Summary
Per-million-token (USD) pricing for the Claude model families, measured 2026-10-08 from the Anthropic platform pricing docs. The page covers four model lines — Opus 5.5, Sonnet 5.5, Fable 5.1, and the Opus 5 family (5 / 4.8 / 4.7) — across input, output, and prompt-cache write/read rates, plus the Sonnet 5.5 breaking changes and related release notes for the yoagent library.
Key Points
- Cache write multipliers are uniform: 5-minute cache writes cost 1.25× input; 1-hour cache writes cost 2× input for every model listed.
- The 5.5 line changes cache-hit pricing: Opus 5.5 and Sonnet 5.5 bill cache hits at 0.05× input, not the usual 0.1×.
- Opus 5.5 is cheaper than Opus 5 ($4/$20 vs $5/$25 in/out), and its 5m cache write ($5) is lower than Opus 5's input price.
- Fable 5.1 is the most expensive tier ($10 in / $50 out), with the cheapest absolute cache-hit read ($0.25) but a very low multiplier (~0.025× input).
- Sonnet 5.5 launched 2026-09-28 with five breaking changes versus Sonnet 5.
- The
yoagentlibrary made cost opt-in and removed default pricing in 0.25.0.
Concepts
- Prompt caching (cache write vs. cache hit pricing)
- Input / output token rates
- Forced
tool_choicesemantics - Extended thinking / effort levels
- Computer-use tool versions (
computer_20251124) ModelConfig::cost,enable_bundled()/enable_live()- Rate-limited retry mapping (
529→RateLimited)