Calibration
AA-GPQA Diamond → Vals GPQA Diamond
Vals GPQA Diamond is estimated from AA-GPQA Diamond with a inverse Michaelis–Menten curve fitted on 47 models measured on both: y = 1.08404·(x − 0.0000) / (0.0000 + 2.0000 − x), R² = 0.53, cross-validated error 7.4 pp. It is used for 86 estimates.
| Estimated model | AA-GPQA Diamond | Vals GPQA Diamond | Source |
|---|---|---|---|
| Apodex 1.1 | 86.4% | 82.4% | estimated ± 7.4 pp, medium confidence |
| Apodex 1.1 Mini | 86.4% | 82.4% | estimated ± 7.4 pp, medium confidence |
| Celeris-1 | 63.1% | 50.0% | estimated ± 7.4 pp, low confidence |
| Claude 3 Haiku | 37.4% | 24.9% | estimated ± 7.4 pp, low confidence |
| Claude 3 Opus | 48.9% | 35.1% | estimated ± 7.4 pp, low confidence |
| Claude 4.1 Opus Thinking | 80.9% | 73.6% | estimated ± 7.4 pp, medium confidence |
| Claude 4 Sonnet | 68.3% | 56.2% | estimated ± 7.4 pp, medium confidence |
| Command A+ | 76.1% | 66.6% | estimated ± 7.4 pp, medium confidence |
| DeepSeek-R1 | 81.3% | 74.2% | estimated ± 7.4 pp, medium confidence |
| DeepSeek R1 Distill Qwen 32B | 61.5% | 48.1% | estimated ± 7.4 pp, low confidence |
| DeepSeek V3 0324 | 65.5% | 52.8% | estimated ± 7.4 pp, medium confidence |
| DeepSeek V3.1 | 73.5% | 63.0% | estimated ± 7.4 pp, medium confidence |
| DeepSeek V3.1 (Reasoning) | 77.9% | 69.2% | estimated ± 7.4 pp, medium confidence |
| DeepSeek V3.2 | 75.1% | 65.2% | estimated ± 7.4 pp, medium confidence |
| Exaone 4.0 1.2B | 42.4% | 29.2% | estimated ± 7.4 pp, low confidence |
| Exaone 4.0 32B | 62.8% | 49.6% | estimated ± 7.4 pp, low confidence |
| Gemini 1.0 Pro | 27.7% | 17.4% | estimated ± 7.4 pp, low confidence |
| Gemini 1.5 Pro | 58.9% | 45.3% | estimated ± 7.4 pp, low confidence |
| Gemini 2.5 Flash | 68.3% | 56.2% | estimated ± 7.4 pp, medium confidence |
| Gemma 3 27B | 42.8% | 29.5% | estimated ± 7.4 pp, low confidence |
| GLM-4.5-Air | 73.3% | 62.7% | estimated ± 7.4 pp, medium confidence |
| GLM-5-Turbo | 84.7% | 79.6% | estimated ± 7.4 pp, medium confidence |
| GLM-5V-Turbo | 80.9% | 73.6% | estimated ± 7.4 pp, medium confidence |
| GPT-4o | 54.3% | 40.4% | estimated ± 7.4 pp, low confidence |
| GPT-4o mini | 42.6% | 29.3% | estimated ± 7.4 pp, low confidence |
| GPT-5.1-Codex | 86.0% | 81.8% | estimated ± 7.4 pp, medium confidence |
| GPT-5.1-Codex-Max | 86.0% | 81.8% | estimated ± 7.4 pp, medium confidence |
| GPT-5.2-Codex | 89.9% | 88.5% | estimated ± 7.4 pp, medium confidence |
| GPT-5.3 Codex | 91.5% | 91.4% | estimated ± 7.4 pp, medium confidence |
| GPT-5 (high) | 85.4% | 80.8% | estimated ± 7.4 pp, medium confidence |
| GPT-5 (medium) | 84.2% | 78.8% | estimated ± 7.4 pp, medium confidence |
| GPT-OSS 120B | 78.2% | 69.6% | estimated ± 7.4 pp, medium confidence |
| GPT-OSS 20B | 68.8% | 56.8% | estimated ± 7.4 pp, medium confidence |
| Granite-4.0-350M | 26.1% | 16.3% | estimated ± 7.4 pp, low confidence |
| Granite-4.0-H-1B | 26.3% | 16.4% | estimated ± 7.4 pp, low confidence |
| Granite-4.0-H-350M | 25.7% | 16.0% | estimated ± 7.4 pp, low confidence |
| Grok 4 | 87.7% | 84.7% | estimated ± 7.4 pp, medium confidence |
| Grok 4.1 Fast | 63.7% | 50.7% | estimated ± 7.4 pp, medium confidence |
| Grok 4.1 Fast (Reasoning) | 85.3% | 80.6% | estimated ± 7.4 pp, medium confidence |
| Grok 4 Fast (Reasoning) | 84.7% | 79.6% | estimated ± 7.4 pp, medium confidence |
| Grok Code Fast 1 | 72.7% | 61.9% | estimated ± 7.4 pp, medium confidence |
| Hy3 | 89.7% | 88.2% | estimated ± 7.4 pp, medium confidence |
| K-Exaone | 78.3% | 69.7% | estimated ± 7.4 pp, medium confidence |
| Kimi K2 | 76.6% | 67.3% | estimated ± 7.4 pp, medium confidence |
| Kimi K2.7 Code | 89.6% | 88.0% | estimated ± 7.4 pp, medium confidence |
| LFM2.5-2.6B | 55.8% | 41.9% | estimated ± 7.4 pp, low confidence |
| LFM2.5-8B-A1B | 51.3% | 37.4% | estimated ± 7.4 pp, low confidence |
| LFM2.5-VL-1.6B-Extract | 28.9% | 18.3% | estimated ± 7.4 pp, low confidence |
| Ling 3.0 Flash VL | 86.2% | 82.1% | estimated ± 7.4 pp, medium confidence |
| Ling 3.0 Tiny | 73.4% | 62.9% | estimated ± 7.4 pp, medium confidence |
| Llama 3.1 405B | 51.5% | 37.6% | estimated ± 7.4 pp, low confidence |
| Llama 4 Maverick | 67.1% | 54.7% | estimated ± 7.4 pp, medium confidence |
| Llama 4 Scout | 58.7% | 45.0% | estimated ± 7.4 pp, low confidence |
| MiMo-V2-Omni | 82.8% | 76.6% | estimated ± 7.4 pp, medium confidence |
| MiMo-V2-Pro | 87.0% | 83.5% | estimated ± 7.4 pp, medium confidence |
| Mistral Large 2 | 48.6% | 34.8% | estimated ± 7.4 pp, low confidence |
| Mistral Large 3 | 68.0% | 55.8% | estimated ± 7.4 pp, medium confidence |
| Mistral Medium 3 | 57.8% | 44.1% | estimated ± 7.4 pp, low confidence |
| Mistral Small 4 | 76.9% | 67.7% | estimated ± 7.4 pp, medium confidence |
| Mistral Small 4 (Reasoning) | 76.9% | 67.7% | estimated ± 7.4 pp, medium confidence |
| Muse Glimmer 30B | 83.5% | 77.7% | estimated ± 7.4 pp, medium confidence |
| Muse Spark 1.3 | 93.5% | 95.2% | estimated ± 7.4 pp, medium confidence |
| Nemotron 3 Nano 30B | 75.7% | 66.0% | estimated ± 7.4 pp, medium confidence |
| Nemotron 3 Super 100B | 80.0% | 72.3% | estimated ± 7.4 pp, medium confidence |
| Nemotron Ultra 253B | 72.8% | 62.0% | estimated ± 7.4 pp, medium confidence |
| North Mini Code | 75.7% | 66.0% | estimated ± 7.4 pp, medium confidence |
| Nova Pro | 49.9% | 36.0% | estimated ± 7.4 pp, low confidence |
| o1-preview | 76.5% | 67.1% | estimated ± 7.4 pp, medium confidence |
| o3 | 82.7% | 76.4% | estimated ± 7.4 pp, medium confidence |
| o3-pro | 84.5% | 79.3% | estimated ± 7.4 pp, medium confidence |
| Phi-4 | 57.5% | 43.7% | estimated ± 7.4 pp, low confidence |
| Phi-4 Multimodal Instruct | 31.5% | 20.3% | estimated ± 7.4 pp, low confidence |
| Quasar 438B | 73.2% | 62.6% | estimated ± 7.4 pp, medium confidence |
| Qwen2.5 Coder 32B Instruct | 41.7% | 28.6% | estimated ± 7.4 pp, low confidence |
| Qwen3.5 397B (Reasoning) | 86.1% | 81.9% | estimated ± 7.4 pp, medium confidence |
| Qwen 3.6 Max (preview) | 88.8% | 86.6% | estimated ± 7.4 pp, medium confidence |
| Qwen3.8 Max Preview | 92.8% | 93.8% | estimated ± 7.4 pp, medium confidence |
| Qwen3 Max | 76.4% | 67.0% | estimated ± 7.4 pp, medium confidence |
| Qwen3-Omni-30B-A3B-Instruct | 62.0% | 48.7% | estimated ± 7.4 pp, low confidence |
| Qwen3-Omni-30B-A3B-Thinking | 72.6% | 61.8% | estimated ± 7.4 pp, medium confidence |
| Sarvam 105B | 73.8% | 63.4% | estimated ± 7.4 pp, medium confidence |
| Sarvam 30B | 63.3% | 50.2% | estimated ± 7.4 pp, medium confidence |
| Solar Pro 2 | 56.1% | 42.3% | estimated ± 7.4 pp, low confidence |
| Solar Pro 3 | 72.4% | 61.5% | estimated ± 7.4 pp, medium confidence |
| Step 3.7 Flash | 80.9% | 73.6% | estimated ± 7.4 pp, medium confidence |
| Ultravox v0.6 Llama 3.3 70B | 49.8% | 35.9% | estimated ± 7.4 pp, low confidence |