Anthropic gates its strongest model
Anthropic has decided not to release one of its most powerful models publicly and is instead offering preview access to a small set of partners, signalling a shift toward controlled distribution of frontier AI. Social posts report the preview is limited to roughly 40 major companies and the firm is backing an initiative called Project Glasswing to channel the model into guarded security and enterprise uses. The move shows that access to cutting-edge capability may become relationship-driven and monitored, not broadly available. (x.com) (x.com)
Anthropic built a new model called Claude Mythos Preview and then did the unusual part: on April 7 it said the model would not be generally available, even though it is now its most capable frontier model and outperforms its earlier Claude Opus 4.6 on many benchmarks. (anthropic.com) Instead of a public launch, Anthropic put Mythos Preview inside a closed program called Project Glasswing and said a limited set of partners would use it for defensive security work. (anthropic.com) The reason is not that the model writes prettier code or faster emails. Anthropic says Mythos Preview can surpass all but the most skilled humans at finding and exploiting software vulnerabilities, which turns a chatbot into something closer to an automated lockpicker for digital systems. (anthropic.com) Anthropic says the model has already found thousands of high-severity vulnerabilities, including flaws in every major operating system and every major web browser. In a separate technical post, the company said it identified and exploited zero-day vulnerabilities, meaning bugs that defenders had not discovered or patched yet. (anthropic.com 1) (anthropic.com 2) That is why the rollout looks more like a weapons-control program than a software launch. Anthropic said more than 40 organizations that build or maintain critical software infrastructure got access, while 12 named launch partners include Amazon Web Services, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, NVIDIA, and Palo Alto Networks. (anthropic.com) Anthropic is also putting money behind the gate. The company said it is committing up to $100 million in Mythos Preview usage credits and another $4 million in direct donations to open-source security organizations, which are the volunteer and nonprofit groups that maintain code used across the internet. (anthropic.com) This did not come out of nowhere. In February 2026, Anthropic updated its Responsible Scaling Policy to version 3.0, which is the company rulebook for what to test, what risks trigger extra safeguards, and when a model should be held back. (anthropic.com 1) (anthropic.com 2) The company’s own system card says Mythos Preview’s capability jump is the reason it chose not to make the model generally available. That is a sharper line than the usual artificial intelligence release pattern, where companies publish first and add restrictions later. (anthropic.com) (axios.com) Outside reporting fills in how unusual that is. The New York Times reported Anthropic was working with 40 companies to explore how the model could prevent cyberattacks, and CNBC reported partners include Microsoft, Amazon, Apple, CrowdStrike, and Palo Alto Networks. (nytimes.com) (cnbc.com) Anthropic’s technical team says more than 99 percent of the vulnerabilities it found have not yet been patched, so it is withholding most details under coordinated disclosure rules. The few details it did share include a now-patched 27-year-old bug in OpenBSD, which is an operating system known for security hardening. (anthropic.com) The bigger change is in who gets the frontier first. If Anthropic’s strongest systems are moving into invite-only programs tied to monitoring, credits, and specific use cases, then access to top-tier artificial intelligence may start looking less like a product launch and more like a private security clearance. (anthropic.com 1) (anthropic.com 2)
