Claude Opus 4.7

2026-04-16 · Publisher : Anthropic · Domain : Artificial Intelligence

Anthropic rend Claude Opus 4.7 généralement disponible : un net progrès sur Opus 4.6 en ingénierie logicielle difficile, avec une vision haute résolution et un niveau d'effort xhigh, au même prix.

What changes

  • Gains sur les tâches de génie logiciel les plus difficiles et les travaux longs.
  • Vision : images jusqu'à 2 576 pixels sur le grand côté (environ 3,75 mégapixels).
  • Nouveau niveau d'effort xhigh entre high et max ; budgets de tâches en bêta.
  • Nouveau tokenizer produisant environ 30 % de tokens en plus ; température, top_p et top_k non par défaut renvoient une erreur 400.

Anthropic makes Claude Opus 4.7 generally available: a clear step up from Opus 4.6 in difficult software engineering, with high-resolution vision and an xhigh effort level, at the same price. ### What changes - Gains on the hardest software engineering tasks and on long-running work. - Vision: images up to 2,576 pixels on the long edge (about 3.75 megapixels). - New xhigh effort level between high and max; task budgets in beta. - New tokenizer producing about 30% more tokens; non-default temperature, top_p and top_k return a 400 error. ### Specifications - Context window and maximum output: not specified in the blog post. - Knowledge cutoff: not disclosed in the blog post. - Availability: Claude products, API, Amazon Bedrock, Vertex AI, Microsoft Foundry. API identifier: `claude-opus-4-7`. ### Pricing | Item | Pricing (USD per million tokens) | | --- | --- | | Input | 5 | | Output | 25 | | Cache write 5 min / 1 h | 6.25 / 10 | | Cache read | 0.50 | | Batch (input / output) | 2.50 / 12.50 | ### Benchmarks The blog post mostly presents charts; figures cited by partners (results reported by the vendor in the announcement): | Evaluation | Opus 4.7 | Opus 4.6 | | --- | --- | --- | | CursorBench | 70% | 58% | | XBOW, visual acuity | 98.5% | 54.5% | | Harvey BigLaw Bench (high effort) | 90.9% | not recorded | | Databricks OfficeQA Pro | 21% fewer errors | baseline | ### Key takeaways Opus 4.7 changes the behavior of the API (tokenizer, sampling parameters): a migration calls for cost and prompt testing.

Links

Tags : LLM, Anthropic, Claude, Opus, Codage, Vision