6 hours ago
- Claude Opus 5 is a significant improvement over Opus 4.8, with gains in deep reasoning, agentic tasks, long-horizon tasks, and test-time compute scaling.
- Thinking is enabled by default on Claude Opus 5; disabling thinking is restricted to effort levels high or below, and attempting to disable it with xhigh or max effort returns a 400 error.
- The model supports a 1M token context window, 128k max output tokens, and an effort ladder from low to max for controlling reasoning depth.
- Mid-conversation tool changes are in beta, allowing tools to be added or removed between turns while preserving the prompt cache.
- A new 'default' mode for fallbacks applies Anthropic's recommended fallback models by refusal category, requiring a beta header.
- The minimum cacheable prompt length is reduced from 1,024 to 512 tokens, enabling caching for shorter prompts without code changes.
- Fast mode is in research preview for Claude Opus 5 on the API only, priced at $10 per million input tokens and $50 per million output tokens.
- In agentic sessions, the model narrates progress more often, delegates to subagents more readily, and verifies its own work, so verification instructions from earlier models cause over-verification.
- Claude Opus 5 delivers frontier intelligence at half the cost of Claude Fable 5, with pricing unchanged from Opus 4.8 at $5 per million input tokens and $25 per million output tokens.
- The model is available on Anthropic API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry; migration requires updating the model ID and reviewing behavior changes.