| Claude Mythos | GPT-5.4 Codex | |
|---|---|---|
| arenaElo | 1478 | 1408 |
| humanEval | 98.7 | 95.5 |
| sweBench | 93.9 | 70 |
| mmlu | 94.2 | 88.6 |
| gpqa | 94.6 | 60.8 |
| aime | 97.6 | 86 |
| Price in/out ($/M) | $25 / $125 | $9 / $36 |
GPT-5.4 Codex held the HumanEval crown at 95.5%. Mythos pushes it to 98.7% and posts 93.9% on SWE-bench Verified, resolving roughly 19 of 20 real GitHub issues end-to-end. Codex still ships in a usable CLI today, integrates with Codex cloud and costs ~3× less per token. For agencies and SaaS teams shipping production code this quarter, Codex remains the practical pick. Mythos is the right answer when SWE-bench reliability matters more than availability.