Update models

Bring src/models.ts and the per-provider defaults back in sync with what the providers actually offer today. This is an interactive procedure: research online, then propose a concrete diff and confirm the judgment calls (defaults, removals) with the user before editing.

Ground rules (read first)

• A wrong price is worse than undefined. Never guess a price. If you can't verify a number on the official page, leave that model out of the table (the cost helpers then return undefined, which is the intended safe outcome).
• Keep the table small. It is the most commonly used models per provider (current + recent generations + the cheap tiers), not an exhaustive catalog. Don't add every niche/specialized model.
• Confirm the judgment calls. Changing a DEFAULT_MODEL and removing a pricing entry are product decisions — surface a recommendation and use AskUserQuestion before doing either. Adding missing models and correcting a clearly-wrong price can be done directly (still report them).

Key files

• src/models.ts — the MODELS pricing table, PRICING_VERIFIED_ON, and the findModel resolver. Read the file header and the findModel doc-comment for the full pricing/resolution rules before editing.
• src/providers/anthropic.ts — const DEFAULT_MODEL (no embeddings).
• src/providers/openai.ts — const DEFAULT_MODEL + DEFAULT_EMBED_MODEL.
• src/providers/google.ts — const DEFAULT_MODEL + DEFAULT_EMBED_MODEL.
• src/providers/__tests__/<provider>.test.ts — the default-path tests assert the default model in the request URL/body and the result (model: field).
• README.md, CLAUDE.md, CHANGELOG.md — doc references to defaults.

Official sources

Prefer WebFetch on the official pages (exact ids + prices); use WebSearch to discover what's new and to find deprecation dates.

• Anthropic pricing: https://platform.claude.com/docs/en/about-claude/pricing
• Anthropic deprecations: https://platform.claude.com/docs/en/about-claude/model-deprecations
• OpenAI pricing: https://developers.openai.com/api/docs/pricing
• Google (Gemini) pricing: https://ai.google.dev/gemini-api/docs/pricing
• Google (Gemini) deprecations: https://ai.google.dev/gemini-api/docs/deprecations
• Also search for each provider's model-deprecation / retirement schedule.

The current month is later than the training cutoff — do not answer from memory, always fetch. Model families move fast (e.g. tier names can change: OpenAI's 5.6 line is Sol/Terra/Luna, not mini/nano).

Procedure

1. Snapshot the current state

Read the MODELS table and each provider's DEFAULT_MODEL/DEFAULT_EMBED_MODEL. Note PRICING_VERIFIED_ON. List what the library currently knows per provider.

2. Research each provider

For Anthropic, OpenAI, and Google gather, from the official pages:

• Current model ids exactly as used in the API (not marketing names).
• Input / output / cached-input price per 1M tokens; any tiered pricing (e.g. Gemini 2.5 Pro > 200k prompt tokens); cache-write rate (Anthropic).
• Which models are GA vs preview, and which are deprecated/retired with their shutdown dates.

3. Reconcile → propose a plan

Produce a concise diff plan, per provider:

• Add — current, common models missing from the table (respect "keep it small").
• Fix price — any table entry whose price ≠ the official page.
• Remove — retired/shut-down models (see step 6 for the nuance).
• Default — your recommended DEFAULT_MODEL and why.

Present this to the user before making the judgment-call edits.

4. Apply pricing-table edits (src/models.ts)

Follow the table's conventions (they are load-bearing — see the file header):

• USD per 1M tokens. Keys are grouped by provider with a short comment.
• Exact keys. The resolver matches k-<DIGIT>… only, so word-suffix siblings each need their own exact key (-mini, -nano, -pro, -sol, -terra, -luna, -flash-lite, …). A dated snapshot (…-2026-07-09) resolves to its base via the digit rule — don't add snapshot keys.
• cachedInputPerMTok only when the page publishes it; omit otherwise (the cost helpers fall back to the full input rate — never fabricate a discount).
• cacheWriteInputPerMTok — Anthropic only (1.25× input = the 5-minute TTL).
• highTier for tiered models (aboveInputTokens + input/output/cached rates).
• Embedding models: input-only, outputPerMTok: 0, full exact key.
• PRICING_VERIFIED_ON: bump to today only if you re-verified all three providers' prices this run. If you checked only one provider, leave it (a table-wide stamp must not over-claim).

5. Defaults — recommend + confirm

For each provider, recommend a DEFAULT_MODEL. Good default = GA (not preview), current generation, sensible price/capability for the provider's "workhorse" tier, and not scheduled to retire soon. A same-price newer-generation successor to the existing default is the ideal move. Use AskUserQuestion to let the user pick (offer your recommendation first, marked "Recommended").

When a default changes, update all of these together:

1. src/providers/<provider>.ts — DEFAULT_MODEL.
2. src/providers/__tests__/<provider>.test.ts — the default-path test's URL, request-body model, and result model assertions. Keep any explicit-model override test pointed at a different model so it still proves overrides work.
3. README.md config example (and any prose reference to the default).
4. CLAUDE.md — the provider's one-line entry.

6. Removals — retired / not recommended

The table exists for cost calculation, and usage logs reference models long after they stop being callable — so keeping a still-relevant entry is often correct even when a model can't be newly selected. Remove an entry only when it is fully retired/shut down (or the user wants the table trimmed). Removing loses cost calc for historical usage of that model — surface that tradeoff and confirm with the user (AskUserQuestion, per-model or as a batch) before deleting. Never remove a model that is still a DEFAULT_MODEL — repoint the default first (step 5). If a model is deprecated-but-active (e.g. 404s for new keys yet still billed for existing ones until a future shutdown date), prefer to keep it with a dated retirement comment rather than remove it.

7. Changelog

Add entries under ## [Unreleased] in CHANGELOG.md (Keep a Changelog format): ### Added for new models, ### Changed for default/price changes, ### Removed for deletions. Keep it short. Do not add a dated version heading (that's a release-time step).

8. Verify (edit loop)

Run as you go, then once at the end on the touched files:

yarn eslint <changed files>
yarn typecheck
yarn prettier --write <changed files>
yarn test google models cost openai anthropic   # the suites this touches

Optional resolver sanity check via the built bundle (confirms new keys resolve and word-suffix siblings don't mis-resolve):

yarn build
node -e 'import("./dist/index.js").then(({findModel})=>{for(const m of ["<new-id>","<new-id>-2026-01-01"])console.log(m,"->",findModel(m)?.id)})'

9. Report

Summarize per provider: models added, prices corrected, models removed, and the new default (with the reasoning). State whether PRICING_VERIFIED_ON was bumped and why/why not. Remind the user nothing was committed (unless they asked).