Changelog

2026-09-22 — Claude Opus 5.5

Added model ID claude-opus-5-5 across Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry.
Added 1M-token context and 128K maximum output.
Added adaptive thinking as an always-on capability with effort control; the default effort is medium.
Reduced API pricing to $4 per million input tokens and $20 per million output tokens compared with Opus 5.
Added support for task budgets, mid-conversation system messages, prompt caching, Batch, Files, PDF, vision, server-side/client-side tools, and on-demand compaction.
Updated computer-use requirements on Claude API and Google Cloud: use computer_toolset_20260801 instead of the older computer_20251124 tool.
Added migration notes for forced tool choice, thinking blocks, streaming text between tool calls, and effort calibration.

For model releases, check the Claude model selection guide first. For Opus 5.5 upgrades, follow the migration guide and verify the official release notes.