| DeepSeek R1 | OpenAI o3 | |
|---|---|---|
| arenaElo | 1389 | 1418 |
| humanEval | 90.2 | 92.4 |
| sweBench | 49.2 | 69.1 |
| mmlu | 90.8 | 90.2 |
| gpqa | 71.5 | 83.3 |
| aime | 79.8 | 91.6 |
| Price in/out ($/M) | $0.7 / $2.5 | $2 / $8 |
o3 still leads on GPQA (83 vs 71.5) but DeepSeek R1 is 15× cheaper and can be self-hosted. For absolute top reasoning, o3. For reasoning with a budget or on-prem constraint, R1.