| Claude Code | GPT-5.4 Codex | |
|---|---|---|
| arenaElo | 1420 | 1408 |
| humanEval | 93.1 | 95.5 |
| sweBench | 64.3 | 70 |
| mmlu | 92.3 | 88.6 |
| gpqa | 65.2 | 60.8 |
| aime | 88 | 86 |
| Price in/out ($/M) | $15 / $75 | $9 / $36 |
GPT-5.4 Codex scores higher on HumanEval in isolation, but Claude Code's orchestration, hooks and memory give Opus 4.7 the edge on long runs. If your workflow fits inside a CLI, Claude Code wins on wall-clock time; if you integrate into your own harness, pick 5.4 Codex directly.