Migrating: Sonnet 5 to Sonnet 5.5

Free on the rate card, not free in the code

Sonnet 5.5 kept Sonnet 5's per-token prices and tokenizer, so this move changes nothing on an invoice for identical traffic — and may lower it, because the minimum cacheable prefix dropped and shorter prompts now cache. Everything that costs effort is in the request code, where several changes return a 400 on code that worked yesterday.

What breaks

The off switch. thinking: disabled is rejected. Try adaptive thinking at low effort first; where a route genuinely must stay thinking-off, send between_tools at effort high or below, with no other field in the thinking object. Any client-side retry that re-sends the same body to another model must drop between_tools, because no other model accepts it. Forced tool choice is rejected, on token counting as well as messages. Thinking blocks belong to Sonnet 5.5 and to the conversation that produced them — no other model can read one. Computer use needs the toolset on the Claude API and Google Cloud. The advisor tool rejects several older advisor models — Opus 4.8, Opus 4.7 and Sonnet 5 among them.

What changes without breaking

The effort levels were recalibrated. The default is still high, so nothing errors, but a level name no longer means the amount of thinking it meant on Sonnet 5 — carry the old value over and you are running a different configuration under a familiar label. Anthropic's starting points are medium for agentic coding and low for chat. Text the model writes between tool calls also moves into thinking blocks, which are empty by default, so an interface that showed progress notes can go quiet until it opts into them.

Related

See what actually changed from Sonnet 5 to Sonnet 5.5 and between_tools is not supported for this model.