1.google/Gemma-4-31B-IT-NVFP4
acc52.5%±3.6%
GPQA Diamond
KnowledgeScienceReasoning
GPQA Diamond is the 198-question hardest tier of GPQA, selected by two filters applied jointly: both expert annotators must agree on the correct answer, and at most one of three skilled non-expert validators with unrestricted web access could solve the question. The result is the highest-confidence "Google-proof" subset of the parent benchmark and the variant most commonly reported on frontier-model leaderboards and in research papers. It is the canonical headline number when GPQA is cited.
Source paperLatest run: 2026-05-25
Benchmark results
Switch between the canonical ranking, release-date performance view, and score-size tradeoff.
28 models
2.google/gemma-4-31B-it
acc52.0%±3.6%
3.google/gemma-4-26B-A4B-it
acc41.4%±3.5%
4.google/gemma-3-27b-it
acc39.4%±3.5%
5.Qwen/Qwen3-Next-80B-A3B-Instruct
acc38.9%±3.5%
6.microsoft/phi-4
acc35.9%±3.4%
7.google/gemma-3-12b-it
acc35.9%±3.4%
8.microsoft/phi-4-mini-instruct
acc33.8%±3.4%
9.Qwen/Qwen3-4B-AWQ
acc32.3%±3.3%
10.Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8
acc30.8%±3.3%
11.nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4
acc29.8%±3.3%
12.Qwen/Qwen3.5-35B-A3B
acc29.8%±3.3%
13.Qwen/Qwen3-32B
acc29.3%±3.2%
14.Qwen/Qwen3-14B
acc29.3%±3.2%
15.openai/gpt-oss-120b
acc28.8%±3.2%
16.Qwen/Qwen3.6-27B
acc28.8%±3.2%
17.MiniMax/MiniMax-M2.1-AWQ
acc28.8%±3.2%
18.Qwen/Qwen3.5-122B-A10B-NVFP4
acc28.3%±3.2%
19.zai-org/GLM-4.5V-FP8
acc28.3%±3.2%
20.Qwen/Qwen3-8B
acc27.8%±3.2%
21.microsoft/phi-4-reasoning-plus
acc27.8%±3.2%
22.Qwen/Qwen3-4B
acc26.8%±3.2%
23.openai/gpt-oss-20b
acc26.8%±3.2%
24.Qwen/Qwen3.6-35B-A3B
acc26.3%±3.1%
25.MiniMax/MiniMax-M2-AWQ
acc25.8%±3.1%
26.microsoft/phi-4-mini-reasoning
acc25.8%±3.1%
27.Qwen/Qwen3-235B-A22B-Thinking-AWQ-2507
acc25.3%±3.1%
28.zai-org/GLM-4.5-Air-FP8
acc24.2%±3.1%
Caveats
At N=198 the variance is severe. A swing of one correctly-answered question is approximately 0.5 percentage points, so reported gaps under roughly 3pp should be treated as within sampling noise unless they are accompanied by confidence intervals or multi-seed runs. Many published comparisons do not provide either, so cross-paper rankings on GPQA Diamond should be read with caution. The expert-agreement filter that produces the Diamond tier may also introduce a selection effect: by requiring two experts to agree, the subset may favor questions with relatively unambiguous textbook answers and exclude genuinely frontier-research edge cases where experts legitimately disagree. Contamination risk is the same as GPQA Main since the Diamond split was published simultaneously, and top reasoning models now exceed 85%, approaching the ~74% adjusted expert ceiling and signalling imminent saturation.
How to cite
Citation
FrozeBench. "GPQA Diamond." https://frozebench.com/benchmarks/gpqa-diamond. Retrieved 2026-06-04.
BibTeX
@misc{frozebench_gpqa_diamond,
title = {GPQA Diamond},
howpublished = {\url{https://frozebench.com/benchmarks/gpqa-diamond}},
year = {2026},
note = {FrozeBench. Retrieved 2026-06-04.}
}URL
https://frozebench.com/benchmarks/gpqa-diamond