| GPT-5 Codex | Claude Opus 4.7 | |
|---|---|---|
| arenaElo | 1395 | 1420 |
| humanEval | 94.8 | 93.1 |
| sweBench | 68 | 64.3 |
| mmlu | 87.2 | 92.3 |
| gpqa | 58.4 | 65.2 |
| aime | 85 | 88 |
| Price in/out ($/M) | $8 / $32 | $15 / $75 |
GPT-5 Codex leads HumanEval by a nose (94.8 vs 93.1) and is cheaper. Opus 4.7 has a larger context (1M vs 400K) and better agent stability over multi-hour runs. For one-shot PRs, Codex. For repo-wide refactors and agents, Opus.