# We've launched Claude Haiku 5.5 (claude-haiku-5-5)

By Steven Van · 2026-10-07

The model has a 1M token context window, up to 128,000 output tokens and configurable thinking effort.

A 1M token context window and up to 128,000 output tokens come with Claude Haiku 5.5, [Anthropic](<https://creatorstoolbox.com/tools/anthropic>)'s new model for high-volume and latency-sensitive work. The release adds adaptive thinking and makes Haiku's thinking effort configurable for the first time.

## Effort controls thinking

Medium effort is the default in the Claude API and Claude Code. Low is the cheapest and fastest setting, intended for chat, short tool tasks and simple requests. High suits knowledge work, longer agent tasks and strict instruction following. Anthropic recommends testing xhigh and max against evaluations to determine whether the quality gains justify their cost and longer responses.

Thinking is on by default and counts toward the output token limit, so requests need room for both thinking and the visible reply. Developers can disable thinking at low, medium and high effort, but doing so at xhigh or max returns a 400 error.

## Migration requires API changes

Existing Haiku 4.5 prompts should perform well without changes, but application code can break. Manual thinking budgets using budget\_tokens now return a 400 error, responses can begin with thinking blocks, and the same text counts as more tokens. Output limits sized for Haiku 4.5 requests without thinking may cut replies off.

## Guidance for tools and agents

Anthropic's [prompting guide](<https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-haiku-5-5>) recommends adaptive thinking when combining JSON output with custom tools, because disabling thinking can cause the model to skip a needed tool call. It also recommends supplying today's date for search tasks and explicitly instructing coding agents to verify changes before reporting completion.

Long agent prompts at low effort can lead to early stopping. Raising effort or adding explicit completion instructions can reduce this, at a higher token cost. Applications should deliver mid-task user input as user text outside tool results and handle responses with stop\_reason set to "refusal", since the model has no server-side fallback.

## Where it is available

Claude Haiku 5.5 is available on the Claude API, Claude in Amazon Bedrock, Claude Platform on AWS, Claude on Google Cloud and Claude in Microsoft Foundry.

![Anthropic](<https://www.google.com/s2/favicons?domain=anthropic.com&sz=128>)

Anthropic

Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.

[View Anthropic →](<https://creatorstoolbox.com/tools/anthropic>)

[Original source](<https://platform.claude.com/docs/en/release-notes/overview#october-7-2026>)

[Read We've launched Claude Haiku 5.5 (claude-haiku-5-5) on Creators Toolbox](<https://creatorstoolbox.com/blog/anthropic-we-ve-launched-claude-haiku-5-5-claude-haiku-5-5>)

---
Canonical source: https://creatorstoolbox.com/blog/anthropic-we-ve-launched-claude-haiku-5-5-claude-haiku-5-5