claude-haiku-5-5 — pricing, context window and limits

The cheapest model per token — until the prompt gets long

Haiku 5.5 is the current Haiku, and its base rates are the lowest in the lineup. It is also the first Claude model priced by prompt length: once a request's prompt passes the threshold in the price table above, every token in that request — input, output and cache — moves to the second rate card. The prompt length counts all of the request's input, cache reads and writes included, so a long, heavily cached system prompt can push a request over the line even when most of it is a cache hit. Each request is priced on its own, which means the question to ask of a workload is not "how many tokens a month" but "how long is each prompt".

Sonnet-class limits on a Haiku

The context window and output cap match the newest Sonnet and Opus models, a large step up from Haiku 4.5. The minimum cacheable prefix dropped to the lowest in the lineup, shared with the newest Opus, Sonnet and Fable models, so prompts that were too short to cache on Haiku 4.5 now create cache entries with no code change.

Thinking is on unless you lower it

Adaptive thinking runs by default, and effort arrives on the Haiku line for the first time, with a default of medium. A response can therefore begin with thinking blocks, and a small max_tokens sized for a one-word answer can be spent on thinking before any text appears. Lowering effort is the documented lever; thinking: {type: "disabled"} is still accepted, but only at effort high or below. A forced tool_choice is accepted, and the response then skips thinking.

What returns a 400 that Haiku 4.5 accepted

budget_tokens, non-default temperature/top_p/top_k, and assistant prefill are all rejected. Computer use on the Claude API and Google Cloud needs the computer toolset. Editing earlier turns invalidates replayed thinking blocks, so a harness that rewrites history has to become append-only. The model can also decline a request with a refusal stop reason, and unlike the larger models there is no server-side fallback to retry it.

Count again before you budget

Its tokenizer counts noticeably more tokens than Haiku 4.5's for the same text, so recount prompts with count_tokens against this model rather than reusing old figures.

Related

See migrating: Haiku 4.5 to Haiku 5.5, Sonnet 5.5 vs Haiku 5.5 and the API pricing table, which lists both rate cards.