Claude Mythos

Updated 2026-04-09

Claude Mythos Preview is Anthropic’s newest frontier model, released on April 7, 2026, and the first in the 10-trillion-parameter class. It is so capable in cybersecurity tasks that Anthropic did not release it publicly, but only made it available to selected partners through Project Glasswing.

Model Profile

  • Parameters: about 10 trillion, the first model known at this scale
  • Training cost: estimated at $10 billion
  • SWE-bench: 94%, the highest known score on this coding benchmark
  • Release model: no public release, controlled access through Project Glasswing

According to Anthropic, these capabilities did not come from targeted security training, but from broader advances in code, reasoning, and autonomy.

Cybersecurity Capabilities

The difference from Opus 4.6 is not gradual. It is categorical. In Anthropic’s internal testing:

  • Opus 4.6: roughly 0% success rate in autonomous exploit development
  • Mythos Preview: 181 working exploits on the same Firefox test, versus 2 for Opus

Mythos Preview can find and exploit zero-day vulnerabilities in all major operating systems and browsers, including complex JIT heap sprays, kernel KASLR bypasses, and multi-stage ROP chains.

Concrete examples:

  • OpenBSD SACK bug, 27 years old: a flaw in the TCP SACK implementation combining two edge cases with an integer overflow, enabling remote denial of service
  • FFmpeg bug, 16 years old: a vulnerability in one of the most heavily tested open-source media projects in the world, including daily industrial fuzzing

Non-experts were able to ask Mythos overnight to find remote-code-execution vulnerabilities and wake up to a finished, working exploit in the morning.

Access Policy

Mythos Preview is not publicly available. Anthropic initially granted access to twelve partners:

Amazon, Cisco, Apple, Google, Microsoft, NVIDIA, JPMorgan Chase, CrowdStrike, Palo Alto Networks, AWS, The Linux Foundation, and Broadcom.

Anthropic is also providing $100 million in compute credits for Project Glasswing.

The logic is simple: defenders should be able to find and patch vulnerabilities before models with similar capabilities are broadly available.

Controversy

AI Anthropomorphization is especially visible in the Mythos presentation. @atmoio argues that the “Impressions” section of the system card, 20 pages of employee reactions to model output, is marketing rather than science, and that the model produces consciousness-like language partly because Anthropic’s own blog posts are in the training corpus.

Nina Schick reads Mythos as an inflection point and asks rhetorically whether this is not simply very expensive AGI. That remains contested. Mythos is a model with extraordinary capability, but not an autonomous system.

Connections

Sources