Home / Benchmarks / GPQA Diamond
Benchmark leaderboard
GPQA Diamond
Graduate-level science questions written to resist web lookup; Diamond is the hardest 198-question subset. Scores via the BenchLM GPQA leaderboard, whose ledger notes most rows are Diamond while some are GPQA main-set — treat close gaps as directional.
| # | Model | Maker | Score | Measured | Link |
|---|---|---|---|---|---|
| 1 | GPT-6 Astra | OpenAI | 96.0 / 100 | Oct 14, 2026 | Source |
| 2 | Gemini 3.8 Flash | Google DeepMind | 95.0 / 100 | Oct 14, 2026 | Source |
| 3 | GPT-5.6 Sol | OpenAI | 94.6 / 100 | Oct 14, 2026 | Source |
| 4 | Gemini 3.1 Pro Preview | Google DeepMind | 94.3 / 100 | Oct 5, 2026 | Source |
| 5 | Claude Fable 5.1 | Anthropic | 93.7 / 100 | Sep 3, 2026 | Source |
| 6 | Kimi K3 | Moonshot AI | 93.5 / 100 | Oct 14, 2026 | Source |
| 7 | Claude Opus 5 | Anthropic | 93.2 / 100 | Sep 3, 2026 | Source |
| 8 | GPT-5.6 Terra | OpenAI | 92.9 / 100 | Oct 14, 2026 | Source |
| 9 | Qwen3.8-Max | Alibaba Qwen | 92.6 / 100 | Oct 14, 2026 | Source |
| 10 | GPT-5.6 Luna | OpenAI | 92.3 / 100 | Oct 14, 2026 | Source |
| 11 | DeepSeek V4.1 Flash | DeepSeek | 90.9 / 100 | Oct 14, 2026 | Source |
| 12 | Kimi K2.6 | Moonshot AI | 90.5 / 100 | Oct 14, 2026 | Source |
| 13 | Grok 4.3 | xAI (SpaceXAI) | 90.1 / 100 | Oct 14, 2026 | Source |
| 14 | DeepSeek V4 Pro (0813) | DeepSeek | 90.1 / 100 | Oct 14, 2026 | Source |
| 15 | Qwen3.5-397B-A17B | Alibaba Qwen | 88.4 / 100 | Oct 14, 2026 | Source |
| 16 | Nemotron 3 Ultra 550B-A55B | NVIDIA | 87.0 / 100 | Oct 14, 2026 | Source |
| 17 | Gemma 4 31B | Google DeepMind | 84.3 / 100 | Oct 14, 2026 | Source |
| 18 | Gemma 4 26B A4B | Google DeepMind | 82.3 / 100 | Oct 14, 2026 | Source |
| 19 | Nemotron 3 Nano 30B-A3B | NVIDIA | 73.0 / 100 | Oct 14, 2026 | Source |
Scores are the benchmark owners’ published results, recorded with the date we verified them. This benchmark feeds the League Score — see methodology.