On August 5, 2025, Anthropic released Claude Opus 4.1, an incremental update to Opus 4 that focuses on agentic tasks, real-world coding, and improved reasoning for detail tracking and agentic search.
Availability and pricing
Opus 4.1 is available now to paid Claude users and through Claude Code. It is also accessible via the API and is offered on Amazon Bedrock and Google Cloud Vertex AI. Anthropic states that pricing remains the same as for Opus 4.
Performance and benchmarks
Anthropic reports that Opus 4.1 achieves 74.5% on SWE-bench Verified for coding performance. Overall, Opus 4.1 shows improvements across most capabilities relative to Opus 4, with especially notable gains in multi-file code refactoring.
Partner findings highlighted in the announcement include:
- GitHub notes that Opus 4.1 improves across most capabilities compared to Opus 4, with particularly strong performance in multi-file refactoring.
- Rakuten Group reports that Opus 4.1 can pinpoint exact corrections within large codebases without introducing unnecessary changes or bugs, which their engineers prefer for routine debugging tasks.
- Windsurf reports that Opus 4.1 delivers approximately a one standard deviation improvement over Opus 4 on their junior developer benchmark, a jump they compare to the advance from Sonnet 3.7 to Sonnet 4.
Detail tracking and agentic search
The update emphasizes better tracking of details and enhanced agentic search capabilities—features that matter for complex, multi-step problem solving, research and data analysis workflows.
Benchmarks and extended thinking notes
Anthropic describes Claude models as hybrid reasoning systems and clarifies that the reported benchmark scores were obtained with and without extended thinking. They specify which results used extended thinking (up to 64K tokens):
- No extended thinking: SWE-bench Verified, Terminal-Bench.
- Benchmarks reported with extended thinking: TAU-bench, GPQA Diamond, MMMLU, MMMU, AIME.
For TAU-bench, the methodology included a prompt addendum to both the Airline and Retail Agent Policy instructing Claude to better leverage its reasoning while using extended thinking and tool use. The model was encouraged to write down its thoughts separately during multi-turn trajectories; to accommodate the additional steps the maximum number of model completions was raised from 30 to 100, though most trajectories completed under 30 steps.
For SWE-bench, Anthropic used the same simple scaffold across the Claude 4 family that provides two tools: a bash tool and a file editing tool that operates via string replacements. They no longer include the third "planning tool" that was used in Sonnet 3.7. Scores for the Claude 4 models are reported out of the full 500 problems (OpenAI scores are reported on a 477-problem subset).
Recommendation for developers
Anthropic recommends upgrading from Opus 4 to Opus 4.1 for all uses. Developers can access the model via the API with the identifier:
claude-opus-4-1-20250805
Anthropic also points readers to the system card, model page, pricing page and documentation for further details, and requests user feedback to guide future improvements.
Data sources and references
The announcement references several sources used for comparison and methodology, including OpenAI's o3 launch post and system card, Gemini 2.5 Pro model card, and Anthropic's own Sonnet 3.7 and Claude 4 launch posts. The company also provides notes on how benchmark reporting and extended thinking were handled in their evaluations.
