OpenAI
o3-mini benchmark results
12 results on 6 benchmarks from 4 sources.
Also listed as: o3-mini-high · o3-mini-2025-01-31_low · o3-mini-2025-01-31_high · o3-mini (high) · o3-mini-2025-01-31_medium · o3-mini (medium) · O3 Mini · Agentless Lite + O3 Mini (20250214) · O3-Mini-2025-01-31 (Low) · o3-mini-2025-01-31__low
All benchmark results
| Benchmarks | Score | Rank | Setting | Source type | Date |
|---|---|---|---|---|---|
| LMArena Text | 1364 | 203 | Overall, style control on; 18589 votes; 95% CI 1358–1369“o3-mini-high” | Official leaderboard | Oct 8, 2026 |
| LMArena Text | 1348 | 218 | Overall, style control on; 56750 votes; 95% CI 1344–1352 | Official leaderboard | Oct 8, 2026 |
| LiveCodeBench | 51.1% | 14 | Code generation, 182 problems dated 2025-01-01…2025-04-07“O3-Mini-2025-01-31 (Low)” | Official leaderboard | Apr 7, 2025 |
| GPQA Diamond | 77.0% | 128 | Epoch AI run, accuracy; effort: high | Independent run | Feb 13, 2025 |
| GPQA Diamond | 74.3% | 144 | Epoch AI run, accuracy; effort: medium | Independent run | Jan 31, 2025 |
| GPQA Diamond | 68.2% | 165 | Epoch AI run, accuracy | Independent run | Jul 20, 2026 |
| SWE-bench Verified | 42.4% | 77 | agent: Agentless Lite | Official leaderboard | Feb 14, 2025 |
| AIME (OTIS Mock AIME 2024–2025) | 76.9% | 117 | Epoch AI run, OTIS Mock AIME 2024–2025; effort: high | Independent run | Feb 27, 2025 |
| AIME (OTIS Mock AIME 2024–2025) | 63.9% | 142 | Epoch AI run, OTIS Mock AIME 2024–2025; effort: medium | Independent run | Feb 25, 2025 |
| AIME (OTIS Mock AIME 2024–2025) | 44.4% | 179 | Epoch AI run, OTIS Mock AIME 2024–2025 | Independent run | Jul 20, 2026 |
| MATH Level 5 | 96.5% | 10 | Epoch AI run, accuracy; effort: high | Independent run | Feb 13, 2025 |
| MATH Level 5 | 95.2% | 15 | Epoch AI run, accuracy; effort: medium | Independent run | Jan 31, 2025 |
FAQ
Where do these o3-mini scores come from?
From Epoch AI Benchmarking Hub, LiveCodeBench leaderboard, LMArena leaderboard dataset, and SWE-bench leaderboards (swebench.com). Each row links to the original leaderboard, dataset or report, and vendor-reported numbers are labelled.