Claude Mythos

Claude Mythos is the most specialised model Anthropic has built, and the only one it does not offer to the general public. It is a frontier system tuned for cybersecurity research and scientific discovery, and access is controlled under an internal program the company calls Project Glasswing. The reason is straightforward. The same abilities that let Mythos find software flaws before attackers do could also be misused, so Anthropic gates who can run it.

Mythos came to wider attention when Anthropic disclosed that the model had uncovered thousands of previously unknown vulnerabilities across widely used software. That result is what makes it valuable to defenders and also what makes it sensitive. Rather than release it openly, the company limits Mythos to vetted security teams, selected researchers and its own internal red teaming, with usage logged and reviewed.

Technically, Mythos is built on the same foundation as the rest of the current lineup and shares the safety training that shapes Claude Opus 4.8 and Claude Fable 5. What sets it apart is specialisation. It has been trained and evaluated heavily on code analysis, exploit reasoning, protocol behaviour and scientific problem solving, so it goes deeper on those tasks than a general model would while being deliberately constrained elsewhere.

For most people, Mythos is not a model you will use directly, and that is by design. If you are building an ordinary product, the right choices are Opus 4.8 for hard reasoning and agentic coding, Fable 5 for creative and long form work, or Claude Haiku 4.5 for fast, high volume tasks. Mythos exists for a narrow set of defensive and research uses where the upside of finding a flaw first outweighs the risk of the capability itself.

Anthropic has framed Mythos as a test of how to ship a genuinely dual use model responsibly. The gating, the logging and the staged access are meant to let defenders benefit while keeping the same tools away from bad actors. Whether that balance holds as the model improves is one of the open questions in frontier AI safety, and we cover it as it develops.

Access to Mythos follows a responsible disclosure model. Findings are reported to affected vendors and given time to patch before any detail becomes public, and the teams granted access agree to use the model for defence rather than attack. Anthropic has said it treats Mythos as a deliberately narrow deployment, widening access only as it builds confidence that the safeguards hold. For everyone outside that program, the practical takeaway is simple: the defensive benefits reach you through more secure software, not through direct use of the model.

You can read our reporting on the Mythos security debate, or step back and compare the models that are open to everyone on the Claude model family page. Access requests for Mythos go through Anthropic directly and are reviewed case by case rather than granted through the standard developer console.