Source · docs
Claude API pricing
Summary
Anthropic's pricing reference for every Claude model and paid feature, read on
2026-10-08 after the Haiku 5.5 launch. Besides adding Haiku 5.5, it shows cache reads on
Opus 5.5 and Sonnet 5.5 at 5% of the base input price instead of the usual 10%. The page
has no date of its own; date_published is the Haiku 5.5 launch day, when it was last
known to change.
Key points
- Cache read multipliers: 0.1x base input on most models, 0.05x on Opus 5.5 and Sonnet 5.5 ($0.20 and $0.10 per 1M), 0.025x on Fable 5.1 and Mythos 5.1 ($0.25 per 1M). Cache writes stay at 1.25x (5 minutes) and 2x (1 hour).
- Simon Willison reports that the Sonnet 5.5 cache read price was halved alongside the Haiku 5.5 launch; the docs do not give the date of the change.
- Haiku 5.5 is the only current model with long-context pricing: over 100K prompt tokens everything costs 5x. Other Claude 4.6+ models bill the full 1M window at one rate.
- Sonnet 5's $2 / $10 introductory price is now permanent; the planned rise to $3 / $15 on 2026-09-01 did not happen.
- Batch API: 50% off input and output. US-only inference (
inference_geo: "us"): 1.1x. - Fast mode (research preview, Claude API only): Opus 5.5 at $8 / $40 per 1M.
- Managed Agents: tokens at model rates plus $0.08 per session-hour of running time.
- Web search $10 per 1,000 searches; web fetch has no extra charge; code execution is free when used with the 2026-02-09 or later web search or fetch tools.
Why it matters
Cheaper cache reads on the two main workhorse models make long, cached agent sessions noticeably cheaper. Haiku 5.5's tiered price breaks the simple "one price for the whole window" rule that held for recent Claude models.