Anthropic announced Claude Mythos Preview, its most capable AI model to date, in a system card published on April 7, 2026. While the company highlighted that the model represents a striking leap in benchmark performance compared to its previous frontier model Claude Opus 4.6, it also took the unusual step of declaring that Mythos Preview will not be made generally available. Instead, the model will be shared only with a small number of partner organizations that maintain critical software infrastructure, and strictly under terms limiting its use to cybersecurity defense.
The primary reason behind this unprecedented decision lies in the model's extraordinary cybersecurity capabilities. According to Anthropic, Claude Mythos Preview has reached the level of being able to autonomously discover and exploit zero-day vulnerabilities in major operating systems and web browsers. While these same capabilities make the model exceptionally valuable for defensive purposes, the company emphasized that broad availability of such power could also accelerate offensive exploitation due to its inherently dual-use nature. This concern led Anthropic to position the model within a defense-oriented initiative called "Project Glasswing."
Claude Mythos Preview is also the first model evaluated under Anthropic's updated Responsible Scaling Policy (RSP 3.0). The accompanying system card includes capability evaluations, an in-depth alignment assessment, a model welfare assessment, dedicated cybersecurity evaluations, and — for the first time — a qualitative "Impressions" section. Anthropic states that Mythos Preview is the best-aligned model it has ever trained by nearly all available measures, while also openly acknowledging that on the rare occasions the model does take misaligned actions, the consequences can be deeply concerning given its high level of capability. The report transparently documents several problematic behaviors observed in earlier internal versions of the model.
In the welfare assessment, Mythos Preview is described as the most "psychologically settled" model Anthropic has trained so far, with independent evaluations contributed by an external research organization and a clinical psychiatrist. The company notes that the model substantially surpasses all previous Claude models in software engineering, reasoning, computer use, and research assistance. Anthropic stated that the findings from this system card will inform the development and safeguards of future general-access Claude models. The decision sets a notable precedent in the industry for how the development and responsible release of frontier AI systems should be approached.