AnthropicReleases

Illustration: a single blank keyboard key in 3/4 view, outlined in light grey on black.
Illustration

Haiku 5.5 is out, and Haiku 4.5 code can return 400s

Before you swap the model ID, strip budget_tokens, temperature, top_p, top_k and prefill, or requests fail.

You can move high-volume work to claude-haiku-5-5 today, but check your request code first. Anthropic says code written for Haiku 4.5 "can break" on Haiku 5.5 with 400 errors.

What changed#

Anthropic launched Claude Haiku 5.5 on Oct 7, 2026. It has a 1M token context window, 128k max output tokens, and adaptive thinking with the effort parameter. It runs on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry.

Before (Haiku 4.5): manual thinking budgets, sampling parameters and prefill worked. After (Haiku 5.5): each of these can return a 400 error, and adaptive thinking is on by default.

Why it matters#

Haiku is the model people put in cheap, high-volume pipelines. Those pipelines often set temperature or prefill a JSON opener. A blind model swap breaks them.

How to use it#

  1. Remove thinking: {type: "enabled", budget_tokens: N}. Control depth with effort instead.
  2. Remove temperature, top_p and top_k.
  3. Remove assistant prefill. Use structured outputs for JSON (not on Bedrock for this model).
  4. Computer use: switch computer_20250124 to computer_toolset_20260801 (Claude API, Google Cloud) or computer_20251124 (Bedrock).
  5. Expect responses to start with thinking blocks. Set thinking.display: "summarized" if you want their text.
  6. Don't change system, tools or earlier turns and then replay old thinking blocks; that can also return a 400.
Search this repo for every Anthropic Messages API call that uses claude-haiku-4-5. For each, list file and line, and flag any use of budget_tokens, temperature, top_p, top_k, assistant prefill, or computer_20250124. Propose the Haiku 5.5 replacement for each.

Good to know#

  • Thinking text "counts as more tokens," per the release notes, so watch cost on the first day.
  • Claude Code users: an open GitHub issue (one reporter, Oct 9) says --thinking disabled sends no thinking field for Haiku 5.5, so the model thinks anyway. Not confirmed by Anthropic.
  • Watch next: the release notes and the Claude Code changelog for a fix to that flag.

Sources#

Last checked: 2026-10-12

Sources

  1. docs.anthropic.com
  2. anthropic.com
  3. github.com

Something wrong? Tell the bot