AI benchmarks and leaderboards are different things
Learn the difference between a benchmark, an evaluation run and a leaderboard, and use each to answer the right question.
Model reports
GPQA Diamond
198 questions in the Diamond split, so run-to-run variance is wide and a single reported number hides it. Approaching saturation at the 2026 frontier, where reported scores cluster in the high 80s and 90s.
Evaluation context and evidence
- Source
- Model reports
- Publisher
- Not recorded
- Benchmark release
- 2023-11-20
- Collected
- Not recorded
- Modality
- Not recorded
These are source-reported observations, not independent reruns by this site. Check the task version, evaluation protocol, unit and reporting date before comparing models. Results from different setups are not a universal ranking.
Original sources
- Original source
- paper
- Claude 3.7 Sonnet
- Claude Opus 4 and Sonnet 4
- DeepSeek-R1
- DeepSeek-V3
- DeepSeek-V4-Pro
- DeepSeek-V4
- Gemini 2.5 Pro
- Gemini 3.1 Pro
- Gemma 4 (31B)
- Llama 4
- Mistral Large 3 (675B)
- Mistral Medium 3
- Kimi K3
- GPT-4.1
- GPT-5.6
- GPT-5
- o3 and o4-mini
- Qwen3.5-397B-A17B
- Qwen3.8-27B
- Qwen3
- Hy4 preview
- Grok 4.5
- Grok 4
- GLM-5.1
- GLM-5.2
- GLM-5
- GLM-5
Reported results 21
Shown in source observation order, not ranked. Dates retain the precision provided by the source.
| Model | Value | Unit | Protocol | Reported date and basis | Source |
|---|---|---|---|---|---|
| Claude Opus 4 | 79.6 | percent | extended thinking up to 64K tokens, no parallel test-time compute | 2025-05-22 · document_publication | Claude Opus 4 and Sonnet 4 |
| Claude Sonnet 4 | 75.4 | percent | extended thinking up to 64K tokens, no parallel test-time compute | 2025-05-22 · document_publication | Claude Opus 4 and Sonnet 4 |
| DeepSeek-R1 | 71.5 | percent | Pass@1, averaged over k=64 samples, temperature 0.6, top-p 0.95, 32K output limit | 2025-01-22 · document_publication | DeepSeek-R1 |
| DeepSeek-V3 | 59.1 | percent | Pass@1, 8K output limit | 2024-12-27 · document_publication | DeepSeek-V3 |
| DeepSeek-V4-Pro | 72.9 | percent | non-think, pass@1 | 2026-04-22 · document_publication | DeepSeek-V4-Pro |
| DeepSeek-V4-Pro | 89.1 | percent | think high, pass@1 | 2026-04-22 · document_publication | DeepSeek-V4-Pro |
| DeepSeek-V4-Pro | 90.1 | percent | think max, pass@1 | 2026-04-22 · document_publication | DeepSeek-V4-Pro |
| DeepSeek-V4-Flash | 88.1 | percent | think max, pass@1 | 2026-06-25 · document_publication | DeepSeek-V4 |
| Claude 4 Opus | 79.6 | percent | single attempt, pass@1 | 2025-06-17 · document_publication | Gemini 2.5 Pro |
| Claude 4 Sonnet | 75.4 | percent | single attempt, pass@1 | 2025-06-17 · document_publication | Gemini 2.5 Pro |
| Gemini 2.5 Pro | 86.4 | percent | single attempt, pass@1 | 2025-06-17 · document_publication | Gemini 2.5 Pro |
| Grok 3 Beta Extended Thinking | 80.2 | percent | single attempt, pass@1 | 2025-06-17 · document_publication | Gemini 2.5 Pro |
| o3 high | 83.3 | percent | single attempt, pass@1 | 2025-06-17 · document_publication | Gemini 2.5 Pro |
| Gemini 3.1 Pro | 94.3 | percent | Thinking (High), No tools | 2026-02-19 · document_publication | Gemini 3.1 Pro |
| Gemma 4 (31B) | 84.3 | percent | as reported | 2026-03-11 · document_publication | Gemma 4 (31B) |
| Kimi K3 | 93.5 | percent | reasoning=max, top-p 0.95 | 2026-06-13 · document_publication | Kimi K3 |
| Qwen3-235B-A22B (Thinking) | 71.1 | percent | averaged over 10 samples | 2025-05-14 · document_publication | Qwen3 |
| Hy4 preview | 92.3 | percent | as printed in the report's benchmark table | 2026-08-28 · document_publication | Hy4 preview |
| GLM-5.1 | 86.2 | percent | as reported | 2026-04-03 · document_publication | GLM-5.1 |
| GLM-5.2 | 91.2 | percent | reasoning-tasks settings, temp 1.0, top_p 0.95, max gen 163840 tokens | 2026-06-16 · document_publication | GLM-5.2 |
| GLM-5 | 86.0 | percent | reasoning-tasks settings, temp 1.0, top_p 0.95 | 2026-02-11 · document_publication | GLM-5 |