Problem#
The rate cards for the two newest Anthropic models have no cache_read_tokens rate. The other 10 Anthropic cards have one.
| Model | cache_write_tokens | cache_write_1h_tokens | cache_read_tokens |
|---|
claude-opus-5-5 | ✓ | ✓ | missing |
claude-fable-5-1 | ✓ | ✓ | missing |
Both cards come from platform.claude.com/docs/en/about-claude/pricing.md (derivation: docs-derived, hash cd36fcc6…, extracted 2026-09-23). Anthropic prices cache reads on every current model, so these cards cannot price a request that reads from the cache.
Impact#
TanStack AI writes pricing.input.cached from the card and uses it to add Anthropic's cache-control options type. Without a cache-read rate, the new models lose both.
Ask#
Fix the docs-derived Anthropic parse so it picks up the cache-read column for new rows. Add a fixture that fails when a card has cache writes but no cache read.
Acceptance#
claude-opus-5-5 and claude-fable-5-1 cards include cache_read_tokens.
POST /v1/estimate with cache_read_tokens returns a price for both.