Anthropic's Claude Opus 5.5 is now generally available in Microsoft Foundry, hosted on Azure, giving developers and enterprises access to Anthropic's most capable Opus model directly inside Microsoft's AI platform.
The model is built for work that extends well past a single prompt: building features across a codebase, tracing production issues, synthesizing long documents and spreadsheets, and producing memos, reports or presentations that require less editing than outputs from Opus 5. It finds root causes before making changes, checks its own work, and explains changes in plain language.
A notable addition is improved agentic communication. As an agent works through dozens of steps, Claude Opus 5.5 surfaces what it did, what it found, what decisions it made, and where it needs input from the user, along with a summary at the end of a long-running task - intended to reduce the need to inspect every step manually.
The model also introduces adaptive thinking, letting developers specify an "effort" level rather than manually setting a thinking-token budget, so the model scales its reasoning to the task automatically. Alongside the model, Anthropic is shipping beta API capabilities - asynchronous compaction, keep-tail compaction, and the ability to change tools mid-conversation while preserving thinking and prompt caching - aimed at agents that run for extended periods without needing to rebuild application state.
Pricing for Claude Opus 5.5 (Global Standard, US DataZone) is $4 per million input tokens and $20 per million output tokens, with cache hits at $0.20 per million tokens, cache writes at $5 per million, and one-hour cache writes at $8 per million. The same pricing applies to a Long Context variant. Both are generally available now, hosted on Azure.
Anthropic is also expanding safeguards for the model: Claude Opus 5.5 is the first Opus model to use safety classifiers in areas including cybersecurity, biology, AI development, and distillation, first introduced with Claude Fable 5.1. Common developer, educational, and knowledge-work requests remain unaffected, but certain higher-risk or dual-use requests may be routed to a different Claude model with additional safeguards.
