Claude Mythos

Claude Mythos Preview was announced on April 7, 2026, after details leaked from an unsecured Anthropic staging endpoint weeks earlier. Anthropic described it as a new class of intelligence above their existing Opus tier, with particular strength in long-horizon reasoning, autonomous coding, and cybersecurity tasks. In pre-release testing it identified thousands of previously unknown zero-day vulnerabilities across major operating systems and browsers, including flaws that had survived decades of human security review.

Because of those offensive capabilities, Anthropic chose not to release Mythos publicly. Instead, access is restricted to Project Glasswing, an invitation-only consortium of organizations including cloud providers, security firms, and government-adjacent bodies. Members can use the model specifically for defensive security work, in exchange for sharing safety telemetry with Anthropic. The UK AI Security Institute independently evaluated Mythos and confirmed it completes difficult multi-step attack simulations at rates no prior model achieved.

For builders, the practical implication is indirect: Mythos is not available to call via the public API today. But the capabilities Anthropic is testing in Mythos tend to migrate into the next public model generation. If you see a dramatic jump in Claude's reasoning or security-analysis quality in a future Opus release, Mythos is likely the source. It also signals that AI-assisted vulnerability discovery is shifting from a research curiosity to a genuine infrastructure concern, which matters for anyone shipping production software.