| Claude Mythos | Claude Opus 4.7 | |
|---|---|---|
| arenaElo | 1478 | 1420 |
| humanEval | 98.7 | 93.1 |
| sweBench | 93.9 | 64.3 |
| mmlu | 94.2 | 92.3 |
| gpqa | 94.6 | 65.2 |
| aime | 97.6 | 88 |
| Price in/out ($/M) | $25 / $125 | $5 / $25 |
Mythos pushes SWE-bench Verified from 53.4% (Opus 4.6) to 93.9% and USAMO from 42.3% to 97.6%, a 4.3× leap on the model-performance trendline. Opus 4.7 stays the practical default: it is generally available, costs 5× less ($5 / $25 vs $25 / $125 per M tokens) and ships through every major cloud. Pick Mythos only if you have a Project Glasswing seat or Anthropic green-lit your use case; otherwise stay on Opus 4.7.