| Gemini 3 Ultra | Gemini 3 Pro | |
|---|---|---|
| arenaElo | 1441 | 1405 |
| humanEval | 91 | 87.3 |
| sweBench | 66 | 63 |
| mmlu | 93.4 | 91.1 |
| gpqa | 72.6 | 62.5 |
| aime | 93 | 90 |
| Price in/out ($/M) | $18 / $72 | $7 / $21 |
Ultra adds ~2 points on MMLU, ~10 on GPQA and a 3M context window. For most production workloads, Pro already hits the Pareto frontier. Move to Ultra only when you genuinely process 1M+ token inputs or need a quality ceiling.