Claude Sonnet 5.5 runs adaptive thinking on every request unless you say otherwise, defaults to high effort, and no longer accepts thinking disabled or budget_tokens. Teams porting latency sensitive Sonnet 5 code should switch to thinking between_tools, set effort explicitly, and take advantage of a prompt caching minimum that drops to 512 tokens.
Anthropic released Claude Sonnet 5.5 on September 28, and its documentation makes one change that will quietly slow down a lot of migrated code: requests without a thinking field now use adaptive thinking. On the Claude API the default effort is high. If your Sonnet 5 integration relied on fast, non reasoning responses for classification, routing, or extraction, simply swapping the model ID to claude-sonnet-5-5 will give you a model that thinks before it answers on every call.
The old off switch is gone. Anthropic's migration guide says to replace thinking type disabled with thinking type between_tools, which is now the lowest thinking setting. It works only at low, medium, and high effort, so it cannot be combined with xhigh or max. With between_tools, the model skips up front reasoning, which is what most agent loops and pipelines want when speed matters more than depth.
Thinking budgets are also retired. A request that sends thinking type enabled with budget_tokens returns a 400 error on Sonnet 5.5. Control depth through output_config effort instead, with five levels: low, medium, high, xhigh, and max. Anthropic says the effort levels were recalibrated relative to Sonnet 5, so a setting that worked before is not a reliable guide. Re run your effort sweep against your own evaluation set and re baseline cost before you lock anything in.
Two smaller details are worth knowing. Thinking text is omitted by default, so if your interface shows reasoning summaries you now need to set display to summarized. And the minimum cacheable prompt drops to 512 tokens, down from 1,024 on Sonnet 5, which means shorter system prompts and tool definitions can now be cached. Teams that gave up on caching small prompts should revisit it.
The practical pattern is simple. Decide per workload whether you want reasoning at all. For fast paths, send between_tools with effort low or medium. For hard problems, leave adaptive thinking on and choose the effort level deliberately rather than inheriting high by accident.
Try it today
- Change the model ID to claude-sonnet-5-5 and audit every request that previously sent thinking type disabled.
- Replace thinking disabled with thinking type between_tools, and pair it with effort low, medium, or high only.
- Remove any thinking budget_tokens settings, which now return a 400 error, and set output_config effort explicitly.
- Re run your effort sweep on your own evaluation set, since Anthropic recalibrated the levels from Sonnet 5.
- If you display reasoning, set display to summarized, and re enable prompt caching for prompts of 512 tokens or more.