GPQA Diamond gpqa Leaderboard
Graduate-level Google-Proof Q&A across biology, physics, chemistry - PhD-grade questions. Secondary board (demoted from core per the audit; HLE is the frontier successor as GPQA approaches saturation). · Metric: Accuracy (higher is better) · 🟢 Updated 17h ago
| # | Model | Accuracy | Paper |
|---|---|---|---|
| 1 | Gemini 3.1 Pro Preview | 94.10 | — |
| 2 | GPT-5.6 Sol (max) | 94.10 | — |
| 3 | GPT-5.5 (xhigh) | 93.50 | — |
| 4 | Kimi K3 | 93.50 | — |
| 5 | GPT-5.5 (high) | 93.20 | — |
| 6 | GPT-5.6 Sol (xhigh) | 93.10 | — |
| 7 | Grok 4.5 (high) | 93.10 | — |
| 8 | MiniMax-M3 | 92.90 | — |
| 9 | GPT-5.6 Sol (high) | 92.80 | — |
| 10 | Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback) | 92.60 | — |
| 11 | GPT-5.5 (medium) | 92.60 | — |
| 12 | GPT-5.6 Sol (medium) | 92.60 | — |
| 13 | GPT-5.6 Terra (max) | 92.50 | — |
| 14 | Qwen3.7 Max | 92.30 | — |
| 15 | Gemini 3.5 Flash (high) | 92.20 | — |
| 16 | Gemini 3.5 Flash (medium) | 92.10 | — |
| 17 | Claude Opus 4.8 (Adaptive Reasoning, Max Effort) | 92.00 | — |
| 18 | GPT-5.4 (xhigh) | 92.00 | — |
| 19 | GPT-5.3 Codex (xhigh) | 91.50 | — |
| 20 | Claude Opus 4.7 (Adaptive Reasoning, Max Effort) | 91.40 | — |
| 21 | Claude Sonnet 5 (Adaptive Reasoning, Max Effort) | 91.10 | — |
| 22 | GPT-5.6 Luna (max) | 91.10 | — |
| 23 | Grok 4.20 0309 v2 (Reasoning) | 91.10 | — |
| 24 | Kimi K2.6 | 91.10 | — |
| 25 | GPT-5.5 (low) | 91.00 | — |
| 26 | Gemini 3 Pro Preview (high) | 90.80 | — |
| 27 | GPT-5.6 Terra (xhigh) | 90.80 | — |
| 28 | DeepSeek V4 Pro (Reasoning, High Effort) | 90.50 | — |
| 29 | GPT-5.2 (xhigh) | 90.30 | — |
| 30 | Grok 4.3 (high) | 90.10 | — |
| 31 | SuperGPQA: Scaling LLM Evaluation across 285 Graduate Disciplines | 61.82 | — |
| 32 | AutoTrainess: Teaching Language Models to Improve Language Models Autonomously | 35.49 | link |
| 33 | When Does Combining Language Models Help? A Co-Failure Ceiling on Routing, Voting, and Mixture-of-Agents Across 67 Frontier Models | 0.13 | link |