ac-ks4no2). - **Breaking:** KIE `Seedance25RequestSchema` and `Seedance25InputSchema` (`bytedance/seedance-2-5`) now default `input.generate_audio` to `false`, not `true`, and the `modelInputSchemas["bytedance/seedance-2-5"]` descriptor says so. `.schema.parse()` output and `apicity describe` change. KIE's documented default is still `true`; every recorded seedance-2-5 run sent `false` (3 of 3 in the 2026-10-24 HAR learning pass), so the schema now follows them. createTask still sends the request as written, so an omitted field still gets KIE's `true`: send `generate_audio: false` for a silent clip. The price does not depend on it (ac-lsk3d7). - **Breaking:** KIE `GrokVideo15PreviewRequestSchema` and `GrokVideo15PreviewInputSchema` now default `input.nsfw_checker` to `false`, not `true`, as KIE's docs and live model schema say (read 2026-10-25). `.schema.parse()` output and `apicity describe` change. createTask still sends the request as written, so an omitted field gets KIE's own default. Pass `nsfw_checker: true` to keep KIE's content filter on (ac-lsk3d7, folded from ac-xgp38d). - The 29 provider packages that ship zod schemas now depend on `zod@^4.6.6`, up from `^4.4.3`. 4.6.6 is the first zod release patched for GHSA-mrhx-c7xc-m3w5, so an install can no longer resolve a vulnerable zod. 4.6.6 also rejects an array over its `max` before parsing its elements, so a union schema such as `AlibabaMultimodalGenerationRequestSchema` reports an oversized `content` as one root `invalid_union` (ac-eix0vh). - Root README rewritten; detail moved to docs/paygate.md, docs/middleware.md, docs/providers.md, CONTRIBUTING.md and MCP.md. - `apicity commands`, `apicity providers` and `apicity describe`, and the `apicity ` and endpoint `--help` forms that reuse them, now answer the success envelope `{"ok": true, "data": …}` with `--json` or a non-terminal stdout, like every other command, instead of printing their result at the top level. Read the result under `.data`: `jq length` on that output now counts the envelope's keys. They also accept `--quiet` for the data alone. - **Breaking:** the OTP pay gate is opt-in in `@apicity/cost`, `@apicity/kie`, `@apicity/xai` and `@apicity/cli`. A paid endpoint now dispatches when no pay-gate secret is configured. Configuring one (`paygate: { secret }`, or the CLI's `--paygate-secret-file` / `APICITY_PAYGATE_SECRET_FILE`) arms the OTP gate, whose behavior is unchanged. An OTP passed to an unarmed provider or CLI is an error (`paygate-not-configured`), and a supplied but empty secret still fails closed. What counts as paid does not change. The CLI's doctor reports an unset secret file as the gate being off, at `ok`, and an empty secret file is a `usage` error. Migration: if you relied on paid endpoints failing closed by default, configure the secret. - `@apicity/cli`: env-file `op://` lines now resolve, for the addressed provider's variables, instead of being skipped. References outside a provider's credential variables (`S3_REGION`, `B2_ENDPOINT`, …) are still neither resolved nor exported, and a reference that cannot be resolved is exit 3 `auth`, naming the variable and the reference. - A 1Password token without a vault is accepted: it authenticates `op://` references, and only the vault convention needs both. A vault without a token is still exit 1 `usage`. - `configured` in `apicity providers`, `commands` and `describe`, and doctor's Providers row, now counts an env-file literal, an `op://` reference and the vault convention (vault and token both configured), including a pair saved in the env file. Those commands still never run `op`. - `apicity doctor`'s 1Password CLI row reads the vault and token from flags, the environment and the env file. It reports a token alone as `ok`, a vault alone as an `error`, and with both lists the vault once and warns, by name, about provider variables it lacks. It names where the token came from and never prints a literal one. - `SETUP_FORMS` gains `"1password"`; `apicity setup --remove` and the other forms' `--remove` never open the env file. - The Claude Code plugin's SessionStart hook reads `apicity providers --json` instead of `apicity doctor --json`, so it never reaches 1Password. The line it prints is unchanged. - `isProviderConfigured(name, env, sources?)`: with `sources` from the new `readCredentialSources`, it applies the meaning above; without, it answers exactly as before. - `resolveCredentials` resolves `op://` references ahead of the vault convention, accepts a token without a vault, and takes two more seams, `listItemTitles` and `injectSecrets`, which `EndpointOptions` passes through. - `OpInject` takes an optional second parameter, the token the `op` child receives as `OP_SERVICE_ACCOUNT_TOKEN`. - `SubprocessOptions` gains an optional `env`, the variables layered over the child's inherited environment, and `SubprocessResult` an optional `notFound`, set when the binary is missing. ### Deprecated - `@apicity/fireworks` audio: `inference.v1.audio.transcriptions` and its `.streaming`, `translations`, and `batch.transcriptions`, `batch.translations` and `batch.get`. Fireworks deprecated audio inference on 2026-06-10 with no shutdown date, and withdrew its rates (last listed 2026-04-29, gone by 2026-05-06), so their `@apicity/cost` rows stay unpriced. On 2026-11-05 `audio-prod` answered 401 to a valid key, `audio-turbo` and the streaming host answered 503, and `audio-batch` still accepted an empty submission. The methods stay callable and send requests unchanged (ac-8l1x3z). - `@apicity/xai` `XAI_GROK_IMAGINE_IMAGE_QUALITY` (from `@apicity/xai/zod`). xAI retired the `grok-imagine-image-quality` slug on 2026-11-02 (https://docs.x.ai/developers/migration/imagine-image-quality-nov-2). The slug still resolves: `grok-imagine-image-2.0` serves it with `quality: "low"` and bills it at the 2.0 low rate, so the constant stays exported. Name `grok-imagine-image-2.0` and pass `quality` instead. The xAI README's Imagine Files example now does that (ac-3h5kfe). ### Removed - **Breaking:** withdrawn KIE Midjourney. `@apicity/kie` no longer exports `post.api.v1.mj.generate` or `get.api.v1.mj.recordInfo`, nor the `Mj*` schemas and types. On 2026-11-12 KIE's Midjourney docs pages served the same soft 404 as a made-up page, and Midjourney was absent from `GET /api/v1/models` and the public price feed. KIE offers no replacement. The matching `@apicity/cost` `PAID_ENDPOINTS` entry and billing rows are gone. Ships in 0.28.0 (ac-3b4swu). - **Breaking:** retired Fireworks image workflows. `@apicity/fireworks` no longer exports `inference.v1.workflows.textToImage`, `kontext` or `getResult`, nor the `FireworksTextToImage*`, `FireworksKontext*` and `FireworksGetResult*` schemas and types. On 2026-11-05 every real FLUX model answered a valid key with the same 404 as a made-up model, and Fireworks deprecated image generation on 2026-06-10. The matching `@apicity/cost` billing rows are gone. Ships in 0.27.0 (ac-8l1x3z). - **Breaking:** retired KIE Gemini 2.5 Pro and Gemini 3.5 Flash. `@apicity/kie` no longer exports `gemini25Pro.post.v1.chat.completions`, `gemini35FlashOpenai.post.v1.chat.completions`, or `gemini.post.v1.models.gemini35Flash.streamGenerateContent`, nor `createGemini25ProProvider`, `createGemini35FlashOpenaiProvider`, or the `KieGemini25Pro*` and `KieGemini35Flash*` schemas and types. On 2026-11-05 each route answered HTTP 410 and KIE's docs mark them deprecated. The matching `@apicity/cost` billing rows are gone. Use Gemini 3 Pro, Gemini 3.8 Flash, or the native 3.6 / 3.7 / 3.8 stream. Ships in 0.27.0 (ac-62hbwl). - **Breaking:** retired KIE Seedance 1.5 Pro. `@apicity/kie` no longer has a schema or guard for `bytedance/seedance-1.5-pro`, nor the `Seedance15Pro*` schemas and types, and `@apicity/cost` no longer prices it. On 2026-11-05 a `createTask` with no prompt answered `code: 410` "retired upstream and is no longer available on Kie.ai". ByteDance deprecated the model on 2026-09-15 and shuts it down on 2026-11-11. Use Seedance 2.5, as KIE suggests, or Seedance 2.0 Mini, ByteDance's named replacement. Ships in 0.27.0 (ac-x0qlnu). - **Breaking:** retired fal Sora 2: `@apicity/fal` `run.sora2.textToVideo` and `run.sora2.imageToVideo` with their `FalSora2*` exports, and the matching `@apicity/cost` pricing, slug and display entries. OpenAI removed Sora 2 from its API on 2026-09-24, and fal's gateway now answers 404 for the application. - **Breaking:** retired KIE `sora-watermark-remover`, its `SoraWatermarkRequestSchema`, and the `SoraWatermarkRequest`, `SoraWatermarkRequestInput`, and `SoraWatermarkParsedRequest` exports. Its cost pricing, slug, and display entries are removed as well. - The MCP server: the CLI's `mcp` subcommand and its compatibility bin. MCP support ended with `@apicity/mcp-server` 0.11.2; use the `apicity` CLI and agent skill (see MCP.md). - The MCP-era eager registry surface of `@apicity/cli`: the `makeToolName` and `toSnakeCase` exports and the `Endpoint` fields `toolName`, `fn`, `jsonSchema`, `example` and `pathParams`. `buildRegistry()` stays as the bulk resolver behind the registry tests and `scripts/diag.mjs`. ### Fixed - Seedance 1.0 Lite/Pro estimates now use the wire duration or its billed five-second default; cost-only hints and legacy top-level durations cannot underquote an omitted duration. - Fal area-billed models now use only documented fixed defaults and warn when an omitted image size cannot be priced safely. - Kling 3.0 pricing applies the sound tier when `multi_shots` promotes an omitted sound setting. - ElevenLabs text-to-speech descriptions shown by `apicity describe` now disclose the 10000-character limit that applies when `model_id` is omitted. - Provider-scoped lint and preflight support endpoint-less packages such as `@apicity/cost` and `@apicity/mcp-server` without accepting misspelled names. - The local dependency audit can inventory the full workspace without overflowing Node's default subprocess buffer. - Generated KIE documentation retains the media URL upload guidance added for pre-upload asset identifiers. - The dependency graph pins `body-parser` to the patched 1.20.6 and 2.3.0 release lines for CVE-2026-12590. ## [0.1.0] - 2026-05-16 ### Added - First stable release of the @apicity/\* monorepo.