Claude Opus 4.7 rolls out

Anthropic released Claude Opus 4.7 on April 16, making its newest flagship model available across Claude’s web app, mobile and desktop apps, and developer API. (anthropic.com) Anthropic said Opus 4.7 improves on Opus 4.6 in advanced software engineering, especially on harder coding tasks that run for longer and need less supervision. The company also said the model is now available through Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. (anthropic.com) For developers, the launch adds the API model `claude-opus-4-7`, a 1 million-token context window, up to 128,000 output tokens, a new “xhigh” effort setting, and a beta feature called task budgets for long agent loops. Anthropic kept pricing at $5 per million input tokens and $25 per million output tokens, the same as Opus 4.6. (platform.claude.com; anthropic.com) A context window is the amount of text, code, or other material a model can keep in working memory during one session. Agentic coding is the newer push to let a model plan, call tools, inspect files, and keep working across many steps instead of answering one prompt at a time. (platform.claude.com; anthropic.com) Anthropic tied the release to a second story: what it chose not to ship. The company said Opus 4.7 is “less broadly capable” than Claude Mythos Preview and that it deliberately tried during training to reduce some cybersecurity capabilities before release. (anthropic.com; cnbc.com) That decision follows Project Glasswing, announced on April 7, which gives a limited group of partners access to Mythos Preview for defensive security work on critical software. Anthropic said launch partners include Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks. (anthropic.com) Anthropic said Mythos Preview will not be made generally available for now, and described its longer-term goal as learning how to deploy “Mythos-class models” safely at scale. In a technical write-up, the company said Mythos Preview was capable in testing of finding and exploiting zero-day vulnerabilities in major operating systems and web browsers. (anthropic.com) Opus 4.7 is the first public model Anthropic says it is using to test automated cyber safeguards in the real world. The company said those safeguards are designed to detect and block requests that suggest prohibited or high-risk cybersecurity use, while a new Cyber Verification Program is meant to give approved security researchers access for vulnerability research, penetration testing, and red-teaming. (anthropic.com) The model also expands Claude’s computer-use push with higher-resolution image handling: Anthropic raised the maximum image size to 2,576 pixels, or 3.75 megapixels, from 1,568 pixels, or 1.15 megapixels. The company said that change is aimed at screenshot reading, document understanding, and on-screen actions that depend on precise coordinates. (platform.claude.com) The release lands two months after Anthropic introduced Opus 4.6 in February and one week after it unveiled Project Glasswing. Opus 4.7 is now the company’s most capable generally available model, while its most powerful cyber model remains behind a tighter gate. (anthropic.com; anthropic.com; anthropic.com)