| GPT-5.4 Codex | Codestral | |
|---|---|---|
| arenaElo | 1408 | 1254 |
| humanEval | 95.5 | 89.7 |
| sweBench | 70 | 35 |
| mmlu | 88.6 | 73.2 |
| gpqa | 60.8 | 41 |
| aime | 86 | 30 |
| Price in/out ($/M) | $9 / $36 | $0.3 / $0.9 |
GPT-5.4 Codex tops the HumanEval chart but at 30× Codestral's price. Codestral remains the right pick for IDE autocomplete and high-volume inline use. 5.4 Codex shines on autonomous PR generation where a single success pays for the call.