Graduate-Level Google-Proof Q&A (GPQA)
GPT-6 Astra leads the GPQA leaderboard on BenchLM's October 2026 update with 96%, ahead of Sakana Fugu-Ultra (95.5%) and Sakana Fugu (95.5%), across 143 models.
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.
Top models on GPQA — October 17, 2026
As of October 17, 2026, GPT-6 Astra leads the GPQA leaderboard with 96% , followed by Sakana Fugu-Ultra (95.5%) and Sakana Fugu (95.5%).
GPT-6 Astra
OpenAI
Sakana Fugu-Ultra
Sakana AI
Sakana Fugu
Sakana AI
143 modelsKnowledgeGPQA follows the active BenchAlign v5.8 policy.MethodologyRefreshingUpdated October 17, 2026
Leaderboard (143 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | GPT-6 AstraOpenAI | 96% | Not reported | Closed |
| 2 | Sakana Fugu-UltraSakana AI | 95.5% | Not reported | Closed |
| 3 | Sakana FuguSakana AI | 95.5% | Not reported | Closed |
| 4 | GPT-5.6 SolOpenAI | 94.6% | Not reported | Closed |
| 5 | Claude Opus 4.7 (Adaptive)Anthropic | 94.2% | Not reported | Closed |
| 6 | Claude Mythos 5Anthropic | 94.1% | Not reported | Closed |
| 7 | Claude Opus 4.8Anthropic | 93.6% | Not reported | Closed |
| 8 | GPT-5.5OpenAI | 93.6% | Not reported | Closed |
| 9 | Kimi K3Moonshot AI | 93.5% | Not reported | Pending |
| 10 | GPT-5.6 TerraOpenAI | 92.9% | Not reported | Closed |
| 11 | GPT-5.4OpenAI | 92.8% | Not reported | Closed |
| 12 | Ornith-1.5-397BOrnith AI | 92.8% | Not reported | Open |
| 13 | Qwen3.8 MaxAlibaba | 92.6% | Not reported | Open |
| 14 | Qwen3.7 MaxAlibaba | 92.4% | Not reported | Closed |
| 15 | GPT-5.2OpenAI | 92.4% | Not reported | Closed |
| 16 | GPT-5.6 LunaOpenAI | 92.3% | Not reported | Closed |
| 17 | Hy4 previewTencent | 92.3% | Not reported | Open |
| 18 | Gemini 3.5 FlashGoogle | 92.2% | Not reported | Closed |
| 19 | Qwen3.8-Flash-NextAlibaba | 91.7% | Not reported | Open |
| 20 | Claude Opus 4.6Anthropic | 91.3% | Not reported | Closed |
| 21 | GLM-5.2Z.AI | 91.2% | Not reported | Open |
| 22 | Qwen3.8-Omni-FlashAlibaba | 91% | Not reported | Closed |
| 23 | DeepSeek V4.1 Flash 1017DeepSeek | 90.9% | Not reported | Open |
| 24 | Kimi K2.6Moonshot AI | 90.5% | Not reported | Open |
| 25 | Qwen3.6 PlusAlibaba | 90.4% | Not reported | Closed |
| 26 | Qwen3.7 PlusAlibaba | 90.3% | Not reported | Closed |
| 27 | DeepSeek V4 Pro 0813DeepSeek | 90.1% | Not reported | Open |
| 28 | Grok 4.3xAI | 90.1% | Not reported | Closed |
| 29 | Claude Sonnet 4.6Anthropic | 89.9% | Not reported | Closed |
| 30 | Interfaze BetaInterfaze | 89.9% | Not reported | Closed |
| 31 | Inkling-SmallThinking Machines Lab | 89.5% | Not reported | Open |
| 32 | Qwen3.8-27BAlibaba | 89.2% | Not reported | Open |
| 33 | Ornith-1.5-35B-A3BOrnith AI | 89.2% | Not reported | Open |
| 34 | DeepSeek V4 Pro (High)DeepSeek | 89.1% | Not reported | Open |
| 35 | Qwen3.5 397BAlibaba | 88.4% | Not reported | Open |
| 36 | DeepSeek V4 Flash 0731DeepSeek | 88.1% | Not reported | Open |
| 37 | GPT-5.4 miniOpenAI | 88% | Not reported | Closed |
| 38 | InklingThinking Machines Lab | 87.9% | Not reported | Open |
| 39 | Qwen3.6-27BAlibaba | 87.8% | Not reported | Open |
| 40 | Kimi K2.5 (Reasoning)Moonshot AI | 87.6% | Not reported | Closed |
| 41 | Kimi K2.5Moonshot AI | 87.6% | Not reported | Open |
| 42 | DeepSeek V4 Flash (High)DeepSeek | 87.4% | Not reported | Open |
| 43 | Hy3 PreviewTencent | 87.2% | Not reported | Open |
| 44 | Claude Opus 4.5Anthropic | 87% | Not reported | Closed |
| 45 | Nemotron 3 UltraNVIDIA | 87% | Not reported | Open |
| 46 | Qwen3.5-122B-A10BAlibaba | 86.6% | Not reported | Open |
| 47 | Ornith-1.5-9BOrnith AI | 86.4% | Not reported | Open |
| 48 | GLM-5Z.AI | 86% | Not reported | Open |
| 49 | Qwen3.6-35B-A3BAlibaba | 86% | Not reported | Open |
| 50 | Ternary Bonsai 2 27BPrism ML | 85.8% | Not reported | Open |
| 51 | GLM-4.7Z.AI | 85.7% | Not reported | Open |
| 52 | Qwen3.5-27BAlibaba | 85.5% | Not reported | Open |
| 53 | Ling 3.0 FlashInclusionAI | 85.0% | Not reported | Open |
| 54 | Gemma 4 31BGoogle | 84.3% | Not reported | Open |
| 55 | Qwen3.5-35B-A3BAlibaba | 84.2% | Not reported | Open |
| 56 | MAI-Thinking-1Microsoft | 84.2% | Not reported | Closed |
| 57 | Ling 3.0 Flash FP8InclusionAI | 84% | Not reported | Open |
| 58 | MiMo-V2-FlashXiaomi | 83.7% | Not reported | Open |
| 59 | Claude Sonnet 4.5Anthropic | 83.4% | Not reported | Closed |
| 60 | Gemini 2.5 ProGoogle | 83% | Not reported | Closed |
| 61 | GPT-5.4 nanoOpenAI | 82.8% | Not reported | Closed |
| 62 | Gemma 4 26B A4BGoogle | 82.3% | Not reported | Open |
| 63 | Qwen3 235B 2507 (Reasoning)Alibaba | 81.1% | Not reported | Open |
| 64 | DeepSeek-R1-0528DeepSeek | 81% | Not reported | Open |
| 65 | Grok 3 MinixAI | 80.3% | Not reported | Closed |
| 66 | DeepSeek V3.1 (Reasoning)DeepSeek | 80.1% | Not reported | Open |
| 67 | GPT-OSS 120BOpenAI | 80.1% | Not reported | Open |
| 68 | Claude 4 Opus ThinkingAnthropic | 79.6% | Not reported | Closed |
| 69 | Nemotron 3 Super 100BNVIDIA | 79.2% | Not reported | Open |
| 70 | o1-proOpenAI | 79% | Not reported | Closed |
| 71 | Gemma 4 12BGoogle | 78.8% | Not reported | Open |
| 72 | Claude 3.7 Sonnet ThinkingAnthropic | 78.2% | Not reported | Closed |
| 73 | o4-miniOpenAI | 77.6% | Not reported | Closed |
| 74 | Qwen3 235B 2507Alibaba | 77.5% | Not reported | Open |
| 75 | o3-miniOpenAI | 77.2% | Not reported | Closed |
| 76 | Nemotron Ultra 253BNVIDIA | 76.0% | Not reported | Open |
| 77 | o1OpenAI | 75.7% | Not reported | Closed |
| 78 | Nemotron 3.5 Lightning 30B A3B NVFP4NVIDIA | 75.6% | Not reported | Open |
| 79 | Grok 3 [Beta]xAI | 75.4% | Not reported | Closed |
| 80 | GLM-4.7-FlashZ.AI | 75.2% | Not reported | Open |
| 81 | Kimi K2Moonshot AI | 75.1% | Not reported | Closed |
| 82 | Claude 4 OpusAnthropic | 74.9% | Not reported | Closed |
| 83 | DeepSeek V3.1DeepSeek | 74.9% | Not reported | Open |
| 84 | LFM2.5-24B-A2BLiquidAI | 74.4% | Not reported | Open |
| 85 | Mercury 2Inception | 73.6% | Not reported | Closed |
| 86 | Nemotron 3 Nano 30BNVIDIA | 73% | Not reported | Open |
| 87 | DeepSeek V4 ProDeepSeek | 72.9% | Not reported | Open |
| 88 | Nemotron 3 Nano Omni 30B A3BNVIDIA | 72.2% | Not reported | Open |
| 89 | DeepSeek-R1DeepSeek | 71.5% | Not reported | Open |
| 90 | GPT-OSS 20BOpenAI | 71.5% | Not reported | Open |
| 91 | Ministral 3 14B (Reasoning)Mistral | 71.2% | Not reported | Open |
| 92 | DeepSeek V4 FlashDeepSeek | 71.2% | Not reported | Open |
| 93 | GPT-5 nanoOpenAI | 71.2% | Not reported | Closed |
| 94 | Qwen3 235B-A22B (Reasoning)Alibaba | 71.1% | Not reported | Open |
| 95 | ZAYA1-8BZyphra | 71% | Not reported | Open |
| 96 | Magistral MediumMistral | 70.8% | Not reported | Closed |
| 97 | GPT-4.5OpenAI | 69.5% | Not reported | Closed |
| 98 | Phi-4-reasoning-plusMicrosoft | 68.9% | Not reported | Open |
| 99 | Qwen3 32B (Reasoning)Alibaba | 68.4% | Not reported | Open |
| 100 | Magistral SmallMistral | 68.2% | Not reported | Open |
| 101 | Claude 3.7 SonnetAnthropic | 68% | Not reported | Closed |
| 102 | Granite 4.2 30BIBM | 66.4% | Not reported | Open |
| 103 | GPT-4.1OpenAI | 66.3% | Not reported | Closed |
| 104 | Phi-4-reasoningMicrosoft | 65.8% | Not reported | Open |
| 105 | QwQ-32BAlibaba | 65.6% | Not reported | Open |
| 106 | Gemini 2.0 FlashGoogle | 65.2% | Not reported | Closed |
| 107 | Gemini 2.5 Flash-LiteGoogle | 64.6% | Not reported | Closed |
| 108 | GPT-4.1 miniOpenAI | 64.2% | Not reported | Closed |
| 109 | Granite 4.2 8BIBM | 64.1% | Not reported | Open |
| 110 | Qwen3 235B-A22BAlibaba | 62.9% | Not reported | Open |
| 111 | Qwen2.5-MaxAlibaba | 60.1% | Not reported | Closed |
| 112 | o1-miniOpenAI | 60% | Not reported | Closed |
| 113 | Claude 3.5 SonnetAnthropic | 59.4% | Not reported | Closed |
| 114 | DeepSeek V3DeepSeek | 59.1% | Not reported | Open |
| 115 | Ling 2.6 FlashInclusionAI | 59% | Not reported | Open |
| 116 | Gemma 4 E4BGoogle | 58.6% | Not reported | Open |
| 117 | Mellum2-12B-A2.5B-ThinkingJetBrains | 57.6% | Not reported | Open |
| 118 | ZAYA1-74B-PreviewZyphra | 57.3% | Not reported | Open |
| 119 | Phi-4Microsoft | 56.1% | Not reported | Open |
| 120 | Grok 2xAI | 56% | Not reported | Closed |
| 121 | Granite 4.2 3BIBM | 54.8% | Not reported | Open |
| 122 | Qwen3 32BAlibaba | 54.6% | Not reported | Open |
| 123 | Grok 2 minixAI | 51% | Not reported | Closed |
| 124 | Llama 3.3 70BMeta | 50.5% | Not reported | Open |
| 125 | Gemini 2.0 Flash-LiteGoogle | 50.5% | Not reported | Closed |
| 126 | GPT-4.1 nanoOpenAI | 50.3% | Not reported | Closed |
| 127 | Gemini 1.5 FlashGoogle | 50% | Not reported | Closed |
| 128 | Llama 3.1 70BMeta | 46.7% | Not reported | Open |
| 129 | Qwen2.5-VL-32BAlibaba | 46% | Not reported | Open |
| 130 | Mistral Small 3.1Mistral | 46.0% | Not reported | Open |
| 131 | Gemma 4 E2BGoogle | 43.4% | Not reported | Open |
| 132 | Soofi S 30B-A3BSoofi Project | 43.4% | Not reported | Open |
| 133 | Claude 3.5 HaikuAnthropic | 41.6% | Not reported | Closed |
| 134 | Mellum2-12B-A2.5B-InstructJetBrains | 40.9% | Not reported | Open |
| 135 | LFM2.5-350MLiquidAI | 30.6% | Not reported | Open |
| 136 | Llama 3.1 8BMeta | 30.4% | Not reported | Open |
| 137 | Phi-3.5-miniMicrosoft | 30.4% | Not reported | Open |
| 138 | Granite-4.0-1BIBM | 29.9% | Not reported | Open |
| 139 | Granite-4.0-H-1BIBM | 29.7% | Not reported | Open |
| 140 | Granite-4.0-H-350MIBM | 26.1% | Not reported | Open |
| 141 | LFM2.5-VL-450MLiquidAI | 25.7% | Not reported | Open |
| 142 | LFM2.5-230MLiquidAI | 25.4% | Not reported | Open |
| 143 | Granite-4.0-350MIBM | 24.1% | Not reported | Open |
How to read this leaderboard
GPQA is a four-choice biology, physics and chemistry benchmark written and validated by domain experts. The published family contains a 546-question Extended set, a 448-question Main set and the stricter 198-question Diamond subset.
Variant check: At the September 22, 2026 editorial check, the GPQA source ledger was mixed: most underlying reports said GPQA Diamond, while some identified only GPQA. The table does not expose a complete per-row subset or protocol label.
The table does not expose complete per-row variant or evaluation-setup labels. Consult the original provider or benchmark report before comparing results, and compare only runs that establish the same subset, prompting, tool access and scoring protocol.
Operator receipt: 143 sourced rows are currently displayable on this page; the leading published row is GPT-6 Astra at 96%.
Provenance: At the September 22, 2026 editorial check, the table was assembled from provider and secondary reports. It was not an official benchmark-owner leaderboard, and the source ledger contained no benchmark-owner or independent result rows for this lane.
Source freshness: The primary paper was submitted on November 20, 2023 and published at COLM 2024. The official repository and gated dataset define the benchmark; this page's refresh date records changes to the source ledger, not a new GPQA release.
Honest limit: The mixed source ledger prevents a controlled whole-table comparison, and the gated questions should not be reproduced. Treat the sorted values as directional until the original provider or benchmark report confirms the subset and setup for the particular results being compared.
According to BenchLM.ai, GPT-6 Astra leads the GPQA benchmark with a score of 96%, followed by Sakana Fugu-Ultra (95.5%) and Sakana Fugu (95.5%). The top models are clustered within 0.5 points, suggesting this benchmark is nearing saturation for frontier models.
143 models have been evaluated on GPQA. The benchmark falls in the Knowledge category. GPQA follows the active BenchAlign v5.8 policy. See the methodology for the current scoring and qualification rules.
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.
Freshness and provenance
Version
GPQA Diamond
Refresh cadence
Static
Staleness state
Refreshing
Question availability
Public benchmark set
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.
Questions
What does GPQA measure?
GPQA is a four-choice science benchmark written and validated by domain experts. The published family includes a 546-question Extended set, a 448-question Main set, and the stricter 198-question Diamond subset. A result is interpretable only when its source identifies which subset and evaluation setup it used.
Are all rows on this page GPQA Diamond results?
No. At the September 22, 2026 editorial check, underlying reports were mostly labelled GPQA Diamond, while some identified only GPQA. The table does not expose a variant label for every row. Consult the original provider or benchmark report, and compare results only when they establish the same subset and setup.
Where should I verify a GPQA result?
Use the paper, official repository, and gated dataset for the benchmark definition. At the September 22, 2026 editorial check, this source ledger was assembled from provider and secondary reports, not an official owner-run leaderboard. The dataset authors use controlled access and contamination markers, so this page does not reproduce benchmark questions.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 5,600+ readers.
One email each week. Unsubscribe anytime.