Knowledge · Benchmark profile
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 27 Jul 2026 · Methodology 1.8.0
Benchmark score on Graduate-Level Google-Proof Q&A
Sakana Fugu leads at 95.5%, followed by Sakana Fugu-Ultra (95.5%) and GPT-5.6 Sol (94.6%).
10 ranked models · higher is better · labels show rank and score
| Model | Rank | Provider | Score |
|---|---|---|---|
| Sakana Fugu | #1 | Sakana AI | 95.5% |
| Sakana Fugu-Ultra | #2 | Sakana AI | 95.5% |
| GPT-5.6 Sol | #3 | OpenAI | 94.6% |
| Claude Opus 4.7 (Adaptive) | #4 | Anthropic | 94.2% |
| Claude Mythos 5 | #5 | Anthropic | 94.1% |
| Claude Opus 4.8 | #6 | Anthropic | 93.6% |
| GPT-5.5 | #7 | OpenAI | 93.6% |
| Kimi K3 | #8 | Moonshot AI | 93.5% |
| GPT-5.6 Terra | #9 | OpenAI | 92.9% |
| GPT-5.4 | #10 | OpenAI | 92.8% |
One best score per model · higher is better
| Rank | Model | Provider | License | Evidence use | Score |
|---|---|---|---|---|---|
| #1 | Sakana Fugu sakana-fugu | Sakana AI | closed | Estimated reference | 95.5% |
| #2 | Sakana Fugu-Ultra sakana-fugu-ultra | Sakana AI | closed | Estimated reference | 95.5% |
| #3 | GPT-5.6 Sol gpt-5.6-sol | OpenAI | closed | Reference only | 94.6% |
| #4 | Claude Opus 4.7 (Adaptive) claude-opus-4-7-max | Anthropic | closed | Estimated reference | 94.2% |
| #5 | Claude Mythos 5 claude-mythos-5 | Anthropic | closed | Reference only | 94.1% |
| #6 | Claude Opus 4.8 claude-opus-4-8 | Anthropic | closed | Reference only | 93.6% |
| #7 | GPT-5.5 gpt-5.5 | OpenAI | closed | Estimated reference | 93.6% |
| #8 | Kimi K3 kimi-k3 | Moonshot AI | closed | Reference only | 93.5% |
| #9 | GPT-5.6 Terra gpt-5.6-terra | OpenAI | closed | Estimated reference | 92.9% |
| #10 | GPT-5.4 gpt-5.4 | OpenAI | closed | Reference only | 92.8% |
| #11 | GPT-5.2 gpt-5.2 | OpenAI | closed | Estimated reference | 92.4% |
| #12 | Qwen3.7-Max qwen3.7-max | Alibaba Cloud | closed | Reference only | 92.4% |
| #13 | GPT-5.6 Luna gpt-5.6-luna | OpenAI | closed | Estimated reference | 92.3% |
| #14 | Gemini 3.5 Flash gemini-3.5-flash | closed | Estimated reference | 92.2% | |
| #15 | Claude Opus 4.6 claude-opus-4-6 | Anthropic | closed | Reference only | 91.3% |
| #16 | GLM-5.2 glm-5-2 | Z.AI | open | Estimated reference | 91.2% |
| #17 | Kimi K2.6 kimi-k2.6 | Moonshot AI | open | Estimated reference | 90.5% |
| #18 | Qwen3.6 Plus qwen3-6-plus | Alibaba Cloud | closed | Reference only | 90.4% |
| #19 | Qwen3.7-Plus qwen3.7-plus | Alibaba Cloud | closed | Reference only | 90.3% |
| #20 | DeepSeek V4 Pro (Max) deepseek-v4-pro-max | DeepSeek | open | Estimated reference | 90.1% |
| #21 | Grok 4.3 grok-4.3 | xAI | closed | Reference only | 90.1% |
| #22 | Claude Sonnet 4.6 claude-sonnet-4-6 | Anthropic | closed | Reference only | 89.9% |
| #23 | Interfaze Beta interfaze-beta | Interfaze | closed | Estimated reference | 89.9% |
| #24 | DeepSeek V4 Pro (High) deepseek-v4-pro-high | DeepSeek | open | Estimated reference | 89.1% |
| #25 | Qwen3.5 397B qwen3.5-397b-a17b | Alibaba Cloud | open | Estimated reference | 88.4% |
| #26 | DeepSeek V4 Flash (Max) deepseek-v4-flash-max | DeepSeek | open | Estimated reference | 88.1% |
| #27 | GPT-5.4 mini gpt-5.4-mini | OpenAI | closed | Estimated reference | 88% |
| #28 | Inkling inkling | Thinking Machines Lab | open | Reference only | 87.9% |
| #29 | Qwen3.6 27B qwen3.6-27b | Alibaba Cloud | open | Estimated reference | 87.8% |
| #30 | Kimi K2.5 kimi-k2.5 | Moonshot AI | open | Reference only | 87.6% |
| #31 | Kimi K2.5 (Reasoning) kimi-k2-5-reasoning | Moonshot AI | closed | Estimated reference | 87.6% |
| #32 | DeepSeek V4 Flash (High) deepseek-v4-flash-high | DeepSeek | open | Estimated reference | 87.4% |
| #33 | Hy3 Preview hy3-preview | Tencent | open | Estimated reference | 87.2% |
| #34 | Claude Opus 4.5 claude-opus-4-5 | Anthropic | closed | Reference only | 87% |
| #35 | Nemotron 3 Ultra nemotron-3-ultra-550b | NVIDIA | open | Estimated reference | 87% |
| #36 | Qwen3.5-122B-A10B qwen3-5-122b-a10b | Alibaba Cloud | open | Reference only | 86.6% |
| #37 | GLM-5 glm-5 | Z.AI | open | Reference only | 86% |
| #38 | Qwen3.6-35B-A3B qwen3-6-35b-a3b | Alibaba Cloud | open | Estimated reference | 86% |
| #39 | GLM-4.7 glm-4.7 | Z.AI | open | Reference only | 85.7% |
| #40 | Qwen3.5 27B qwen3.5-27b | Alibaba Cloud | open | Reference only | 85.5% |
| #41 | Gemma 4 31B gemma-4-31b | open | Reference only | 84.3% | |
| #42 | MAI-Thinking-1 mai-thinking-1 | Microsoft | closed | Estimated reference | 84.2% |
| #43 | Qwen3.5-35B-A3B qwen3-5-35b-a3b | Alibaba Cloud | open | Reference only | 84.2% |
| #44 | MiMo-V2-Flash mimo-v2-flash | Xiaomi | open | Reference only | 83.7% |
| #45 | Claude Sonnet 4.5 claude-sonnet-4-5 | Anthropic | closed | Estimated reference | 83.4% |
| #46 | Gemini 2.5 Pro gemini-2.5-pro | closed | Reference only | 83% | |
| #47 | GPT-5.4 nano gpt.5.4.nano | OpenAI | closed | Reference only | 82.8% |
| #48 | o1-pro o1-pro | OpenAI | closed | Estimated reference | 79% |
| #49 | Gemma 4 12B gemma-4-12b | open | Estimated reference | 78.8% | |
| #50 | Qwen3 235B 2507 qwen3-235b-2507 | Alibaba Cloud | open | Estimated reference | 77.5% |
| #51 | o3-mini o3-mini | OpenAI | closed | Reference only | 77.2% |
| #52 | o1 o1 | OpenAI | closed | Estimated reference | 75.7% |
| #53 | DeepSeek V4 Pro deepseek-v4-pro | DeepSeek | open | Reference only | 72.9% |
| #54 | Nemotron 3 Nano Omni 30B A3B nemotron-3-nano-omni-30b-a3b | NVIDIA | open | Estimated reference | 72.2% |
| #55 | DeepSeek V4 Flash deepseek-v4-flash | DeepSeek | open | Estimated reference | 71.2% |
| #56 | ZAYA1-8B zaya1-8b | Zyphra | open | Estimated reference | 71% |
| #57 | GPT-4.1 gpt-4-1 | OpenAI | closed | Reference only | 66.3% |
| #58 | GPT-4.1 mini gpt-4-1-mini | OpenAI | closed | Estimated reference | 64.2% |
| #59 | Claude 3.5 Sonnet claude-3-5-sonnet | Anthropic | closed | Estimated reference | 59.4% |
| #60 | DeepSeek V3 deepseek-v3 | DeepSeek | open | Reference only | 59.1% |
| #61 | Ling 2.6 Flash ling-2-6-flash | InclusionAI | open | Estimated reference | 59% |
| #62 | Gemma 4 E4B gemma-4-e4b | open | Estimated reference | 58.6% | |
| #63 | Mellum2-12B-A2.5B-Thinking mellum2-12b-a2-5b-thinking | JetBrains | open | Estimated reference | 57.6% |
| #64 | ZAYA1-74B-Preview zaya1-74b-preview | Zyphra | open | Estimated reference | 57.3% |
| #65 | GPT-4.1 nano gpt-4-1-nano | OpenAI | closed | Estimated reference | 50.3% |
| #66 | Gemma 4 E2B gemma-4-e2b | open | Estimated reference | 43.4% | |
| #67 | Soofi S 30B-A3B soofi-s-30b-a3b | Soofi Project | open | Estimated reference | 43.4% |
| #68 | Mellum2-12B-A2.5B-Instruct mellum2-12b-a2-5b-instruct | JetBrains | open | Estimated reference | 40.9% |
| #69 | LFM2.5-VL-450M lfm2-5-vl-450m | LiquidAI | open | Estimated reference | 25.7% |
| #70 | LFM2.5-230M lfm2-5-230m | LiquidAI | open | Estimated reference | 25.4% |
The top of this snapshot is led by Sakana Fugu at 95.5%; third place is 0.90 points behind. The top-10 spread is 2.7 points.
About Graduate-Level Google-Proof Q&A
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. Results stay tied to the exact model variant and evaluation system. Multiple systems for the same model use the best published score on this page; overall Lumina scoring uses the median of ranking-eligible rows.
Open benchmark source ↗FAQ
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.
Sakana Fugu by Sakana AI currently leads with 95.5%.
70 models in the LuminaBench cohort have a qualifying score on this benchmark.
Related
2026 · 21 results · Reference
Artificial Analysis MMLU-Pro2026 · 8 results · Reference
Massive Multitask Language Understanding2020 · 16 results · Reference
SuperGPQA: Scaling LLM Evaluation Across 285 Graduate Disciplines2025 · 38 results · Reference
Massive Multitask Language Understanding Professional2024 · 85 results · Reference
AGIEval2026 · 6 results · Reference
