Accurate Anthropic Messages prompt caching pricing hotfix¶

An urgent hotfix resolves pricing discrepancies for POST /v1/messages using Anthropic prompt caching, ensuring complete consistency between API response usage metrics, public catalog pricing rules, and ledger transaction records.

Fixed issues¶

1-hour and 5-minute cache pricing tiers¶

Corrected pricing calculations for explicit 1-hour cache write tokens (cache_creation_1h_input_tokens), which previously applied an incorrect base rate.
Enforced strict rate separation between standard 5-minute ephemeral cache creation (cache_creation_input_tokens) and 1-hour persistent cache blocks.

Mixed token requests¶

Fixed partial usage accounting loss in complex requests combining standard input tokens, cache read tokens, and new cache write blocks simultaneously.
Ledger debit records in balance history now use the exact same normalized pricing formula as returned in the API response.

Catalog and debit synchronization¶

Resolved discrepancies between cache creation rates displayed in the public catalog and actual transaction ledger debits.

Improvements¶

Transaction history transparency¶

Detail views for Claude model requests in the History section now clearly break down prompt caching token metrics and exact billing structures.

No changes to client SDKs, request formats, or Anthropic caching headers are needed.
Standard Anthropic prompt caching parameters work as documented in the API reference.