| OpenAI o4 | Claude Opus 4.7 | |
|---|---|---|
| arenaElo | 1438 | 1420 |
| humanEval | 94 | 93.1 |
| sweBench | 72 | 64.3 |
| mmlu | 92.5 | 92.3 |
| gpqa | 87.1 | 65.2 |
| aime | 94 | 88 |
| Price in/out ($/M) | $12 / $48 | $15 / $75 |
o4 is state-of-the-art on GPQA (87% vs 65%) and wins one-shot math and science. Opus 4.7 runs faster, has 4× the context and beats o4 on multi-hour agentic workflows. Pick o4 for hard single answers; pick Opus 4.7 for agents that must ship.