Anthropic shipped Claude Opus 4.7 today, calling it the company’s most capable Opus model yet and also making it generally available worldwide; the release includes automated safeguards that detect and block prohibited or high‑risk cybersecurity requests. On SWE-bench Multilingual, Opus 4.7 scored 80.5%, compared with Opus 4.6’s 77.8%. On GDPVal‑AA, Opus 4.7 scored 1,753 Elo, compared with GPT‑5.4’s 1,674 Elo.
Anthropic framed Opus 4.7 as its most capable Opus release and published benchmark figures to demonstrate improvements. The company also described safeguards and reported experimenting with measures to reduce cyber capabilities during training.
Claude Opus 4.7 has demonstrated notable improvements in several benchmark tests. On OfficeQA Pro, a benchmark evaluating question-answering capability, Opus 4.7 achieved a score of 80.6%, significantly outperforming its predecessor, Opus 4.6, which scored 57.1%. Additionally, Opus 4.7 surpassed competitors with GPT-5.4 scoring 51.1% and Gemini 3.1 Pro at 42.9%.
In terms of long-term coherence, as measured by Vending-Bench 2 which assesses the ability to maintain a coherent line of reasoning over extended interactions, Opus 4.7 managed a money balance of $10,937 compared to $8,018 for Opus 4.6. Furthermore, in the GDPVal-AA benchmark, which assesses model ability in artificial intelligence tasks using the Elo rating system, Opus 4.7 scored 1,753 Elo against GPT-5.4, which scored 1,674 Elo. These scores emphasize the advancements Claude Opus 4.7 has made in both its processing capabilities and overall performance.
Claude Opus 4.7 launches with automated safeguards that detect and block prohibited or high‑risk cybersecurity requests. The release includes mechanisms that identify such requests and prevent the model from responding to them. Anthropic said these safeguards are built into the model as automated controls.
Anthropic confirmed it conducted experiments to selectively reduce Opus 4.7’s cyber capabilities during training. The company described those steps as experimental efforts to differentially reduce cyber capabilities for the model. Anthropic offers a Cyber Verification Program that provides access to the automated safeguards.
Mythos Preview remains restricted to vetted security firms. The UK’s AI Security Institute evaluated Mythos as the first AI to complete ‘The Last Ones,’ a 32-step corporate network attack simulation that typically takes human red teams 20 hours. Anthropic continues to limit access to Mythos Preview to vetted security firms.
Anthropic describes Claude Opus 4.7 as its most powerful model publicly available and highlights benchmark improvements over prior Opus versions and competing models across multiple standardized tests. The company presented improvements in question‑answering performance, multilingual ability and long‑term coherence as part of its published benchmark results.
The release includes automated safeguards that detect and block prohibited or high‑risk cybersecurity requests, Anthropic said it experimented with efforts to differentially reduce cyber capabilities during training, and the company offers a Cyber Verification Program while keeping Mythos Preview restricted to vetted security firms.
