# Static model catalog for the openai connector — flagship reasoning models
# plus the everyday chat workhorse, in the vendor's own model ids. Pricing is
# cents per million tokens.
# Prompt-cache prices (`cacheRead` = cached input, `cacheWrite` = cache writes)
# and the gpt-5.6-sol pair ($4/$20) read from developers.openai.com/api/docs/pricing
# on 2026-09-20 (Standard, short context). gpt-5.5's cached-input figure is the
# OpenRouter listing's (the vendor page no longer lists the model);
# gpt-5.5-pro and gpt-5.3-chat-latest publish no cache price.
# Sol 6.1: https://developers.openai.com/api/docs/models/gpt-6.1-sol
# (2026-10-03). Tools require Responses; Chat Completions accepts no tools.
# Reasoning cannot be disabled: neither none nor minimal is supported.
# Codex subscription availability: https://learn.chatgpt.com/docs/models
# The GPT-6 family (Astra, Sol, Luna) read from the same pages and from the
# vendor's latest-model guide on 2026-10-03: "GPT-6 Astra and GPT-6.1 Sol
# support Chat Completions, but tool calling requires Responses", and "GPT-6
# Sol and GPT-6 Luna support function calling in Chat Completions only with
# reasoning_effort: none". So Astra declares the Responses tool API (direct
# chat calls it there), while Sol and Luna declare off + toolsRequireOff like
# the 5.6 trio. Astra and Sol bill prompts over 272K input tokens at 2x input
# and cache rates and 1.5x output, which this per-token schema cannot express;
# the short-context prices are listed. Luna's cache writes are 1.25x its input
# rate, as the vendor states it.
- id: gpt-6.1-sol
  provider: openai
  tags:
    - chat
  supportsTools: true
  toolCallingApi: responses
  supportsVision: true
  reasoning:
    knob: effort
  contextWindow: 1050000
  maxOutputTokens: 128000
  pricing:
    inputCentsPerMillion: 200
    outputCentsPerMillion: 1000
    cacheReadCentsPerMillion: 10
    cacheWriteCentsPerMillion: 250
- id: gpt-6-astra
  provider: openai
  tags:
    - chat
  supportsTools: true
  toolCallingApi: responses
  supportsVision: true
  reasoning:
    knob: effort
  contextWindow: 1050000
  maxOutputTokens: 128000
  pricing:
    inputCentsPerMillion: 1000
    outputCentsPerMillion: 5000
    cacheReadCentsPerMillion: 100
    cacheWriteCentsPerMillion: 1250
- id: gpt-6-sol
  provider: openai
  tags:
    - chat
  supportsTools: true
  supportsVision: true
  reasoning:
    knob: effort
    off: none
    toolsRequireOff: true
  contextWindow: 1050000
  maxOutputTokens: 128000
  pricing:
    inputCentsPerMillion: 200
    outputCentsPerMillion: 1000
    cacheReadCentsPerMillion: 20
    cacheWriteCentsPerMillion: 250
- id: gpt-6-luna
  provider: openai
  tags:
    - chat
  supportsTools: true
  supportsVision: true
  reasoning:
    knob: effort
    off: none
    toolsRequireOff: true
  contextWindow: 1050000
  maxOutputTokens: 128000
  pricing:
    inputCentsPerMillion: 10
    outputCentsPerMillion: 50
    cacheReadCentsPerMillion: 1
    cacheWriteCentsPerMillion: 12.5
- id: gpt-5.5
  provider: openai
  tags:
    - chat
  supportsTools: true
  supportsVision: true
  reasoning:
    knob: effort
    # Probed 2026-08-14 via the provider's own 400: "Function tools with
    # reasoning_effort are not supported for gpt-5.5 in /v1/chat/completions.
    # To use function tools, use /v1/responses or set reasoning_effort to
    # 'none'." — chat always carries tools, so every turn (Default step AND
    # sticky effort picks) must send none; the picker offers no levels.
    off: none
    toolsRequireOff: true
  # Vendor model page: 1,050,000-token context, 128,000-token output.
  contextWindow: 1050000
  maxOutputTokens: 128000
  pricing:
    inputCentsPerMillion: 500
    outputCentsPerMillion: 3000
    cacheReadCentsPerMillion: 50
- id: gpt-5.5-pro
  provider: openai
  tags:
    - chat
  supportsTools: true
  supportsVision: true
  reasoning:
    knob: effort
    # Family behavior (see gpt-5.5 above); the pro tier may turn out to be
    # served on /v1/responses only — a 404 here means repoint, not retune.
    off: none
    toolsRequireOff: true
  contextWindow: 400000
  maxOutputTokens: 128000
  pricing:
    inputCentsPerMillion: 3000
    outputCentsPerMillion: 18000
# The chat-tuned variant ships under the vendor's rolling "-latest" alias —
# the bare id 404s ("The model `gpt-5.3-chat` does not exist…", probed
# 2026-08-14). Non-reasoning: takes max_completion_tokens AND a custom
# temperature.
- id: gpt-5.3-chat-latest
  provider: openai
  tags:
    - chat
  supportsTools: true
  supportsVision: true
  contextWindow: 128000
  maxOutputTokens: 16384
  pricing:
    inputCentsPerMillion: 175
    outputCentsPerMillion: 1400
# The 5.6 frontier trio — ids, sizes, and prices from the vendor's model
# catalog page (sol $5/$30, terra $2/$12, luna $0.20/$1.20 per million;
# 1.05M context / 128K output each). The page frames them as Responses-API
# models, but /v1/chat/completions serves them too — its parameter 400
# ("Function tools with reasoning_effort are not supported for gpt-5.6-sol…")
# proves the route and names the restriction, so all three declare
# off + toolsRequireOff. Codex variants are Responses-only and NOT listed;
# gpt-5.6-luna-pro exists only in the openrouter catalog, not here.
- id: gpt-5.6-sol
  provider: openai
  tags:
    - chat
  supportsTools: true
  supportsVision: true
  reasoning:
    knob: effort
    off: none
    toolsRequireOff: true
  contextWindow: 1050000
  maxOutputTokens: 128000
  pricing:
    inputCentsPerMillion: 400
    outputCentsPerMillion: 2000
    cacheReadCentsPerMillion: 40
    cacheWriteCentsPerMillion: 500
- id: gpt-5.6-terra
  provider: openai
  tags:
    - chat
  supportsTools: true
  supportsVision: true
  reasoning:
    knob: effort
    off: none
    toolsRequireOff: true
  contextWindow: 1050000
  maxOutputTokens: 128000
  pricing:
    inputCentsPerMillion: 200
    outputCentsPerMillion: 1200
    cacheReadCentsPerMillion: 20
    cacheWriteCentsPerMillion: 250
- id: gpt-5.6-luna
  provider: openai
  tags:
    - chat
  supportsTools: true
  supportsVision: true
  reasoning:
    knob: effort
    off: none
    toolsRequireOff: true
  contextWindow: 1050000
  maxOutputTokens: 128000
  pricing:
    inputCentsPerMillion: 20
    outputCentsPerMillion: 120
    cacheReadCentsPerMillion: 2
    cacheWriteCentsPerMillion: 25
# Text-to-speech: powers "Read replies aloud" / "Speak out loud". The context
# window is nominal — synthesis takes character-capped requests, not context.
- id: gpt-4o-mini-tts
  provider: openai
  tags:
    - text-to-speech
  supportsTools: false
  supportsVision: false
  contextWindow: 2000
  tts:
    defaultVoice: alloy
    voicesByLocale:
      en: alloy
      de: onyx
      fr: nova
    audioFormat: mp3
    centsPerMillionCharacters: 1200
# Transcription: powers browser dictation's server fallback and audio/video
# attachment transcription, both through `/audio/transcriptions`. Without an
# entry carrying this tag `resolveTranscriptionModel` finds nothing and every
# upload fails `NO_TRANSCRIPTION_MODEL` — the capability exists only because
# it is declared here, exactly as with the embedding widths below.
#
# whisper-1 rather than the gpt-4o-transcribe family: the request asks for
# `response_format=verbose_json` and the paragraphizer needs the per-segment
# timestamps it returns, which the newer transcribe models do not serve.
# Pricing is per audio MINUTE, which this catalog's per-token schema cannot
# express — so it is omitted and the ledger records minutes at a zero cost
# estimate. The context window is nominal: transcription takes a file.
- id: whisper-1
  provider: openai
  tags:
    - transcription
  supportsTools: false
  supportsVision: false
  contextWindow: 2000
# Image generation: the models agents call through the generate_image tool,
# over /v1/images/generations (and /v1/images/edits when the agent passes
# reference images). They write images, never text, so they carry no chat
# tag and every chat list skips them. Prices read from
# developers.openai.com/api/docs/pricing on 2026-09-29 (Standard): `input` is
# the text-input price, `imageInput` the reference-image input price and
# `output` the image-output price, per million tokens — the ledger costs a
# call from the token counts the response reports. The context window is
# nominal: generation takes a prompt of at most 32,000 characters.
- id: gpt-image-1
  provider: openai
  tags:
    - image-generation
  supportsTools: false
  supportsVision: true
  outputsMedia: true
  contextWindow: 32000
  pricing:
    inputCentsPerMillion: 500
    outputCentsPerMillion: 4000
    imageInputCentsPerMillion: 1000
- id: gpt-image-1-mini
  provider: openai
  tags:
    - image-generation
  supportsTools: false
  supportsVision: true
  outputsMedia: true
  contextWindow: 32000
  pricing:
    inputCentsPerMillion: 200
    outputCentsPerMillion: 800
    imageInputCentsPerMillion: 250
# Embedding model — the everyday knowledge-indexing workhorse. 1536 native
# dimensions, under pgvector's HNSW index limit.
- id: text-embedding-3-small
  provider: openai
  tags:
    - embedding
  supportsTools: false
  supportsVision: false
  contextWindow: 8191
  pricing:
    inputCentsPerMillion: 2
    outputCentsPerMillion: 0
  embedding:
    dimensions: 1536
    recommended: true