Anthropic's new Claude Opus 4.7 is its best public AI model — but not its most powerful

Anthropic released Claude Opus 4.7, its most capable generally available AI model, while acknowledging the model is "less broadly capable" than Claude Mythos Preview, a more powerful system the company has declined to release publicly.

According to Anthropic, the release marks an advancement over Claude Opus 4.6 in several areas, among them software engineering, adherence to instructions, and the ability to carry out practical tasks. It also supports higher-resolution images — up to 2,576 pixels on the long edge, more than three times the limit of prior Claude models — and includes a new "xhigh" effort level that gives users finer control over the tradeoff between reasoning depth and response speed.

During training, Anthropic took deliberate steps to pull back on what the model can do in cybersecurity contexts, a process the company characterized as working to "differentially reduce" those capabilities relative to Mythos Preview. Built into the release are automated protections intended to intercept queries that fall into prohibited or elevated-risk cybersecurity categories before they can be acted on. Researchers and practitioners who want to apply the model to sanctioned security work have been pointed toward a dedicated application pathway, the Cyber Verification Program.

The release connects directly to Project Glasswing, Anthropic's cybersecurity initiative announced earlier this month, which brought in partners including AWS, Apple $AAPL, Microsoft $MSFT, Google $GOOGL, and Cisco $CSCO to test Claude Mythos Preview. That model, described in internal materials as "by far the most powerful AI model we've ever developed," has been made available only to a select group of companies. Anthropic has said its goal is to use what it learns from deploying less capable models — such as Opus 4.7 — to work toward a broader release of Mythos-class systems.

The existence of Mythos first became known after draft materials describing the model were left in a publicly accessible data store on Anthropic's website. Those documents described it as more advanced in cybersecurity tasks than any competing AI model and warned it could allow attacks to scale faster than defenders could respond.

Pricing for the new model holds at $5 per million input tokens and $25 per million output tokens, matching what Anthropic charged for Opus 4.6. Access is available through the company's consumer-facing Claude products, its API, and a range of third-party cloud infrastructure including Amazon $AMZN Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry.

Anthropic also noted that Mythos Preview remains the best-aligned model the company has trained according to its own evaluations, and that Opus 4.7's alignment assessment concluded the model is "largely well-aligned and trustworthy, though not fully ideal in its behavior."

Two additional developer features launched alongside the model: task budgets in public beta, which allow developers to guide how Claude allocates token spend across longer runs, and a new "/ultrareview" command in Claude Code that produces a dedicated review session to flag bugs and design issues.