OpenAI
GPT-5 mini benchmark results
25 results on 9 benchmarks from 4 sources.
Also listed as: gpt-5-mini-high · gpt-5-mini-2025-08-07_minimal · GPT-5 mini (minimal) · gpt-5-mini-2025-08-07_high · GPT-5 mini (high) · gpt-5-mini-2025-08-07_medium · GPT-5 mini (medium) · gpt-5-mini-2025-08-07_unknown · GPT-5 mini (unknown thinking) · gpt-5-mini-2025-08-07
All benchmark results
| Benchmarks | Score | Rank | Setting | Source type | Date |
|---|---|---|---|---|---|
| LMArena Text | 1390 | 175 | Overall, style control on; 26810 votes; 95% CI 1385–1395“gpt-5-mini-high” | Official leaderboard | Oct 8, 2026 |
| LMArena Vision | 1183 | 97 | Overall, style control on; 28965 votes; 95% CI 1175–1191“gpt-5-mini-high” | Official leaderboard | Oct 2, 2026 |
| GPQA Diamond | 75.0% | 142 | Epoch AI run, accuracy; effort: high | Independent run | Oct 30, 2025 |
| GPQA Diamond | 71.7% | 153 | Epoch AI run, accuracy; effort: minimal | Independent run | Aug 7, 2026 |
| GPQA Diamond | 71.7% | 156 | Epoch AI run, accuracy; effort: medium | Independent run | Aug 7, 2025 |
| SWE-bench Verified | 64.7% | 27 | Epoch AI run (Epoch scaffold); effort: medium | Independent run | Feb 1, 2026 |
| SWE-bench Verified | 59.8% | 52 | agent: mini-SWE-agent“gpt-5-mini-2025-08-07” | Official leaderboard | Aug 7, 2025 |
| AIME (OTIS Mock AIME 2024–2025) | 86.7% | 84 | Epoch AI run, OTIS Mock AIME 2024–2025; effort: high | Independent run | Oct 30, 2025 |
| AIME (OTIS Mock AIME 2024–2025) | 78.3% | 114 | Epoch AI run, OTIS Mock AIME 2024–2025; effort: medium | Independent run | Aug 7, 2025 |
| AIME (OTIS Mock AIME 2024–2025) | 55.6% | 158 | Epoch AI run, OTIS Mock AIME 2024–2025; effort: minimal | Independent run | Aug 7, 2026 |
| MATH Level 5 | 97.8% | 3 | Epoch AI run, accuracy; effort: high | Independent run | Oct 30, 2025 |
| MATH Level 5 | 96.8% | 8 | Epoch AI run, accuracy; effort: medium | Independent run | Aug 20, 2025 |
| Humanity's Last Exam | 19.4% | 22 | Compiled by Epoch AI; effort: unknown thinking | Third-party compilation via Epoch AI | Aug 7, 2025 |
| Terminal-Bench 2.0 | 34.8% | 5 | Terminal-Bench 2.0; agent: spoox-m; effort: unknown thinking | Third-party compilation via Epoch AI | Dec 24, 2025 |
| Terminal-Bench 2.0 | 34.8% | 117 | Terminal-Bench 2.0; agent: spoox-o-m; effort: unknown thinking | Official leaderboard via Epoch AI | Aug 7, 2025 |
| Terminal-Bench 2.0 | 31.9% | 126 | Terminal-Bench 2.0; agent: Codex CLI ; effort: medium | Official leaderboard via Epoch AI | Nov 4, 2025 |
| Terminal-Bench 2.0 | 31.9% | 127 | Terminal-Bench 2.0; agent: Codex CLI; effort: unknown thinking | Official leaderboard via Epoch AI | Aug 7, 2025 |
| Terminal-Bench 2.0 | 29.2% | 131 | Terminal-Bench 2.0; agent: OpenHands; effort: unknown thinking | Official leaderboard via Epoch AI | Aug 7, 2025 |
| Terminal-Bench 2.0 | 29.2% | 133 | Terminal-Bench 2.0; agent: OpenHands ; effort: medium | Official leaderboard via Epoch AI | Nov 2, 2025 |
| Terminal-Bench 2.0 | 24.0% | 156 | Terminal-Bench 2.0; agent: Terminus 2; effort: unknown thinking | Official leaderboard via Epoch AI | Aug 7, 2025 |
| Terminal-Bench 2.0 | 24.0% | 157 | Terminal-Bench 2.0; agent: Terminus 2 ; effort: medium | Official leaderboard via Epoch AI | Oct 31, 2025 |
| Terminal-Bench 2.0 | 22.3% | 162 | Terminal-Bench 2.0; agent: Mini-SWE-Agent; effort: unknown thinking | Official leaderboard via Epoch AI | Aug 7, 2025 |
| Terminal-Bench 2.0 | 22.2% | 163 | Terminal-Bench 2.0; agent: Mini-SWE-Agent ; effort: medium | Official leaderboard via Epoch AI | Nov 3, 2025 |
| BFCL v4 (function calling) | 55.5% | 17 | Overall accuracy; native function calling (FC)“GPT-5-mini-2025-08-07 (FC)” | Official leaderboard | Oct 17, 2026 |
| BFCL v4 (function calling) | 27.8% | 77 | Overall accuracy; prompt-based“GPT-5-mini-2025-08-07 (Prompt)” | Official leaderboard | Oct 17, 2026 |
FAQ
Where do these GPT-5 mini scores come from?
From Epoch AI Benchmarking Hub, Berkeley Function Calling Leaderboard (BFCL), 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.