When Anthropic released Claude Mythos Preview in April 2026, it broke every existing AI benchmark in a single release. SWE-bench Verified at 93.9% — previously peaking at ~60% for frontier models. USAMO at 97.6% — university math olympiad problems solved at near-perfect accuracy. SWE-bench Pro at 77.8% and SWE-bench Multimodal at 59.0%. This guide breaks down each score, explains what it actually measures, and contextualizes what these numbers mean for developers.
SWE-bench Verified: 93.9%
SWE-bench Verified is the most credible software engineering benchmark because it uses real GitHub issues from popular open-source repositories. The model receives the issue description and the full codebase, then must generate a patch that passes the existing test suite. 93.9% means Mythos successfully resolves nearly 19 out of 20 real-world GitHub issues — a level of engineering capability that exceeds many junior developers on routine maintenance tasks.
USAMO: 97.6%
The USA Math Olympiad requires proof-based reasoning about abstract mathematical structures — not pattern matching or numerical computation. A 97.6% score represents genuine mathematical reasoning capability, not sophisticated retrieval. For comparison, GPT-5.4 scores approximately 75% on USAMO, and human experts who compete in the Olympiad score around 85%.
SWE-bench Pro: 77.8%
SWE-bench Pro is a harder variant with less test coverage and more ambiguous issue descriptions — closer to real production engineering work. The 77.8% score suggests Mythos maintains strong performance even when the scaffolding of well-specified tests is removed, which is where most models degrade significantly.
Full Benchmark Comparison
| Benchmark | Mythos | Opus 4.6 | GPT-5.4 |
|---|---|---|---|
| SWE-bench Verified | 93.9% | ~58% | ~61% |
| USAMO | 97.6% | ~65% | ~75% |
| SWE-bench Pro | 77.8% | ~42% | ~45% |
| SWE-bench Multimodal | 59.0% | ~30% | ~33% |
Note: Opus 4.6 and GPT-5.4 figures are approximate based on public benchmarks available at time of Mythos release.
What This Means for Developers
The gap between Mythos and other models on SWE-bench (93.9% vs ~60%) is not incremental — it is qualitative. At 93.9%, the model is reliable enough to act as an autonomous agent on real engineering tasks, not just a suggestion engine. The practical implication: AI-assisted code generation is shifting from "AI helps you code faster" to "AI can fix bugs and implement features end-to-end" for a large class of routine engineering work.