Benchmark Registry

GPQA

GPQA results reported for 95 models across 1 version. Updated 2026-10-07.

Report an issue

Versions

1 version, plus 3 variants and configurations · Metric: GPQA Diamond Accuracy

VersionReleasedModelsResults
GPQA DiamondLatest
3 variants of Diamond
VariantConfigurationReleasedModelsResults
GPQA Diamond — Epoch AI harnessEpoch AI harnessNovember 20, 202372124
GPQA Diamond — Mercor single-shotMercor single-shot agentNovember 20, 20233839
GPQA Diamond — Vals AI harnessVals AI harness: mean of zero-shot and few-shot CoTNovember 20, 20237176
November 20, 20235165

Recently reported results: GPQA Diamond

Latest version: GPQA Diamond

ModelScoreSource
GPT-6 Astra96.0%openai.comGPQA Diamond AccuracyReported Launch postOpenAISelf-reportedRecord
MAI-Thinking-184.2%microsoft.aiGPQA Diamond AccuracyReported Technical reportMicrosoft AISelf-reportedMicrosoftRecord
Qwen3.8-Max92.6%huggingface.coGPQA Diamond AccuracyReported Model cardQwenSelf-reportedQwen TeamRecord
Gemma 4 E4B Instruct (thinking mode)58.6%huggingface.coGPQA Diamond AccuracyReported Model cardGoogleSelf-reportedGoogle DeepMindRecord
Gemma 4 26B A4B Instruct (thinking mode)82.3%huggingface.coGPQA Diamond AccuracyReported Model cardGoogleSelf-reportedGoogle DeepMindRecord