Claude Haiku 5.5 is out, and Haiku 4.5 code can break on it

what changed

On 7 October, Anthropic launched Claude Haiku 5.5, tuned for high-volume and latency-sensitive work. It has a 1M token context window, 128k max output tokens, and adaptive thinking with the effort parameter. It is available on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry.

The release notes also list what breaks when you move from Claude Haiku 4.5. Manual extended thinking (budget_tokens), the sampling parameters temperature, top_p and top_k, and assistant message prefill can each return a 400 error. So can the older computer use tool version. On Amazon Bedrock, structured outputs are not available for this model.

why it matters

Haiku is the model teams put on the hot path: classification, routing, extraction, sub-agents. Those are often the oldest prompts in a codebase, and the ones most likely to set a temperature or prefill the assistant turn. A model swap there is not a config change.

Adaptive thinking is on by default, so a response can start with thinking blocks, and that text counts as extra tokens. Cost per call can move even when the price per token does not.

what we would do

Search your code for temperature, top_p, budget_tokens and assistant prefill before you switch. If you run on Bedrock and rely on structured outputs, keep that path on the model you use today until that changes. Then run your evals on both models and compare cost per task, not cost per token. The migration guide lists each change.