GPQA Leaderboard

Graduate-level science questions, "Google-proof". · Prices checked November 17, 2026

GPQA Diamond is a set of graduate-level questions in biology, physics and chemistry, written so that searching the web does not help. It isolates reasoning from recall, which makes it the cleanest signal we track for hard analytical work.

Leaderboard

60 of 102 scored

LeaderboardScores as published by each benchmark. Prices checked November 17, 2026.

GPT-5.6 Sol leads on GPQA Diamond at 94.6%. The median across the 102 models we hold a score for is 50.5%. The cheapest model on this board is Phi-4 at $0.0910/M blended, scoring 56.1%.

#ModelProviderScoreInput /1MOutput /1MBlended 70/30Context
1GPT-5.6 SolOpenAI94.6%$4.00$20.00$8.801,050,000
2Claude Opus 4.8Anthropic93.6%$5.00$25.00$11.001,000,000
3GPT-5.6 TerraOpenAI92.9%$2.00$12.00$5.001,050,000
4GPT-5.6 LunaOpenAI92.3%$0.2000$1.20$0.50001,050,000
5Claude Opus 4.6Anthropic91.31%$5.00$25.00$11.001,000,000
6Claude Sonnet 4.6Anthropic89.9%$3.00$15.00$6.601,000,000
7Claude Opus 4.5Anthropic86.95%$5.00$25.00$11.00200,000
8Gemini 2.5 ProGoogle86.4%$1.25$10.00$3.881,000,000
9GPT-5OpenAI85.7%$1.25$10.00$3.88400,000
10Claude Sonnet 4.5Anthropic83.4%$3.00$15.00$6.60200,000
11o3OpenAI83.3%$2.00$8.00$3.80200,000
12Gemini 2.5 FlashGoogle82.8%$0.3000$2.50$0.96001,000,000
13o4-miniOpenAI81.4%$1.10$4.40$2.09200,000
14DeepSeek-R1-0528DeepSeek81%$0.5500$2.19$1.0464,000
15Claude Sonnet 3.7Anthropic78.2%$3.00$15.00$6.60200,000
16Llama-3.1-Nemotron-Ultra-253BNvidia76.01%$1.60$1.60$1.60128,000
17o1OpenAI75.7%$15.00$60.00$28.50200,000
18Grok 3xAI75.4%$3.00$15.00$6.60131,000
19Grok 3 FastxAI75.4%$5.00$25.00$11.00131,000
20o3-miniOpenAI74.9%$1.10$4.40$2.09200,000
21Claude Haiku 4.5Anthropic73%$1.00$5.00$2.20200,000
22DeepSeek-R1DeepSeek71.5%$0.5500$2.19$1.0464,000
23DeepSeek-R1 (Groq)Groq71.5%$0.7500$0.9900$0.8220128,000
24DeepSeek-R1 (Together)Together AI71.5%$3.00$7.00$4.2064,000
25DeepSeek-R1 (fw)Fireworks AI71.5%$3.00$8.00$4.5064,000
26DeepSeek-R1 (DI)DeepInfra71.5%$0.5500$2.19$1.0464,000
27DeepSeek-R1 (Cerebras)Cerebras71.5%$0.5500$0.9900$0.682064,000
28Llama 4 MaverickMeta69.8%$0.3500$1.00$0.5450128,000
29DeepSeek-V3-0324DeepSeek68.4%$0.2900$1.14$0.5450128,000
30GPT-4.1OpenAI66.3%$2.00$8.00$3.801,000,000
31Grok 3 minixAI66.2%$0.3000$0.5000$0.3600131,000
32Gemini 2.0 FlashGoogle65.2%$0.1000$0.4000$0.19001,000,000
33DeepSeek-R1-Distill-70BDeepSeek65.2%$0.3500$0.8800$0.5090128,000
34Claude Sonnet 3.5Anthropic65%$3.00$15.00$6.60200,000
35GPT-4.1 miniOpenAI65%$0.4000$1.60$0.76001,000,000
36DeepSeek-R1-Distill-32BDeepSeek62.1%$0.2000$0.5000$0.2900128,000
37Qwen2.5-MaxQwen60.1%$1.60$6.40$3.0432,000
38DeepSeek-V3DeepSeek59.1%$0.2574$1.03$0.488864,000
39DeepSeek-R1-Distill-14BDeepSeek59.1%$0.1000$0.3500$0.1750128,000
40Gemini 1.5 ProGoogle58.1%$1.25$5.00$2.382,000,000
41Llama 4 ScoutMeta57.2%$0.1000$0.3000$0.1600512,000
42Nova PremierAmazon57.1%$2.50$12.50$5.50300,000
43Phi-4Microsoft56.1%$0.0700$0.1400$0.091016,000
44Grok 2xAI56%$2.00$10.00$4.40131,000
45Grok 2 VisionxAI56%$2.00$10.00$4.4032,000
46GPT-4oOpenAI53.1%$2.50$10.00$4.75128,000
47GLM-4-PlusZhipu AI51%$7.00$7.00$7.00128,000
48GLM-4VZhipu AI51%$7.00$7.00$7.002,000
49Command ACohere50.8%$2.50$10.00$4.75256,000
50Gemini 2.0 Flash LiteGoogle50.5%$0.0750$0.3000$0.14251,000,000
51Llama 3.3 70BMeta50.5%$0.2300$0.4000$0.2810128,000
52Llama 3.3 70B (Groq)Groq50.5%$0.5900$0.7900$0.6500128,000
53Llama 3.3 70B (Together)Together AI50.5%$0.8800$0.8800$0.8800128,000
54Llama 3.3 70B (DI)DeepInfra50.5%$0.2300$0.4000$0.2810128,000
55Llama-3.3-70B (Cerebras)Cerebras50.5%$0.5900$0.9900$0.7100128,000
56GPT-4.1 nanoOpenAI50.3%$0.1000$0.4000$0.19001,000,000
57Gemini 1.5 FlashGoogle50%$0.0750$0.3000$0.14251,000,000
58Nova ProAmazon50%$0.8000$3.20$1.52300,000
59Qwen2.5-32B-InstructQwen49.5%$0.7000$0.7000$0.7000128,000
60GPT-4 TurboOpenAI49.3%$10.00$30.00$16.00128,000

58 of the 60 scores shown link to the page they were read on (underlined; hover for publisher, variant and date read). The rest have no recorded source yet and are being worked through.

Reasoning models tend to lead this table, and they also tend to be the most expensive per token, because the internal chain of thought they generate before answering is billed as output.

About these scores

  • Graduate-level science questions, "Google-proof".
  • Scores are as published by the model’s provider or a public leaderboard. Provider-reported figures are self-reported and are not independently re-run by us.
  • Where a hosted variant of an open-weights model has no score of its own, it inherits the base model’s, because the weights are the same.
  • Prices beside each score are the provider’s published token rates. How the blended rate is calculated.

Showing the top 60 of 102 models we hold a GPQA Diamond score for.

Frequently asked questions

What does GPQA measure?

Multi-step scientific reasoning at graduate level. The questions are deliberately "Google-proof": domain experts answer roughly two-thirds correctly even with unrestricted web access, so a high score reflects genuine reasoning rather than retrieval or memorisation.

Why do reasoning models score higher on GPQA?

They are trained to work through a problem step by step before answering, which is exactly what these questions require. That extra deliberation costs output tokens, so the models at the top of this table are usually also near the top of the price sheet.

Other leaderboards

Rankings that weigh price too

Leaderboards are recomputed on every deploy from the catalogue. Prices checked November 17, 2026.