// BENCHMARK

GPQA Diamond benchmark

AI model leaderboard for the GPQA Diamond benchmark. Compare how large language models score on GPQA Diamond, see the full ranking, and understand what this AI benchmark measures. Sakana Fugu currently leads with 95.5%. 198 hardest expert-validated graduate-level science questions resistant to web search.

Leaderboard

# Model Organization Score Variant Source
#1 Sakana Fugu Sakana AI 95.5% official ↗
#2 Gemini 3 Pro Google DeepMind 91.9% cited-deepseek-v3-2 official ↗
#3 GPT-5.5 Pro OpenAI 91.2% parallel-compute official ↗
#4 Grok 4.3 xAI 90.1% official ↗
#5 GPT-5.5 OpenAI 88.4% no-tools official ↗
#6 Grok 4 xAI 87.7% official ↗
#7 Claude Opus 4.7 Anthropic 87.1% extended-thinking official ↗
#8 Gemini 3.1 Pro Google DeepMind 87% official ↗
#9 Claude Opus 4.5 Anthropic 87% official ↗
#10 gemini-2.5-pro Google DeepMind 86.4% cited-ring-1t official ↗
#11 GPT-5 OpenAI 86.05% thinking-high-cited-ring-1t official ↗
#12 GPT-5 OpenAI 85.7% cited-deepseek-v3-2 official ↗
#13 GPT-5.4 Thinking OpenAI 85.7% thinking official ↗
#14 Claude Opus 4.6 Anthropic 85.5% extended-thinking official ↗
#15 DeepSeek V4-Pro DeepSeek 85% thinking official ↗
#16 Qwen3.6-Max-Preview Qwen 84.7% official ↗
#17 Kimi K2 Thinking Moonshot 84.5% thinking-no-tools official ↗
#18 Kimi K2 Thinking Moonshot 84.5% thinking-cited-deepseek-v3-2 official ↗
#19 gemini-2.5-pro Google DeepMind 84.4% official ↗
#20 MAI-Thinking-1 Microsoft 84.2% official ↗
#21 gemini-2.5-pro Google DeepMind 84% cited-phi4r official ↗
#22 Claude Sonnet 4.5 Anthropic 83.4% cited-deepseek-v3-2 official ↗
#23 o3 OpenAI 82.7% official ↗
#24 DeepSeek-V3.2 DeepSeek 82.4% thinking official ↗
#25 Claude Sonnet 4.6 Anthropic 81.4% extended-thinking official ↗
#26 DeepSeek-R1-0528 DeepSeek 81.3% official ↗
#27 Qwen3-235B-A22B-Thinking (Jul 2025) Qwen 81.1% official ↗
#28 Qwen3-235B-A22B-Thinking (Jul 2025) Qwen 81.1% thinking-cited-ring-1t official ↗
#29 DeepSeek-V3.1-Terminus DeepSeek 81% thinking-cited-ring-1t official ↗
#30 GPT-5.4 OpenAI 81% official ↗
#31 DeepSeek-R1-0528 DeepSeek 81% reasoning official ↗
#32 gpt-oss-120b OpenAI 80.9% high-with-tools official ↗
#33 gpt-oss-120b OpenAI 80.1% high-no-tools official ↗
#34 gpt-oss-120b OpenAI 80.1% high reasoning, no tools official ↗
#35 DeepSeek-V3.2-Exp DeepSeek 79.9% from-hf-readme official ↗
#36 Claude Opus 4 Anthropic 79.6% official ↗
#37 DeepSeek-V3.1 DeepSeek 79.46% official ↗
#38 Trinity-Large-Thinking Arcee AI 79.4% official ↗
#39 GLM 4.5 Zhipu AI 79.1% official ↗
#40 Gemma 4 12B Google DeepMind 78.8% official ↗
#41 Ring-1T Ant Group 78.63% thinking official ↗
#42 DeepSeek V4-Flash DeepSeek 78.6% thinking official ↗
#43 Qwen3-Max-Thinking Qwen 78.4% thinking official ↗
#44 o3-mini OpenAI 77.7% high-cited-phi4r official ↗
#45 MiniMax-M2 MiniMax 77.7% cited-deepseek-v3-2 official ↗
#46 o1 OpenAI 77.3% cited-s1 official ↗
#47 gpt-oss-120b OpenAI 77.04% official ↗
#48 Claude 3.7 Sonnet Anthropic 76.8% thinking-cited-phi4r official ↗
#49 o1 OpenAI 76.7% cited-phi4r official ↗
#50 DeepSeek-V3.1-Terminus DeepSeek 76.23% non-thinking-cited-ling-1t official ↗
#51 Muse Spark Meta AI 76.2% official ↗
#52 Llama Nemotron Ultra 253B NVIDIA 76.01% reasoning-on-pass@1 official ↗
#53 EXAONE 4.0 (32B) LG AI Research 75.4% reasoning official ↗
#54 GLM-4.5-Air Zhipu AI 75% official ↗
#55 OpenReasoning-Nemotron-32B NVIDIA 74.98% official ↗
#56 gpt-oss-20b OpenAI 74.2% high-with-tools official ↗
#57 Kimi K2 Moonshot 74.2% k2-0905-no-tools-cited-k2-thinking official ↗
#58 Kimi K2 Moonshot 73.93% instruct-0905-cited-ling-1t official ↗
#59 Pangu Pro MoE Huawei 73.7% instruct-pass@1 official ↗
#60 DeepSeek-R1-Zero DeepSeek 73.3% cited-magistral official ↗
#61 LongCat-Flash Meituan Inc 73.23% official ↗
#62 Ling-1T Ant Group 72.98% pass@1 official ↗
#63 Llama Nemotron Super 49B NVIDIA 71.97% reasoning-on-pass@1-avg@4 official ↗
#64 gemini-2.5-pro Google DeepMind 71.81% lowthink-cited-ling-1t official ↗
#65 gpt-oss-20b OpenAI 71.5% high reasoning, no tools official ↗
#66 gpt-oss-20b OpenAI 71.5% high-no-tools official ↗
#67 DeepSeek-R1 DeepSeek 71.5% cited-magistral official ↗
#68 DeepSeek-R1 DeepSeek 71.5% cited-s1 official ↗
#69 GPT-5 OpenAI 71.31% main-cited-ling-1t official ↗
#70 Qwen3-235B-A22B Qwen 71.1% thinking official ↗
#71 Qwen3-235B-A22B-Thinking (Jul 2025) Qwen 71.1% reasoning official ↗
#72 K2 Think Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) 71.08% avg@16 official ↗
#73 DeepSeek-R1 DeepSeek 71.08% official ↗
#74 Magistral Medium 1.1 Mistral AI 70.8% official ↗
#75 MiniMax-M1-80k MiniMax 70% 32 samples official ↗
#76 Llama 4 Maverick Meta AI 69.8% official ↗
#77 Qwen3-32B Qwen 69.7% thinking-cited-ntele-readme official ↗
#78 Phi-4-Reasoning-plus Microsoft 69.3% pass@1-avg@5 official ↗
#79 MiniMax-M1-40k MiniMax 69.2% 32 samples official ↗
#80 Phi-4-Reasoning-plus Microsoft 68.9% reasoning official ↗
#81 Qwen3-32B Qwen 68.4% reasoning official ↗
#82 DeepSeek-V3-0324 DeepSeek 68.4% official ↗
#83 Qwen3-32B Qwen 68.4% thinking official ↗
#84 Magistral Small 1.1 Mistral AI 68.2% official ↗
#85 NTele-R1-32B-V1 ZTE 67.17% from-hf-readme official ↗
#86 Phi-4-Reasoning Microsoft 67.1% pass@1-avg@5 official ↗
#87 QwQ-32B Qwen 66.24% official ↗
#88 DeepSeek-R1-Distill-Llama-70B DeepSeek 66.2% cited-phi4r official ↗
#89 EXAONE Deep 32B LG AI Research 66.1% pass@1-from-hf-readme official ↗
#90 EXAONE Deep 32B LG AI Research 66.1% cited-phi4r official ↗
#91 Qwen3-30B-A3B Qwen 65.8% thinking official ↗
#92 Qwen3-235B-A22B Qwen 65.55% official ↗
#93 gpt-oss-20b OpenAI 65.45% official ↗
#94 DeepSeek-R1-Distill-Llama-70B DeepSeek 65.2% pass@1-cited-exaone official ↗
#95 Claude 3.5 Sonnet Anthropic 65% 1022-cited-mimo official ↗
#96 Claude 3.5 Sonnet Anthropic 65% 1022-diamond-cited-minimax-text official ↗
#97 NVIDIA-Nemotron-Nano-12B-v2 NVIDIA 64.48% from-hf-readme official ↗
#98 NVIDIA-Nemotron-Nano-9B-v2 NVIDIA 64% from-hf-readme official ↗
#99 Qwen3-14B Qwen 64% thinking official ↗
#100 s1 Stanford University 63.6% s1-1-from-hf-readme official ↗
#101 QwQ-32B Qwen 63.6% cited-ntele-readme official ↗
#102 QwQ-32B Qwen 63.3% pass@1-cited-exaone official ↗
#103 Qwen3-235B-A22B Qwen 62.9% non-thinking official ↗
#104 EXAONE Deep 7.8B LG AI Research 62.6% pass@1-from-hf-readme official ↗
#105 DeepSeek-R1-Distill-Qwen-32B DeepSeek 62.1% pass@1-cited-exaone official ↗
#106 DeepSeek-R1-Distill-Qwen-32B Qwen 62.1% pass@1-cited-exaone official ↗
#107 Kimi Linear Moonshot 62.1% sft-avg@8 official ↗
#108 Gemini 2.0 Flash Google DeepMind 62.1% exp-diamond-cited-minimax-text official ↗
#109 DeepSeek-R1-Distill-Qwen-32B DeepSeek 62.1% cited-s1 official ↗
#110 DeepSeek-R1-Distill-Qwen-32B Qwen 62.1% cited-ntele-readme official ↗
#111 DeepSeek-R1-Distill-Qwen-32B Qwen 62.1% cited-s1 official ↗
#112 DeepSeek-R1-Distill-Qwen-32B DeepSeek 62.1% cited-ntele-readme official ↗
#113 Qwen3-8B Qwen 62% thinking official ↗
#114 o1-mini OpenAI 60% cited-phi4r official ↗
#115 Mistral Medium 3 Mistral AI 59.6% cited-magistral official ↗
#116 s1 Stanford University 59.6% s1-from-hf-readme official ↗
#117 QwQ-32B Qwen 59.5% cited-phi4r official ↗
#118 Gemini 1.5 Pro Google DeepMind 59.1% 002-diamond-cited-minimax-text official ↗
#119 DeepSeek-V3 DeepSeek 59.1% diamond-cited-minimax-text official ↗
#120 DeepSeek-R1-Distill-Qwen-14B DeepSeek 59.1% cited-mimo official ↗
#121 DeepSeek-R1-Distill-Qwen-14B Qwen 59.1% cited-mimo official ↗
#122 DeepSeek-V3 DeepSeek 59.1% cited-magistral official ↗
#123 Qwen3-30B-A3B Qwen 58.91% official ↗
#124 Gemma 4 31B Google DeepMind 58.3% official ↗
#125 Llama Nemotron Ultra 253B NVIDIA 56.6% reasoning-off-pass@1 official ↗
#126 DeepSeek-V3 DeepSeek 56.1% pass@1 official ↗
#127 Gemma 4 26B Google DeepMind 55.9% official ↗
#128 Qwen3-4B Qwen 55.9% thinking official ↗
#129 Qwen3-14B Qwen 54.8% non-thinking official ↗
#130 Qwen3-30B-A3B Qwen 54.8% non-thinking official ↗
#131 Qwen3-32B Qwen 54.6% non-thinking official ↗
#132 MiniMax-Text-01 MiniMax 54.4% diamond-from-hf-readme official ↗
#133 EXAONE Deep 2.4B LG AI Research 54.3% pass@1-from-hf-readme official ↗
#134 Llama Nemotron Nano 8B NVIDIA 54.1% reasoning-on-pass@1 official ↗
#135 Llama 4 Scout Meta AI 53.5% pass@1-cited-pangu-pro-moe official ↗
#136 dots.llm1 Rednote 52.6% pass@1 official ↗
#137 EXAONE 4.0 (1.2B) LG AI Research 52% reasoning official ↗
#138 Llama 3.1-405B Meta AI 50.7% instruct-diamond-cited-minimax-text official ↗
#139 DeepSeek-R1-Distill-Qwen-7B DeepSeek 49.1% cited-mimo official ↗
#140 DeepSeek-R1-Distill-Qwen-7B Qwen 49.1% cited-mimo official ↗
#141 Qwen2.5 Instruct (72B) Qwen 49% instruct-diamond-cited-minimax-text official ↗
#142 DeepSeek-R1-Distill-Llama-8B DeepSeek 49% pass@1-cited-exaone official ↗
#143 Mistral Small 3.2 Mistral AI 46.13% instruct-5-shot-cot official ↗
#144 GPT-4o (Nov 2024) OpenAI 46% diamond-cited-minimax-text official ↗
#145 Mistral Small 3.1 Mistral AI 45.96% instruct-5-shot-cot official ↗
#146 Gemma 3 27B Google DeepMind 42.4% official ↗
#147 Qwen3-4B Qwen 41.7% non-thinking official ↗
#148 GPT-4o mini OpenAI 41.1% cited-phi35moe official ↗
#149 Granite-4.0-H-Small IBM 40.63% 0-shot-cot official ↗
#150 Qwen3-1.7B Qwen 40.1% thinking official ↗
#151 Llama Nemotron Nano 8B NVIDIA 39.4% reasoning-off-pass@1 official ↗
#152 Qwen3-8B Qwen 39.3% non-thinking official ↗
#153 Gemini 1.5 Flash (Sep 2024) Google DeepMind 37.5% cited-phi35moe official ↗
#154 Phi-3.5-MoE Microsoft 36.8% instruct-0-shot-cot official ↗
#155 Qwen2-57B-A14B Qwen 34.3% base-from-hf-readme official ↗
#156 DeepSeek-R1-Distill-Qwen-1.5B Qwen 33.8% pass@1-cited-exaone official ↗
#157 DeepSeek-R1-Distill-Qwen-1.5B DeepSeek 33.8% pass@1-cited-exaone official ↗
#158 Llama 3.1-8B Meta AI 33.5% instruct-0-shot-cited-falcon3 official ↗
#159 Granite-4.0-H-Tiny IBM 32.59% 0-shot-cot official ↗
#160 Granite-4.0-H-Micro IBM 32.15% 0-shot-cot official ↗
#161 Qwen2.5-7B Qwen 32% instruct-0-shot-cited-falcon3 official ↗
#162 Falcon3-7B Technology Innovation Institute 31.9% instruct-0-shot official ↗
#163 Qwen2-7B Qwen 31.8% base-from-hf-readme official ↗
#164 Qwen1.5-32B Qwen 30.8% cited-qwen2-57b official ↗
#165 Gemma 2 9B Google DeepMind 29.2% instruct-cited-phi35moe official ↗
#166 Mixtral 8x7B Mistral AI 29.2% cited-qwen2-57b official ↗
#167 Qwen3-1.7B Qwen 28.6% non-reasoning official ↗
#168 Qwen3-1.7B Qwen 28.6% non-thinking official ↗
#169 Mistral NeMo Mistral AI 28.6% instruct-cited-phi35moe official ↗
#170 Qwen3-0.6B Qwen 27.9% thinking official ↗
#171 Qwen1.5-7B Qwen 26.7% base-cited-qwen2-7b official ↗
#172 Llama 3.1-8B Meta AI 26.3% instruct-cited-phi35moe official ↗
#173 Llama 3-8B Meta AI 25.8% base-cited-qwen2-7b official ↗
#174 Gemma 7B Google DeepMind 25.7% base-cited-qwen2-7b official ↗
#175 DCLM 7B Apple 24.75% diamond-from-hf-readme official ↗
#176 Mistral 7B Mistral AI 24.7% base-cited-qwen2-7b official ↗
#177 Qwen3-0.6B Qwen 22.9% non-thinking official ↗
#178 Qwen3-0.6B Qwen 22.9% non-reasoning official ↗
#179 Gemma 3 1B Google DeepMind 19.2% non-reasoning official ↗
#180 Falcon Mamba Technology Innovation Institute 8.05% 0-shot-from-hf-readme official ↗
#181 Llama 3-8B Meta AI 7.38% cited-falcon-mamba official ↗
#182 Llama 3.1-8B Meta AI 6.15% base-cited-falcon-mamba official ↗
#183 Mistral 7B Mistral AI 5.59% v0-1-cited-falcon-mamba official ↗
#184 Gemma 7B Google DeepMind 4.92% cited-falcon-mamba official ↗

Frequently asked questions about GPQA Diamond

What is the GPQA Diamond benchmark?

198 hardest expert-validated graduate-level science questions resistant to web search.

How is the GPQA Diamond benchmark scored?

GPQA Diamond is scored using the accuracy (%) metric, where a higher score is better. The maximum achievable score is 100.000. GenAIList aggregates reported scores from model providers and papers into a single ranked leaderboard.

Which AI model scores highest on GPQA Diamond?

As of the latest reported scores on GenAIList, Sakana Fugu achieves the highest result on GPQA Diamond with a score of 95.5%.

Is a higher GPQA Diamond score better?

Yes. On GPQA Diamond a higher score indicates better performance, so models near the top of the leaderboard are the strongest.