Anthropic Releases Claude Opus 4.7 With Safeguards
The updated model sharpens coding and long-task performance while adding tighter cybersecurity limits.
Topics
Anthropic on Thursday released Claude Opus 4.7, a new version of its flagship generally available AI model, with upgrades focused on coding, instruction-following, image analysis and long-running tasks.
The model is now available across Anthropic’s products, application programming interface and major cloud platforms.
Anthropic said Opus 4.7 performs better than Opus 4.6 on complex software engineering work and is designed to verify its own outputs before responding.
The company also said the model handles higher-resolution images and produces stronger documents and presentations.
The release comes with new safeguards aimed at blocking prohibited or high-risk cybersecurity uses.
Anthropic said it had reduced some of Opus 4.7’s cyber capabilities during training and would use lessons from its real-world deployment to inform any broader release of Mythos-class models.
Early testing indicated stronger results in complex coding tasks, where the system can complete assignments with less need for human supervision. The model is also designed to check its own outputs before returning responses.
Anthropic said Opus 4.7 outperforms its predecessor across internal and third-party benchmarks, including evaluations related to finance and other knowledge-based work.
Despite these improvements, the company said the model is less broadly capable than its more advanced system, Claude Mythos Preview, which remains limited to select users.
Anthropic said it reduced some of Opus 4.7’s cybersecurity capabilities during training and introduced safeguards to block high-risk or prohibited uses.
“We are releasing Opus 4.7 with safeguards that automatically detect and block requests that indicate prohibited or high-risk cybersecurity uses,” the company said in a blog post.
Claude Opus 4.7 is priced the same as the previous version, at $5 per million input tokens and $25 per million output tokens, according to the company.
