Calibration
AA-GPQA Diamond → LABBench2
LABBench2 is estimated from AA-GPQA Diamond with a inverse Michaelis–Menten curve fitted on 6 models measured on both: y = 0.94179·(x − 0.0000) / (0.0000 + 2.0000 − x), R² = 0.44, cross-validated error 1.8 pp. It is used for 127 estimates.
| Estimated model | AA-GPQA Diamond | LABBench2 | Source |
|---|---|---|---|
| A.X K2 | 85.7% | 70.6% | estimated ± 1.8 pp, low confidence |
| Apodex 1.1 | 86.4% | 71.6% | estimated ± 1.8 pp, low confidence |
| Apodex 1.1 Mini | 86.4% | 71.6% | estimated ± 1.8 pp, low confidence |
| Celeris-1 | 63.1% | 43.4% | estimated ± 1.8 pp, low confidence |
| Claude 3 Haiku | 37.4% | 21.7% | estimated ± 1.8 pp, low confidence |
| Claude 3 Opus | 48.9% | 30.5% | estimated ± 1.8 pp, low confidence |
| Claude 4.1 Opus Thinking | 80.9% | 64.0% | estimated ± 1.8 pp, low confidence |
| Claude 4 Sonnet | 68.3% | 48.8% | estimated ± 1.8 pp, low confidence |
| Claude Opus 4.5 | 81.0% | 64.1% | estimated ± 1.8 pp, low confidence |
| Claude Opus 4.6 | 84.0% | 68.2% | estimated ± 1.8 pp, low confidence |
| Claude Opus 4.7 (Adaptive) | 91.4% | 79.3% | estimated ± 1.8 pp, low confidence |
| Command A+ | 76.1% | 57.8% | estimated ± 1.8 pp, low confidence |
| DeepSeek-R1 | 81.3% | 64.5% | estimated ± 1.8 pp, low confidence |
| DeepSeek R1 Distill Qwen 32B | 61.5% | 41.8% | estimated ± 1.8 pp, low confidence |
| DeepSeek V3 | 55.7% | 36.4% | estimated ± 1.8 pp, low confidence |
| DeepSeek V3 0324 | 65.5% | 45.9% | estimated ± 1.8 pp, low confidence |
| DeepSeek V3.1 | 73.5% | 54.7% | estimated ± 1.8 pp, low confidence |
| DeepSeek V3.1 (Reasoning) | 77.9% | 60.1% | estimated ± 1.8 pp, low confidence |
| DeepSeek V3.2 | 75.1% | 56.6% | estimated ± 1.8 pp, low confidence |
| Exaone 4.0 1.2B | 42.4% | 25.3% | estimated ± 1.8 pp, low confidence |
| Exaone 4.0 32B | 62.8% | 43.1% | estimated ± 1.8 pp, low confidence |
| Gemini 1.0 Pro | 27.7% | 15.1% | estimated ± 1.8 pp, low confidence |
| Gemini 1.5 Pro | 58.9% | 39.3% | estimated ± 1.8 pp, low confidence |
| Gemini 2.5 Flash | 68.3% | 48.8% | estimated ± 1.8 pp, low confidence |
| Gemini 2.5 Pro | 84.4% | 68.8% | estimated ± 1.8 pp, low confidence |
| Gemma 3 27B | 42.8% | 25.6% | estimated ± 1.8 pp, low confidence |
| Gemma 4 12B | 75.3% | 56.9% | estimated ± 1.8 pp, low confidence |
| Gemma 4 26B A4B | 79.2% | 61.7% | estimated ± 1.8 pp, low confidence |
| Gemma 4 31B | 85.7% | 70.6% | estimated ± 1.8 pp, low confidence |
| Gemma 4 E2B | 43.3% | 26.0% | estimated ± 1.8 pp, low confidence |
| Gemma 4 E4B | 57.6% | 38.1% | estimated ± 1.8 pp, low confidence |
| GLM-4.5-Air | 73.3% | 54.5% | estimated ± 1.8 pp, low confidence |
| GLM-5 | 82.0% | 65.4% | estimated ± 1.8 pp, low confidence |
| GLM-5-Turbo | 84.7% | 69.2% | estimated ± 1.8 pp, low confidence |
| GLM-5V-Turbo | 80.9% | 64.0% | estimated ± 1.8 pp, low confidence |
| GPT-4.1 | 66.6% | 47.0% | estimated ± 1.8 pp, low confidence |
| GPT-4.1 mini | 66.4% | 46.8% | estimated ± 1.8 pp, low confidence |
| GPT-4.1 nano | 51.2% | 32.4% | estimated ± 1.8 pp, low confidence |
| GPT-4o | 54.3% | 35.1% | estimated ± 1.8 pp, low confidence |
| GPT-4o mini | 42.6% | 25.5% | estimated ± 1.8 pp, low confidence |
| GPT-5.1-Codex | 86.0% | 71.0% | estimated ± 1.8 pp, low confidence |
| GPT-5.1-Codex-Max | 86.0% | 71.0% | estimated ± 1.8 pp, low confidence |
| GPT-5.2-Codex | 89.9% | 76.9% | estimated ± 1.8 pp, low confidence |
| GPT-5.3 Codex | 91.5% | 79.4% | estimated ± 1.8 pp, low confidence |
| GPT-5 (high) | 85.4% | 70.2% | estimated ± 1.8 pp, low confidence |
| GPT-5 (medium) | 84.2% | 68.5% | estimated ± 1.8 pp, low confidence |
| GPT-OSS 120B | 78.2% | 60.5% | estimated ± 1.8 pp, low confidence |
| GPT-OSS 20B | 68.8% | 49.4% | estimated ± 1.8 pp, low confidence |
| Granite-4.0-350M | 26.1% | 14.1% | estimated ± 1.8 pp, low confidence |
| Granite-4.0-H-1B | 26.3% | 14.3% | estimated ± 1.8 pp, low confidence |
| Granite-4.0-H-350M | 25.7% | 13.9% | estimated ± 1.8 pp, low confidence |
| Granite 4.2 30B | 64.4% | 44.7% | estimated ± 1.8 pp, low confidence |
| Granite 4.2 3B | 55.9% | 36.5% | estimated ± 1.8 pp, low confidence |
| Granite 4.2 8B | 63.1% | 43.4% | estimated ± 1.8 pp, low confidence |
| Grok 4 | 87.7% | 73.5% | estimated ± 1.8 pp, low confidence |
| Grok 4.1 Fast | 63.7% | 44.0% | estimated ± 1.8 pp, low confidence |
| Grok 4.1 Fast (Reasoning) | 85.3% | 70.0% | estimated ± 1.8 pp, low confidence |
| Grok 4 Fast (Reasoning) | 84.7% | 69.2% | estimated ± 1.8 pp, low confidence |
| Grok Code Fast 1 | 72.7% | 53.8% | estimated ± 1.8 pp, low confidence |
| Hy3 | 89.7% | 76.6% | estimated ± 1.8 pp, low confidence |
| Hy3 Preview | 89.7% | 76.6% | estimated ± 1.8 pp, low confidence |
| K-Exaone | 78.3% | 60.6% | estimated ± 1.8 pp, low confidence |
| K-EXAONE 2.0 | 82.9% | 66.7% | estimated ± 1.8 pp, low confidence |
| Kimi K2 | 76.6% | 58.5% | estimated ± 1.8 pp, low confidence |
| Kimi K2.5 | 87.9% | 73.8% | estimated ± 1.8 pp, low confidence |
| Kimi K2.5 (Reasoning) | 87.9% | 73.8% | estimated ± 1.8 pp, low confidence |
| Kimi K2.7 Code | 89.6% | 76.4% | estimated ± 1.8 pp, low confidence |
| LFM2.5-2.6B | 55.8% | 36.4% | estimated ± 1.8 pp, low confidence |
| LFM2.5-8B-A1B | 51.3% | 32.5% | estimated ± 1.8 pp, low confidence |
| LFM2.5-VL-1.6B-Extract | 28.9% | 15.9% | estimated ± 1.8 pp, low confidence |
| Ling 2.6 Flash | 59.3% | 39.7% | estimated ± 1.8 pp, low confidence |
| Ling 3.0 Flash FP8 | 85.5% | 70.3% | estimated ± 1.8 pp, low confidence |
| Ling 3.0 Flash VL | 86.2% | 71.3% | estimated ± 1.8 pp, low confidence |
| Ling 3.0 Tiny | 73.4% | 54.6% | estimated ± 1.8 pp, low confidence |
| Llama 3.1 405B | 51.5% | 32.7% | estimated ± 1.8 pp, low confidence |
| Llama 4 Maverick | 67.1% | 47.5% | estimated ± 1.8 pp, low confidence |
| Llama 4 Scout | 58.7% | 39.1% | estimated ± 1.8 pp, low confidence |
| MiMo-V2-Flash | 65.6% | 46.0% | estimated ± 1.8 pp, low confidence |
| MiMo-V2-Omni | 82.8% | 66.5% | estimated ± 1.8 pp, low confidence |
| MiMo-V2-Pro | 87.0% | 72.5% | estimated ± 1.8 pp, low confidence |
| MiniCPM5-2B | 70.2% | 50.9% | estimated ± 1.8 pp, low confidence |
| Mistral Large 2 | 48.6% | 30.2% | estimated ± 1.8 pp, low confidence |
| Mistral Large 3 | 68.0% | 48.5% | estimated ± 1.8 pp, low confidence |
| Mistral Medium 3 | 57.8% | 38.3% | estimated ± 1.8 pp, low confidence |
| Mistral Small 4 | 76.9% | 58.8% | estimated ± 1.8 pp, low confidence |
| Mistral Small 4 (Reasoning) | 76.9% | 58.8% | estimated ± 1.8 pp, low confidence |
| Muse Glimmer 30B | 83.5% | 67.5% | estimated ± 1.8 pp, low confidence |
| Muse Spark 1.3 | 93.5% | 82.7% | estimated ± 1.8 pp, low confidence |
| Nemotron 3.5 Lightning 30B A3B NVFP4 | 74.3% | 55.7% | estimated ± 1.8 pp, low confidence |
| Nemotron 3 Nano 30B | 75.7% | 57.4% | estimated ± 1.8 pp, low confidence |
| Nemotron 3 Nano Omni 30B A3B | 46.9% | 28.9% | estimated ± 1.8 pp, low confidence |
| Nemotron 3 Super 100B | 80.0% | 62.8% | estimated ± 1.8 pp, low confidence |
| Nemotron Ultra 253B | 72.8% | 53.9% | estimated ± 1.8 pp, low confidence |
| North Mini Code | 75.7% | 57.4% | estimated ± 1.8 pp, low confidence |
| Nova Pro | 49.9% | 31.3% | estimated ± 1.8 pp, low confidence |
| o1 | 74.7% | 56.1% | estimated ± 1.8 pp, low confidence |
| o1-preview | 76.5% | 58.3% | estimated ± 1.8 pp, low confidence |
| o3 | 82.7% | 66.4% | estimated ± 1.8 pp, low confidence |
| o3-mini | 74.8% | 56.3% | estimated ± 1.8 pp, low confidence |
| o3-pro | 84.5% | 68.9% | estimated ± 1.8 pp, low confidence |
| Phi-4 | 57.5% | 38.0% | estimated ± 1.8 pp, low confidence |
| Phi-4 Multimodal Instruct | 31.5% | 17.6% | estimated ± 1.8 pp, low confidence |
| Quasar 438B | 73.2% | 54.4% | estimated ± 1.8 pp, low confidence |
| Qwen2.5 Coder 32B Instruct | 41.7% | 24.8% | estimated ± 1.8 pp, low confidence |
| Qwen3.5-122B-A10B | 85.7% | 70.6% | estimated ± 1.8 pp, low confidence |
| Qwen3.5-27B | 85.8% | 70.8% | estimated ± 1.8 pp, low confidence |
| Qwen3.5-35B-A3B | 84.5% | 68.9% | estimated ± 1.8 pp, low confidence |
| Qwen3.5 397B | 86.1% | 71.2% | estimated ± 1.8 pp, low confidence |
| Qwen3.5 397B (Reasoning) | 86.1% | 71.2% | estimated ± 1.8 pp, low confidence |
| Qwen3.6-27B | 84.2% | 68.5% | estimated ± 1.8 pp, low confidence |
| Qwen3.6-35B-A3B | 84.1% | 68.3% | estimated ± 1.8 pp, low confidence |
| Qwen 3.6 Max (preview) | 88.8% | 75.2% | estimated ± 1.8 pp, low confidence |
| Qwen3.7 Plus | 90.0% | 77.1% | estimated ± 1.8 pp, low confidence |
| Qwen3.8-Flash-Next | 92.3% | 80.7% | estimated ± 1.8 pp, low confidence |
| Qwen3.8 Max Preview | 92.8% | 81.5% | estimated ± 1.8 pp, low confidence |
| Qwen3 Max | 76.4% | 58.2% | estimated ± 1.8 pp, low confidence |
| Qwen3-Omni-30B-A3B-Instruct | 62.0% | 42.3% | estimated ± 1.8 pp, low confidence |
| Qwen3-Omni-30B-A3B-Thinking | 72.6% | 53.7% | estimated ± 1.8 pp, low confidence |
| Sarvam 105B | 73.8% | 55.1% | estimated ± 1.8 pp, low confidence |
| Sarvam 30B | 63.3% | 43.6% | estimated ± 1.8 pp, low confidence |
| Solar Pro 2 | 56.1% | 36.7% | estimated ± 1.8 pp, low confidence |
| Solar Pro 3 | 72.4% | 53.4% | estimated ± 1.8 pp, low confidence |
| Solar Pro 4 | 89.1% | 75.7% | estimated ± 1.8 pp, low confidence |
| Step 3.7 Flash | 80.9% | 64.0% | estimated ± 1.8 pp, low confidence |
| Trinity-Large-Preview | 75.2% | 56.7% | estimated ± 1.8 pp, low confidence |
| Trinity-Large-Thinking | 75.2% | 56.7% | estimated ± 1.8 pp, low confidence |
| Ultravox v0.6 Llama 3.3 70B | 49.8% | 31.2% | estimated ± 1.8 pp, low confidence |