Sonnet 5.5 vs Haiku 5.5

The current pair for high-volume work

These are the newest models in their lines, and the choice between them now turns on different things than it did with Haiku 4.5. Both have the same large context window and output cap, the same minimum cacheable prefix and the full effort ladder, so the old reasons to step up to Sonnet — room for a big document set, prompts too short to cache on Haiku — no longer apply. What remains is price, reasoning depth and a handful of request-shape differences.

Haiku is far cheaper, until its prompts get long

On the base rates in the table above, Haiku 5.5 costs a small fraction of Sonnet 5.5 per token. But Haiku 5.5 is priced by prompt length and Sonnet 5.5 is not: once a Haiku request's prompt passes the threshold, the whole request bills at the second rate card, while Sonnet charges the same at any length. Even past the threshold Haiku stays cheaper per token, so the real question is quality, not price — but a workload that sends very long prompts should budget on the higher Haiku card, not the headline one. Count real prompts with count_tokens against each model before comparing invoices, rather than reusing a count taken on another model.

Defaults that differ

Both think by default. Haiku 5.5 defaults to medium effort, Sonnet 5.5 to high, so an unset effort field is not a like-for-like comparison — set it explicitly on both when you test. Turning thinking off works differently too: Haiku 5.5 still accepts thinking: {type: "disabled"} at effort high or below, while Sonnet 5.5 rejects it and offers between_tools instead. Sonnet 5.5 rejects a forced tool_choice; Haiku 5.5 accepts one and skips thinking for that response.

Refusals and fallbacks

Both can decline a request with a refusal stop reason. Sonnet 5.5 supports a server-side fallback on the Claude API; Haiku 5.5 has none, so a declined Haiku request has to be handled in your own code.

Related

See Claude Haiku 5.5, Claude Sonnet 5.5 and the model picker, which prices both at the context size you enter.