最新模型
Claude Opus 5.5 lowers token prices and introduces multiple breaking changes to the agent API
The new version reduces input, output, and cache read prices and makes adaptive thinking always active. Forced tool calls, edits to conversation history, and streamed progress displays all require renewed validation.
Anthropic launched Claude Opus 5.5 on September 22, reducing prices per million input and output tokens to $4 and $20, respectively, with cache reads priced at $0.20. The company says typical tasks cost about 40% less than with Opus 5 under default settings. That figure reflects changes in both pricing and usage and should not be treated as a uniform discount across all workloads. [Official announcement](https://www.anthropic.com/claude-opus-5-5)
For agent developers, the more immediate change is request compatibility. Adaptive thinking is always enabled in the new version; disabling thinking or specifying a legacy manual thinking budget returns an error. Forcing the model to call any tool or a specific tool is also no longer supported. Existing workflows that rely on forced tool calls to produce structured data must switch to structured outputs or strict tool schema validation and verify that a tool was actually called. Strict schema validation does not guarantee a tool call; workflows that depended on that guarantee still need application-level checks and retry handling. Default reasoning effort has also changed from high in the previous generation to medium, so upgrade comparisons must explicitly hold settings constant. [API changes](https://platform.claude.com/docs/en/models/opus-5-5/whats-new-opus-5-5)
Multi-turn conversations face additional state constraints: when previous thinking blocks are sent back, the system prompt, tool set, and messages preceding each block must pass prefix validation. This check is enabled by default for accounts created on or after August 31; older accounts enable it by explicitly setting the relevant option. Rewriting earlier tool results or manually summarizing conversation history may invalidate subsequent thinking blocks. A beta control option can discard invalid blocks so the request can proceed, but the model loses that reasoning, and the cache is rebuilt from the point of modification. [Preserved thinking documentation](https://platform.claude.com/docs/en/build-with-claude/preserved-thinking)
The streaming interface also requires regression testing. Progress text between tool calls now appears in thinking blocks, and the default display settings omit that text. An application may report no errors while its interface stops showing progress. Engineering teams should parse responses by block type and verify display options instead of assuming the first block contains text. [Migration notes](https://platform.claude.com/docs/en/models/opus-5-5/whats-new-opus-5-5)
On performance, the model scores 58 on the Artificial Analysis Intelligence Index, but testing used the highest reasoning effort and default fallback settings, which differ from the settings underlying the official default-cost figures. This result alone does not establish performance on Traditional Chinese tasks, within a fixed latency budget, or with fallback disabled. During adoption, teams should measure task success rates, total token usage, cache hits, and fallback activations together to determine whether lower prices translate into a lower cost per completed task. [Independent evaluation](https://artificialanalysis.ai/models/claude-opus-5-5)