Claude Mythos is Anthropic's most capable AI model, released in April 2026 and available only through an invite-only preview called Project Glasswing. It is not publicly accessible on Claude.ai. Mythos scored 93.9% on SWE-bench Verified (real GitHub issue resolution), 97.6% on USAMO (university-level math olympiad), and 59.0% on SWE-bench Multimodal — breaking every existing frontier AI benchmark at the time of release.
Why Mythos Is Different
Previous frontier models like Claude Opus 4.6 and GPT-5.4 plateau around 50-60% on SWE-bench Verified. Mythos's 93.9% represents a qualitative jump in software engineering capability — not an incremental improvement. The model can reason about multi-file codebases, trace complex execution paths, generate working patches, and verify its own output without human guidance on a per-task basis.
Key Benchmark Scores
| Benchmark | Claude Mythos | What it measures |
|---|---|---|
| SWE-bench Verified | 93.9% | Resolving real GitHub issues in open-source codebases |
| USAMO | 97.6% | University-level math olympiad problems |
| SWE-bench Pro | 77.8% | Harder, proprietary-style software engineering tasks |
| SWE-bench Multimodal | 59.0% | Software engineering tasks with visual context |
Project Glasswing Access
Mythos is accessible only through Project Glasswing — Anthropic's invite-only cybersecurity initiative. Partners include 12 major technology companies who use the model for vulnerability research. There is no public waitlist or API access outside this program. Anthropic has not announced a timeline for general availability.
The restricted access reflects concerns about dual-use capabilities: a model that can autonomously find and exploit zero-day vulnerabilities in critical infrastructure is as dangerous as it is useful. Anthropic's approach — deploy to trusted security researchers first, measure outcomes, then expand — mirrors how nuclear and biosecurity capabilities have historically been managed.