Migrating: Sonnet 4.6 to Sonnet 5.5

Two generations in one move

The migration guide points Sonnet 4.6 code at Sonnet 5.5 by way of Sonnet 5's changes. Most of the cost story comes from the first step; most of the error story from the second. Doing both at once saves a round of testing, provided you test for both.

The cost story: cheaper per token, more tokens

Sonnet 5.5 costs less per token than Sonnet 4.6, but it uses the newer tokenizer, which counts noticeably more tokens for the same text. The saving is real only where the rate cut outruns the extra tokens, and that depends on your content — code and non-English text behave differently from prose. Recount real prompts with count_tokens against both ids before quoting a number. The newer model's lower cache minimum pushes the other way, letting short system prompts cache.

The request story: what stops being accepted

Sonnet 4.6 still accepted non-default temperature, top_p and top_k, and the deprecated budget_tokens; all of them return a 400 on Sonnet 5.5. So do thinking: disabled — the lowest setting is now between_tools — and forced tool choice. Omitting thinking flips from no thinking to adaptive thinking, so max_tokens needs room for reasoning it never had to hold before.

The tuning story: effort means something new

Sonnet 4.6 tops out below xhigh; Sonnet 5.5 has the full range, recalibrated relative even to Sonnet 5, so no mapping from your old setting transfers. Start from Anthropic's suggestions — medium for agentic coding, low for chat and extraction — and let an evaluation set decide.

Related

See what actually changed from Sonnet 4.6 to Sonnet 5, then from Sonnet 5 to Sonnet 5.5.