Benchmark profile

Graduate-Level Google-Proof Q&A (GPQA)

A challenging dataset of 448 multiple-choice questions written by domain experts in biology, physics, and chemistry. Designed to be difficult even for skilled non-experts with access to Google.

Data verified July 20, 2026

Sakana Fugu-Ultra leads the GPQA leaderboard on BenchLM's July 2026 update with 95.5%, ahead of Sakana Fugu (95.5%) and GPT-5.6 Sol (94.6%), across 71 tracked models.

Top models on GPQA — July 20, 2026

As of July 20, 2026, Sakana Fugu-Ultra leads the GPQA leaderboard with 95.5% , followed by Sakana Fugu (95.5%) and GPT-5.6 Sol (94.6%).

1Closed

Sakana Fugu-Ultra

Sakana AI

sakana-fugu-ultra

95.5%

Overall —Context 1M

2Closed

Sakana Fugu

Sakana AI

sakana-fugu

95.5%

Overall —Context 1M

3Closed

GPT-5.6 Sol

OpenAI

gpt-5-6-sol

94.6%

Overall 81.96Context 1M

71 modelsKnowledge7% of category scoreRefreshingUpdated July 20, 2026

Leaderboard (71 models)

Score

Sakana Fugu-UltraSakana AI · Closed

95.5%

Sakana FuguSakana AI · Closed

95.5%

GPT-5.6 SolOpenAI · Closed

94.6%

Claude Opus 4.7 (Adaptive)Anthropic · Closed

94.2%

Claude Mythos 5Anthropic · Closed

94.1%

Claude Opus 4.8Anthropic · Closed

93.6%

GPT-5.5OpenAI · Closed

93.6%

Kimi K3Moonshot AI · Closed

93.5%

GPT-5.6 TerraOpenAI · Closed

92.9%

GPT-5.4OpenAI · Closed

92.8%

Qwen3.7 MaxAlibaba · Closed

92.4%

GPT-5.2OpenAI · Closed

92.4%

GPT-5.6 LunaOpenAI · Closed

92.3%

Gemini 3.5 FlashGoogle · Closed

92.2%

Claude Opus 4.6Anthropic · Closed

91.3%

GLM-5.2Z.AI · Open weight

91.2%

Kimi K2.6Moonshot AI · Open weight

90.5%

Qwen3.6 PlusAlibaba · Closed

90.4%

Qwen3.7 PlusAlibaba · Closed

90.3%

DeepSeek V4 Pro (Max)DeepSeek · Open weight

90.1%

Grok 4.3xAI · Closed

90.1%

Claude Sonnet 4.6Anthropic · Closed

89.9%

Interfaze BetaInterfaze · Closed

89.9%

DeepSeek V4 Pro (High)DeepSeek · Open weight

89.1%

Qwen3.5 397BAlibaba · Open weight

88.4%

DeepSeek V4 Flash (Max)DeepSeek · Open weight

88.1%

GPT-5.4 miniOpenAI · Closed

88%

InklingThinking Machines Lab · Open weight

87.9%

Qwen3.6-27BAlibaba · Open weight

87.8%

Kimi K2.5Moonshot AI · Open weight

87.6%

Kimi K2.5 (Reasoning)Moonshot AI · Closed

87.6%

DeepSeek V4 Flash (High)DeepSeek · Open weight

87.4%

Hy3 PreviewTencent · Open weight

87.2%

Claude Opus 4.5Anthropic · Closed

87%

Nemotron 3 UltraNVIDIA · Open weight

87%

Qwen3.5-122B-A10BAlibaba · Open weight

86.6%

GLM-5Z.AI · Open weight

86%

Qwen3.6-35B-A3BAlibaba · Open weight

86%

GLM-4.7Z.AI · Open weight

85.7%

Qwen3.5-27BAlibaba · Open weight

85.5%

Gemma 4 31BGoogle · Open weight

84.3%

MAI-Thinking-1Microsoft · Closed

84.2%

Qwen3.5-35B-A3BAlibaba · Open weight

84.2%

MiMo-V2-FlashXiaomi · Open weight

83.7%

Claude Sonnet 4.5Anthropic · Closed

83.4%

Gemini 2.5 ProGoogle · Closed

83%

GPT-5.4 nanoOpenAI · Closed

82.8%

o1-proOpenAI · Closed

79%

Gemma 4 12BGoogle · Open weight

78.8%

Qwen3 235B 2507Alibaba · Open weight

77.5%

o3-miniOpenAI · Closed

77.2%

o1OpenAI · Closed

75.7%

DeepSeek V4 ProDeepSeek · Open weight

72.9%

Nemotron 3 Nano Omni 30B A3BNVIDIA · Open weight

72.2%

DeepSeek V4 FlashDeepSeek · Open weight

71.2%

GPT-5 nanoOpenAI · Closed

71.2%

ZAYA1-8BZyphra · Open weight

71%

GPT-4.1OpenAI · Closed

66.3%

GPT-4.1 miniOpenAI · Closed

64.2%

Claude 3.5 SonnetAnthropic · Closed

59.4%

DeepSeek V3DeepSeek · Open weight

59.1%

Ling 2.6 FlashInclusionAI · Open weight

59%

Gemma 4 E4BGoogle · Open weight

58.6%

Mellum2-12B-A2.5B-ThinkingJetBrains · Open weight

57.6%

ZAYA1-74B-PreviewZyphra · Open weight

57.3%

GPT-4.1 nanoOpenAI · Closed

50.3%

Soofi S 30B-A3BSoofi Project · Open weight

43.4%

Gemma 4 E2BGoogle · Open weight

43.4%

Mellum2-12B-A2.5B-InstructJetBrains · Open weight

40.9%

LFM2.5-VL-450MLiquidAI · Open weight

25.7%

LFM2.5-230MLiquidAI · Open weight

25.4%

According to BenchLM.ai, Sakana Fugu-Ultra leads the GPQA benchmark with a score of 95.5%, followed by Sakana Fugu (95.5%) and GPT-5.6 Sol (94.6%). The top models are clustered within 0.9 points, suggesting this benchmark is nearing saturation for frontier models.

71 models have been evaluated on GPQA. The benchmark falls in the Knowledge category. This category carries a 12% weight in BenchLM.ai's overall scoring system. Within that category, GPQA contributes 7% of the category score, so strong performance here directly affects a model's overall ranking.

About GPQA

Year

2023

Tasks

448 questions

Format

Multiple choice questions

Difficulty

Graduate level

GPQA questions are crafted by PhD-level domain experts and validated to be answerable by experts but challenging for non-experts even with internet access. This makes it an excellent test of deep scientific knowledge and reasoning.

GPQA: A Graduate-Level Google-Proof Q&A Benchmark

BenchLM freshness & provenance

Version

GPQA Diamond

Refresh cadence

Static

Staleness state

Refreshing

Question availability

Public benchmark set

Refreshing

BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.

FAQ

What does GPQA measure?

A challenging dataset of 448 multiple-choice questions written by domain experts in biology, physics, and chemistry. Designed to be difficult even for skilled non-experts with access to Google.

Which model scores highest on GPQA?

Sakana Fugu-Ultra by Sakana AI currently leads with a score of 95.5% on GPQA.

How many models are evaluated on GPQA?

71 AI models have been evaluated on GPQA on BenchLM.

Compare Top Models on GPQA

Sakana Fugu-Ultra vs Sakana Fugu Sakana Fugu vs GPT-5.6 Sol GPT-5.6 Sol vs Claude Opus 4.7 (Adaptive)Claude Opus 4.7 (Adaptive) vs Claude Mythos 5

Learn More

Read our explainer: GPQA benchmark deep dive

Last updated: July 20, 2026 · BenchLM version GPQA Diamond

Choose a model with this week’s evidence

Join 2,000+ readers for ranking moves, pricing changes, and the claims that still need proof.

One email each week. Unsubscribe anytime.