Anthropic Claude Sonnet 5 rejects legacy thinking and sampling parameters

What changed

Anthropic launched Claude Sonnet 5 (claude-sonnet-5) as a drop-in upgrade path from Claude Sonnet 4.6 with three request-behavior changes:

1. Adaptive thinking is on by default. Requests without a thinking field run with adaptive thinking (unlike Sonnet 4.6, where omitted thinking meant no thinking). Disable with thinking: {type: "disabled"}.
2. Manual extended thinking is removed. thinking: {type: "enabled", budget_tokens: N} returns HTTP 400.
3. Non-default sampling parameters return HTTP 400. Setting temperature, top_p, or top_k to a non-default value is rejected; omit them or use defaults and guide behavior via prompting.

Sonnet 5 also uses a new tokenizer that produces approximately 30% more tokens for the same text (exact increase depends on content). Per-token pricing is lower than Sonnet 4.6 ($2/$10 vs $3/$15 per million tokens), but equivalent-request cost does not drop in direct proportion because of the tokenizer change.

Who is affected

Teams migrating workloads to claude-sonnet-5 from Sonnet 4.6 (or building new integrations against Sonnet 5) that still:

• omit thinking while assuming thinking stays off,
• send manual budget_tokens extended thinking, or
• set non-default temperature / top_p / top_k.

Also affected: cost/token budgeting systems that reuse Sonnet 4.6 token counts without recounting.

What you need to do

Recommended review when adopting Sonnet 5:

1. Change the model ID to claude-sonnet-5.
2. Explicitly set thinking: {type: "disabled"} if you need thinking off; otherwise migrate manual extended thinking to adaptive thinking / effort controls.
3. Remove non-default sampling parameters from requests.
4. Recount prompts with Anthropic’s token counting guidance and revisit max_tokens budgets sized against Sonnet 4.6.

There is no forced retirement date in this notice for Sonnet 4.6; this brief covers migration-time breaking behavior, not a sunset deadline.

Migration effort

Likely medium for production agents that customize thinking budgets or sampling. Likely lower if requests already omit sampling params and do not rely on manual extended thinking—then the main work is model ID swap plus token-budget validation.

Cost impact

Mixed. Official per-token prices are lower than Sonnet 4.6 ($2/$10 vs $3/$15 per million tokens), while the new tokenizer produces approximately 30% more tokens for the same text (workload-dependent). That ~30% figure is a token-count statement, not a universal cost percentage. Net spend direction depends on workload shape; Anthropic explicitly warns equivalent-request cost does not drop in direct proportion to the per-token price cut.