Skip to content
AIRank
en pl es fr de
← Benchmarks

GPQA

18 models ranked

GPQA is a multiple-choice benchmark built by Google DeepMind with 448 graduate-level science questions across physics, chemistry and biology. Its name means the questions are hard enough that search engines cannot answer them, which keeps the benchmark resistant to contamination. Scores reflect how often a model answers these expert questions correctly.

Scores from Hugging Face

#Model
1GPT-599
2Claude 4 Opus98
3Gemini 2.5 Pro97
4Cursor Tab96
5Llama 4 405B95
6Claude Code95
7Grok 394
8Qwen3 235B93
9DeepSeek V392
10GitHub Copilot92
11Mistral Large 391
12Claude 4 Sonnet90
13Aider89
14Command R+88
15Codeium Windsurf88
16Gemini 2.5 Flash87
17GPT-4o mini85
18Continue Dev84

AIRank aggregates these results from the public sources listed on each page. We do not run our own evaluations, and we normalize raw results to a 0–100 scale so tables are comparable.

FAQ

Frequently asked questions

What is GPQA?

GPQA is a multiple-choice benchmark built by Google DeepMind with 448 graduate-level science questions across physics, chemistry and biology. Its name means the questions are hard enough that search engines cannot answer them, which keeps the benchmark resistant to contamination. Scores reflect how often a model answers these expert questions correctly. The GPQA results shown here come from the public scores published by the benchmark’s own team and re-aggregated through sources such as Hugging Face.

How are benchmark scores computed?

Each benchmark reports its results on its own native scale — percentages, pass@1 rates or Elo points. We normalize every benchmark to a shared 0–100 scale using min-max scaling over the models we track, so tables are comparable across benchmarks. The underlying numbers always come from the public results published by the benchmark maintainers, never from proprietary or self-run evaluations.

Why do the same models rank differently across benchmarks?

Benchmarks measure different skills. GPQA tests graduate-level science knowledge, HumanEval and SWE-Bench measure coding ability, and MATH covers step-by-step problem solving. A model can be excellent at code and average at science — that is exactly why we keep a separate leaderboard for each benchmark instead of merging everything into one score.