price-check

price-check

Keeps backend/core/providers/pricing.py (MODEL_PRICING + PRICING_SOURCES) current with published per-token rates and regenerates the reviewable report backend/core/providers/PRICING.md. No provider's models API exposes price, so this skill is the maintenance mechanism that keeps the usage/cost dashboard accurate as new models are added dynamically.

When to run

• During any PR that adds/changes models or touches core/providers/ or core/agents/usage/ (required by the project CLAUDE.md → Model Pricing).
• On request: "check prices", "refresh model pricing", "is pricing current?".

Guard 1 — ship only in the Chatty repo

Before writing or committing anything, confirm this is the Chatty repo:

• git remote -v contains wwilson1017/chatty, OR
• CLAUDE.md starts with # Chatty AND backend/core/providers/pricing.py exists.

If neither holds, STOP: report that price-check only runs in the Chatty repo, and make no changes.

Guard 2 — skip if already verified this chat

If model pricing was already fetched/verified earlier in THIS conversation, or backend/core/providers/PRICING.md already shows Last verified: <today> with no pending diff, do NOT re-fetch. Reuse the in-session results and report "already current". Re-fetching wastes calls and risks rate limits.

Procedure

1. Model set. Union of the keys in MODEL_PRICING and each provider's current model list — the fallback *_MODELS constants, plus (if credentials are configured) what GET /api/providers/{provider}/models returns. This prices newly-surfaced dynamic models, not just legacy ones.
2. Fetch current STANDARD-tier rates (USD per 1M input / output tokens) via WebFetch from the official pages:
   • Anthropic — https://platform.claude.com/docs/en/about-claude/models/overview
   • OpenAI — https://developers.openai.com/api/docs/pricing
   • Google — https://ai.google.dev/gemini-api/docs/pricing
   • Together — https://www.together.ai/pricing Record tier caveats (e.g. Gemini 2.5 Pro is the ≤200K-prompt tier) in a code comment.
3. Diff against the current MODEL_PRICING; classify each model: added / changed / unchanged / unverifiable.
4. Never fabricate. If a model's rate isn't on the official page, leave it OUT of MODEL_PRICING and report it as "unknown" — the usage dashboard already flags unpriced paid models (is_priced / per-agent has_unknown_pricing). Do NOT guess from memory or community posts.
5. Write backend/core/providers/pricing.py:
   • MODEL_PRICING[model] = (input, output)
   • PRICING_SOURCES[model] = (source_url, "<YYYY-MM-DD>") — use the current session date. Keep proven legacy entries (the dashboard prices historical rows by model id via longest-prefix match).
6. Regenerate backend/core/providers/PRICING.md from the updated pricing.py (do not hand-author divergent numbers): a header line Last verified: <YYYY-MM-DD> (price-check) followed by a per-provider table model | input $/M | output $/M | source | verified.
7. Report a concise diff table to the user. When run during a PR, also place the diff in the PR description.

Commit

• Commit pricing.py and PRICING.md together, e.g. Refresh model pricing (price-check <YYYY-MM-DD>).
• No Co-Authored-By or "Generated with Claude Code" footers (repo policy).
• Never commit secrets; never run gcloud secrets/--set-secrets commands.

Notes

• The session date is authoritative for the verified date — read it from the environment context, don't compute it in code.
• This skill is committed into the Chatty repo so it travels with the project; it is the canonical place to evolve the pricing-refresh workflow.