--- name: price-check description: "Look up what a generation actually costs right now from the provider live pricing page, record it with its source, and quote a run before spending anything. Triggers on \"/price-check\", \"what will this cost\", \"quote the month\", \"price this video\", \"compare providers\", \"is sound extra\", \"cost estimate\", or before generate-all and any paid generation. Prices expire after 24h; quotes refuse rather than guess, and a quote is never an approval." argument-hint: "[--model ] [--provider ] [--compare] [--batch] [--staleness]" effort: low user-invocable: true --- # /socialforge:price-check — What this run will cost, looked up today Prices here are never remembered. They are looked up, recorded with the URL they came from, and expire after 24 hours. That is not caution for its own sake. On 2026-07-31 a major video model shipped one day after this plugin's model registry was last reviewed, and the registry never knew. In the same week, a provider already wired into SocialForge was selling video at **$0.01/second** while the old cost table assumed **$0.40**. Any price baked into a plugin is wrong on a timetable nobody controls. ## The rule **Never state a cost you did not look up in this session or the last 24 hours.** Not from this file. Not from the model registry. Not from what you remember a model costing. If `price_book.py` says a price is `unknown` or `stale`, that is the answer — go and read the provider's page, then record what you read. ## Looking a price up 1. Find where to look: ```bash python3 "${CLAUDE_PLUGIN_ROOT}/scripts/price_book.py" --action providers ``` That returns each provider's pricing URL and its auth shape. Two things it will tell you that save time: - **Auth differs per provider.** `Authorization: Bearer …` for WaveSpeed and Kie, `Authorization: Key …` for fal.ai, application-default credentials for Vertex. Getting this wrong is the most common reason a first call fails. - **Kie AI blocks automated fetches** (measured: HTTP 403). Do not report that as "no price found" — ask the user for the rate instead, and record it with the pricing page as the source. 2. Fetch the pricing page with your own web tools and read the actual number. 3. Record it, with the URL you read it from: ```bash python3 "${CLAUDE_PLUGIN_ROOT}/scripts/price_book.py" --action record \ --model "kling-3.0-std" --provider wavespeed \ --unit second --price 0.084 --source https://wavespeed.ai/pricing ``` A record without a source URL is rejected. That is deliberate — a price with no provenance is a guess wearing a number's clothes. ## Quoting before you spend Single item: ```bash python3 "${CLAUDE_PLUGIN_ROOT}/scripts/price_book.py" --action quote \ --model "kling-3.0-std" --provider wavespeed --units 55 ``` A whole planned run — **required before `/socialforge:generate-all`**, which fans out across an entire monthly calendar and is the most expensive command in the product: ```bash python3 "${CLAUDE_PLUGIN_ROOT}/scripts/price_book.py" --action quote-batch --items '[ {"model":"kling-3.0-std","provider":"wavespeed","units":5,"label":"post-01"}, {"model":"seedance-2.0-fast","provider":"wavespeed","units":6,"label":"post-02"} ]' ``` Both exit non-zero when they cannot price the work, so a caller can never mistake a refusal for a costed run. **One unpriced item blocks the whole batch.** A total that quietly omits the three clips nobody could price is worse than no total, because it reads as complete. ## Options that change the price A base rate is not always the whole bill. On at least one wired video model, asking for synchronised audio bills at **1.5× the base per-second rate** — and SocialForge passes `sound` straight through to that API, so a quote taken without it understates a 10-second clip by about half a dollar and a 28-post month by roughly fifteen. Look the surcharge up on the model's page, then pass it: ```bash python3 "${CLAUDE_PLUGIN_ROOT}/scripts/price_book.py" --action quote \ --model "kling-v3.0-pro" --provider wavespeed --units 10 \ --multiplier 1.5 --multiplier-reason "sound=true" ``` In a batch, put it on the item: `{"model": …, "units": 10, "multiplier": 1.5, "multiplier_reason": "sound=true"}`. The quote then shows base, multiplier and reason separately, so the user can see what the option cost them. Multipliers are provider facts and move — look them up, do not carry them in your head. ## Then stop A quote is not consent. `approved_to_run` is always `false`, even for a clean batch. Show the user the number and wait for an explicit go. **One approval covers one run.** If the user approves a batch and then changes the calendar, re-quote and ask again. ## Comparing providers before you route The same model genuinely costs different amounts in different places — Seedance 2.0 Fast has been seen at $0.10/s on one provider and $0.24/s on another. That is a 2.4× difference for identical output. ```bash python3 "${CLAUDE_PLUGIN_ROOT}/scripts/price_book.py" --action compare --model "seedance-2.0-fast" ``` Cheapest first, grouped by unit — a per-image price is never "cheaper" than a per-second one, so they are never ranked against each other. If only one provider is on record, the output says so: one price is not a comparison. Look up a second before telling the user something is the cheapest option. ## Checking what has gone stale ```bash python3 "${CLAUDE_PLUGIN_ROOT}/scripts/price_book.py" --action staleness ``` Run this at the start of a production month. Anything listed as stale needs a fresh lookup before it can appear in a quote. ## Draft cheap, finish properly Iterating a rejected concept at hero-model rates is the quietest way to waste a client's budget. Generate drafts on the cheapest capable model, and only re-run the concept the user actually picked on the expensive one. Route on a price you looked up, not on a tier name — tier labels in the registry describe capability, not cost, and the two do not track each other reliably. ## What this skill will not do - Quote from memory, from the registry, or from this file - Present a partial total as the cost of a run - Treat a quote as approval - Call one provider cheapest on the strength of a single price - Report a fetch failure as an absence of cost