Benchmark
GPQA Diamond Benchmark
What is 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…
- Released
- 2023-11-20
- Evaluates
- science
- Openness
- unknown
- Importer
- Model reports
- Reported scores
- 33
Source provenance
- Original evidence https://arxiv.org/abs/2311.12022
GPQA Diamond paper
GPQA Diamond results and reported scores
Model reports
| Model | Organization | Reported value | Reported | Evidence |
|---|---|---|---|---|
| Bespoke-Stratos-7B | Bespoke Labs | 37.8 | 2025-04-30 | evidence |
| Claude 4 Opus | Anthropic | 79.6 | 2025-06-17 | evidence |
| Claude 4 Sonnet | Anthropic | 75.4 | 2025-06-17 | evidence |
| Claude Opus 4 | Anthropic | 79.6 | 2025-05-22 | evidence |
| Claude Sonnet 4 | Anthropic | 75.4 | 2025-05-22 | evidence |
| DeepSeek-R1 | DeepSeek | 71.5 | 2025-01-22 | evidence |
| DeepSeek-R1-Distill-Llama-8B | DeepSeek | 47.3 | 2025-04-30 | evidence |
| DeepSeek-R1-Distill-Qwen-7B | DeepSeek | 49.5 | 2025-04-30 | evidence |
| DeepSeek-V3 | DeepSeek | 59.1 | 2024-12-27 | evidence |
| DeepSeek-V4-Flash | DeepSeek | 88.1 | 2026-06-25 | evidence |
| DeepSeek-V4-Pro | DeepSeek | 72.9 | 2026-04-22 | evidence |
| DeepSeek-V4-Pro | DeepSeek | 89.1 | 2026-04-22 | evidence |
| DeepSeek-V4-Pro | DeepSeek | 90.1 | 2026-04-22 | evidence |
| DeepSeek-V4.1-Flash | DeepSeek | 90.9 | 2026-09-10 | evidence |
| Gemini 2.5 Pro | 86.4 | 2025-06-17 | evidence | |
| Gemini 3.1 Pro | 94.3 | 2026-02-19 | evidence | |
| Gemma 4 (31B) | 84.3 | 2026-03-11 | evidence | |
| GLM-5 | Z.ai | 86.0 | 2026-02-11 | evidence |
| GLM-5.1 | Z.ai | 86.2 | 2026-04-03 | evidence |
| GLM-5.2 | Z.ai | 91.2 | 2026-06-16 | evidence |
| Grok 3 Beta Extended Thinking | xAI | 80.2 | 2025-06-17 | evidence |
| Hy4 preview | Tencent | 92.3 | 2026-08-28 | evidence |
| Kimi K3 | Moonshot AI | 93.5 | 2026-06-13 | evidence |
| Llama-3.2-3B-Instruct | Meta | 25.3 | 2025-04-30 | evidence |
| o1-mini | OpenAI | 60.0 | 2025-04-30 | evidence |
| o3 high | OpenAI | 83.3 | 2025-06-17 | evidence |
| OpenThinker-7B | OpenThoughts | 42.4 | 2025-04-30 | evidence |
| Phi-4-Mini (base model, 3.8B) | Microsoft | 36.9 | 2025-04-30 | evidence |
| Phi-4-mini-reasoning (3.8B) | Microsoft | 52.0 | 2025-04-30 | evidence |
| Qwen3-235B-A22B (Thinking) | Qwen | 71.1 | 2025-05-14 | evidence |
| Seed2.0 Lite | ByteDance | 66.1 | 2026-02-14 | evidence |
| Seed2.0 Mini | ByteDance | 58.1 | 2026-02-14 | evidence |
| Seed2.0 Pro | ByteDance | 88.9 | 2026-02-14 | evidence |
Scores are partitioned by the source that reported them and are never merged into a single cross-source ranking, because the sources measure different things and say so.
Which sources cite GPQA Diamond?
- DeepSeek-V4.1-Flash DeepSeek, 2026-09-10
- Hy4 preview Tencent, 2026-08-28
- Qwen3.8-27B Qwen, 2026-08-05
- Grok 4.5 xAI, 2026-07-08
- GPT-5.6 OpenAI, 2026-06-26
- DeepSeek-V4 DeepSeek, 2026-06-25
- GLM-5.2 Z.ai, 2026-06-16
- Kimi K3 Moonshot AI, 2026-06-13
- DeepSeek-V4-Pro DeepSeek, 2026-04-22
- GLM-5.1 Z.ai, 2026-04-03
- Gemma 4 (31B) Google, 2026-03-11
- Gemini 3.1 Pro Google, 2026-02-19
30 source documents cite this benchmark.