The model has a 1 million-token context window, a 128,000-token maximum output, and adaptive thinking controlled through an effort parameter. It is available through the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry.
The notes flag migration changes for Haiku 4.5 users. Manual extended thinking via budget_tokens, temperature, top_p, top_k, and assistant prefill can return 400 errors. The computer_20250124 tool also returns 400; Anthropic directs API and Google Cloud users to computer_toolset_20260801, while Bedrock users should use computer_20251124.
Python and TypeScript SDKs now include beta classes for browser-use and computer-use tools. Developers subclass a class and implement a method per tool; the SDK runs the tool loop and applies URL and file policies plus an approval callback. The release notes do not establish Haiku 5.5 pricing, benchmark performance, or task-quality comparisons.