Claude API pricing for every model, newest first

Every row is Anthropic's own price where Anthropic sells the model, in dollars per million tokens. Claude's prompt cache works a little differently from most. You mark where the reusable part of the prompt ends, the first call writes it to the cache, and later calls read it back.

On Claude Fable 5.2 a cache write costs $10 per million, above the normal input rate, and a read costs $0.20, 98% below it. So caching only pays once a prefix gets read back a few times. An agent loop or a long chat does that on every turn, and a one-off question never does. A cached prefix stays warm for five minutes by default, and every read resets that clock.

The Batch API takes 50% off input and 50% off output on Claude Fable 5.2, for $4 and $20 per million. Most batches finish within an hour, so it suits evals, backfills and nightly reports, just not a chat window.

Claude Haiku 5.5 has a second rate for long prompts. Past 100,000 prompt tokens it bills $0.50 input and $2.50 output per million, so check where your longest prompts land before you route them all to the small model. Claude Fable 5.2 reads up to 1M tokens, as much as any Claude model.

Opus 5.5 also has a fast mode that returns tokens much faster at a higher rate. The table shows standard rates only.

Don't pick one model for everything. Send volume work like summaries, routing and subagents to Claude Haiku 5.5, and keep Claude Fable 5.2 for the jobs where a wrong answer costs you an afternoon. Test the split on your own prompts and track the cost per accepted answer, since a retry pays for the whole call again.

Keep the cached prefix byte for byte the same between calls, and send anything that can wait through the Batch API. A Claude Pro or Max plan doesn't cover API calls, so budget for the API on its own.