GPQA Diamond
GPQA Diamond - Google-Proof Q&A benchmark for graduate-level scientific reasoning.
Category
STEM & Reasoning
Score type
percent
Scale
100%
higher is better
Models tested
57
What is GPQA Diamond?
GPQA Diamond is the 'diamond' subset of the Google-Proof Q&A benchmark — the hardest and cleanest part. The questions are written and checked by PhD scientists and are designed so that a skilled non-expert with full internet access can still barely get them right. The model has to answer graduate-level science questions with four multiple-choice options.
What does it test?
GPQA Diamond measures hard reasoning in science, technology, engineering and math. In detail, it checks:
- expert-level physics, chemistry and biology knowledge
- real reasoning instead of memorised question patterns
- resistance to plausible distractors
- reliability without web access (the questions are not Googleable)
How to read the score
The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better.
Frequently asked questions
What is the GPQA Diamond benchmark?
GPQA Diamond is the 'diamond' subset of the Google-Proof Q&A benchmark — the hardest and cleanest part. The questions are written and checked by PhD scientists and are designed so that a skilled non-expert with full internet access can still barely get them right. The model has to answer graduate-level science questions with four multiple-choice options.
What does GPQA Diamond test?
GPQA Diamond mainly measures hard reasoning in science, technology, engineering and math. Concretely, it checks: expert-level physics, chemistry and biology knowledge; real reasoning instead of memorised question patterns; resistance to plausible distractors.
How is the GPQA Diamond score calculated?
The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better. The declared maximum score of this benchmark is 100.0.
Where do I get GPQA Diamond data and scores?
The official source is https://huggingface.co/datasets/Idavidrein/gpqa. The official benchmark page/reference is linked on this page. Every score on Open LLM Radar is also available as open data: download the scores JSON (http://open-llm-radar-api:8000/api/v1/benchmarks/gpqa-diamond/scores) or pull it directly from the Open LLM Radar API with any HTTP client.
Benchmark data & downloads
Where to get, inspect and reuse the data behind this benchmark.
Official benchmark page / paper
The benchmark maintainer's website, paper or leaderboard
Benchmark dataset
The question/task data itself — download and evaluate locally
Benchmark metadata (JSON)
This page's benchmark record from the Open LLM Radar API
All model scores (JSON download)
Full score table as JSON; paginate with limit/offset params
Open LLM Radar API docs
Swagger UI to explore and reuse all radar API endpoints
Model scores 57 models tested
| # | Model | Score | Date | Source | Verified |
|---|---|---|---|---|---|
| 1 |
Kimi K3.1 Pro
Moonshot AI
|
94.2%
|
07.11.2026 | Report | Verified |
| 2 |
Kimi K3
Moonshot AI
|
93.5%
|
18.08.2026 | Report | Verified |
| 3 |
MiniMax-M3
MiniMaxAI
|
93%
|
17.06.2026 | Report | Unverified |
| 4 |
Qwen3.8-2.4T-A95B
Qwen
|
92.6%
|
18.08.2026 | Report | Verified |
| 5 |
Hy4 preview
Tencent
|
92.3%
|
28.08.2026 | Tencent Hy4 preview model card benchmark appendix (https://huggingface.co/tencent/Hy4-preview#benchmark-appendix) | Verified |
| 6 |
DeepSeek-V4.1-Flash-1017
DeepSeek
|
92.2%
|
07.11.2026 | Kimi K3.1 Pro model card comparison column (max effort) https://huggingface.co/moonshotai/Kimi-K3.1-Pro | Unverified |
| 7 |
Qwen3.8-Flash-Next
Qwen Team (Alibaba)
|
91.7%
|
26.08.2026 | Report | Verified |
| 8 |
GLM-5.3
Zhipu AI
|
91.7%
|
28.08.2026 | Tencent Hy4 preview model card benchmark appendix (https://huggingface.co/tencent/Hy4-preview#benchmark-appendix) [provider 91.7; Tencent rerun 91.4] | Verified |
| 9 |
GLM-5.2
Zhipu AI
|
91.2%
|
17.06.2026 | Report | Verified |
| 10 |
Kimi K2.6
Moonshot AI
|
91%
|
04.06.2026 | Report | Unverified |
| 11 |
Hy3
Tencent
|
90.9%
|
28.08.2026 | Tencent Hy4 preview model card benchmark appendix (https://huggingface.co/tencent/Hy4-preview#benchmark-appendix) | Verified |
| 12 |
DeepSeek-V4.1-Flash
DeepSeek
|
90.9%
|
10.09.2026 | Report | Verified |
| 13 |
DeepSeek-V4-Flash-0731
DeepSeek
|
90.8%
|
26.08.2026 | Report | Unverified |
| 14 |
DeepSeek-V4-Pro
DeepSeek
|
90.1%
|
26.06.2026 | Report | Verified |
| 15 |
Nanbeige4.3-33B
Nanbeige
|
89.5%
|
24.10.2026 | Report | Verified |
| 16 |
Qwen3.8-27B
Qwen Team (Alibaba Cloud)
|
89.2%
|
18.08.2026 | Report | Verified |
| 17 |
Ornith-1.5-35B-A3B
Ornith AI
|
89.2%
|
19.08.2026 | Report | Verified |
| 18 |
Muse-Spark-Open-816B-A42B
Meta Superintelligence Lab
|
89%
|
15.10.2026 | Report | Verified |
| 19 |
DeepSeek-V4-Flash
DeepSeek
|
88.1%
|
26.06.2026 | Report | Verified |
| 20 |
Qwen3.6-27B
Qwen Team (Alibaba Cloud)
|
87.8%
|
22.04.2026 | Qwen Team (official model card) | Verified |
| 21 |
DeepSeek-V4-Pro
DeepSeek
|
87.8%
|
04.06.2026 | Report | Unverified |
| 22 |
Kimi K2.5
Moonshot AI
|
87.6%
|
23.08.2026 | Report | Verified |
| 23 |
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16
NVIDIA
|
87%
|
04.06.2026 | Report | Verified |
| 24 |
Qwen3.5-122B-A10B
Qwen Team (Alibaba Cloud)
|
86.6%
|
01.02.2026 | Report | Verified |
| 25 |
NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB
NVIDIA
|
86.43%
|
15.10.2026 | Report | Verified |