The concept has been on this hub before. In “Anthropic Built an AI That Finds Zero-Day Vulnerabilities. It Decided Not to Release It.” and the follow-up deep-dive “When AI Becomes the Best Hacker in the Room”, the hub covered the company’s decision to build but not broadly release a powerful vulnerability-finding model. On April 14, 2026, that model got a name, a coalition, partners, and a price.
Claude Mythos Preview is Anthropic’s formal name for the gated cybersecurity AI. Project Glasswing is the access coalition it sits within. According to Anthropic’s announcement, Microsoft, NVIDIA, and AWS are named coalition partners. Access is restricted to more than 40 defensive organizations and is not available to the general public.
Anthropic states the model has identified thousands of high-severity vulnerabilities that survived decades of human review. That claim has not been independently verified, gated access makes independent evaluation structurally impossible, which is worth noting plainly. Benchmark figures from the announcement include a GPQA score of 0.9 and an internal lead on SWE-Bench Pro, according to Anthropic’s internal evaluation. No third-party assessment exists.
Pricing for the preview has been announced at $25 per million input tokens and $125 per million output tokens, according to Anthropic. Both figures reflect announced pricing and may change at general availability.
What the formalization changes
The prior hub briefs documented the decision architecture: Anthropic built something it considered too dangerous for open release. What’s new today isn’t that decision, it’s that the decision has been institutionalized into a product. Named model. Named coalition. Named partners. Posted pricing. That’s the shift from “we chose not to release it” to “here is the formal structure we’ve built around it.”
For security practitioners and compliance teams, the formalization matters because it defines who is in and who is out. If your organization isn’t among the more than 40 in the Glasswing coalition, Claude Mythos Preview is not accessible to you, regardless of your security posture, budget, or use case. The access architecture itself is the policy decision.
What to watch
The coalition’s composition will expand or contract over time, what criteria Anthropic uses for Glasswing membership, and whether those criteria become public, will define whether the gated model remains an elite tier or evolves into a broader access program. The three named partners (Microsoft, NVIDIA, AWS) are all infrastructure players, not end users, what their specific roles in Glasswing are, beyond the announcement, is still unclear.
Independent evaluation remains structurally impossible as long as access is gated. If a third-party security firm inside the coalition publishes findings, that’s the data point that makes the capability claims assessable. Watch for that.
TJS synthesis
The hub has covered the debate over whether restricting a powerful AI is the right safety choice. Today’s news is that Anthropic has stopped debating and started building the infrastructure around that choice. Project Glasswing is the formalization of a governance position, gated access as safety architecture. Whether that architecture holds, whether it scales, and whether it remains coherent as the model improves are the questions worth watching. The deep-dive below maps what the coalition structure actually means for organizations on the outside of the firewall.