Claude Opus 4.7

Anthropic has made Claude Opus 4.7 generally available (GA).

The core changes are improvements in high-difficulty software engineering, long-running tasks, instruction following, and self-verification capabilities. Anthropic explains that more complex coding tasks can now be entrusted with less oversight burden than before, and emphasizes that consistency and accuracy have improved even in long, multi-step tasks.

Multimodal performance has also been strengthened. In particular, the model reads high-resolution images better and can accept input with a long-edge length of up to 2,576px. This is said to give it an advantage in dense screenshots, complex diagrams, and precise computer-use tasks. For professional documents, slides, and UI generation, it is described as producing results with better taste and more creativity.

Improvements over Opus 4.6 were also confirmed in benchmarks and real-world evaluations.

• On the 93-task coding benchmark, the solve rate rose by 13%, and the model solved 4 additional tasks that Opus 4.6 and Sonnet 4.6 had failed to solve.
• On the internal research-agent benchmark, it showed top-tier efficiency with a combined score of 0.715 across 6 modules, with General Finance in particular improving to 0.813 vs 0.767.
• Across multiple partner evaluations in finance, law, document analysis, long-context reasoning, tool use, and code review, lower error rates, better traceability, and fewer tool errors were reported.

In terms of safety, Anthropic explains that the model maintains a profile similar to Opus 4.6, with some items improving and others slightly weakening. In particular, for potential cybersecurity abuse, automated detection and blocking guardrails have been applied, and legitimate security researchers can join the Cyber Verification Program. This deployment experience is presented as a preparatory step toward the future release of a stronger Mythos-class model.

The release terms are also clear. It is immediately available across Claude products, the API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry, priced the same as Opus 4.6 at $5 per million input tokens and $25 per million output tokens. The API model name is claude-opus-4-7.