Opus 4.7 leaps forward

Anthropic has officially released Claude Opus 4.7. It shows particular improvement in complex software engineering, long-running autonomous tasks, and strict instruction following, and its ability to verify its own outputs has also been strengthened.

Vision performance has also improved. It can now accept images up to 2,576px on the long edge, which benefits dense screenshots, complex diagrams, and precise computer-use tasks. Anthropic noted that for specialized work like UI, slides, and documents, it produces more polished and creative results.

Numbers were also presented. On 93 coding benchmarks, it improved by 13% over Opus 4.6, and in internal research agent evaluations it recorded a top-tier combined score of 0.715 across 6 modules. In the financial module, it improved to 0.813 vs 0.767, and even in low-cost settings it showed efficiency similar to Opus 4.6's mid-tier level.

The key changes are as follows.

• Carries out complex, long-running tasks more persistently to completion
• Stronger literal adherence to instructions, meaning existing prompts may need to be readjusted
• Enhanced use of filesystem-based memory, benefiting multi-session tasks
• Security features have been strengthened, but high-risk cyber uses are automatically blocked

The release is available simultaneously in 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.