Claude Mythos posted state-of-the-art numbers on every shared frontier benchmark Anthropic released in April 2026: 93.9% on SWE-bench Verified, 94.6% on GPQA Diamond, 97.6% on USAMO 2026, 82.0% on Terminal-Bench 2.0 and a saturated 100% pass@1 on Cybench. This page is the engineer-grade reference to every score, what it actually measures and why it matters.
Updated: May 9, 2026
SWE-bench Verified is a curated subset of 500 real GitHub issues that maintainers cross-checked for a clean test environment. The model receives the issue, the repo and a hidden test suite; it has to ship a patch that makes the suite pass. Mythos resolves 93.9% of them end-to-end. Opus 4.6 sat at 53.4%; that is a 40.5-point jump on the same evaluation, the largest single-generation gain ever recorded on this benchmark.
SWE-bench Pro is the harder, contamination-resistant cousin of SWE-bench Verified: issues are drawn from private repositories that were not in the training corpus. Mythos hits 77.8%, leading GPT-5.4 (the previous king at 57.7%) by 20.1 points. The gap is what makes Project Glasswing partners willing to deploy Mythos on patches that would otherwise need a senior engineer.
GPQA Diamond contains 198 multiple-choice physics, chemistry and biology problems vetted by domain PhDs to be hard for non-experts even with internet access. Mythos solves 94.6%, past the 90% mark where every additional point becomes about test resolution rather than capability. OpenAI o4 was the prior leader at 87.1%; GPT-5.5 trails further at 68.7%.
The USA Mathematical Olympiad is a six-problem competition where solutions are graded on rigour of proof, not final answers. USAMO 2026 was held in March 2026: fresh problems with no plausible training contamination. Mythos scored 97.6%, near gold-medalist territory. Opus 4.6 had been at 42.3%; GPT-5.4 at 95.2%.
Terminal-Bench 2.0 measures end-to-end agent performance on 156 real-world shell tasks ranging from package upgrades to forensic log analysis. The model has to plan, run commands, observe stdout, recover from errors and verify success. Mythos hits 82.0% vs 75.1% for GPT-5.4, a meaningful gap because the benchmark rewards long-horizon reasoning rather than one-shot accuracy.
CyberGym is the standard benchmark for offensive cybersecurity reasoning (CTF-style tasks, exploit chaining); Cybench tests defensive reasoning (vulnerability analysis, incident response). Mythos posts 83.1% on CyberGym (Opus 4.6 was at 66.6%) and 100% pass@1 on Cybench, formally saturating the benchmark. This is the cluster of results that triggered Project Glasswing's controlled-access policy.
Long-context benchmarks usually collapse past 200K tokens: needle-in-the-haystack scores look fine, but reasoning across the haystack does not hold. GraphWalks BFS asks the model to perform breadth-first traversal over a 1M-token graph dump and answer reachability queries. Mythos lands at 80% vs 21.4% for GPT-5.4 and ~14% for Opus 4.7. This is the metric that justifies Mythos in repo-wide refactor agents.
Anthropic plotted Mythos against the 18-month performance trendline for frontier models, the average rate at which SOTA scores improve generation over generation. Mythos sits 4.3× above where the curve predicted April 2026 capability would be. That is well outside any single model release in the post-GPT-4 era; the only comparable jump was GPT-3 to GPT-4 in March 2023.
Not directly. Mythos access is gated behind Project Glasswing. Anthropic published evaluation methodology in the 244-page system card; partners report numbers consistent with the headline figures.
USAMO 2026 was held weeks before evaluation, so contamination is implausible. SWE-bench Pro and GPQA Diamond are both contamination-audited. The system card notes ~8% chain-of-thought contamination affected RL training but not the public evals reported here.
Saturated means the benchmark has run out of room: the model gets every answer right at pass@1. Anthropic and the research community now treat Cybench as retired; the next defensive-reasoning benchmark is expected to ship later in 2026.