Calibration
GDPval-AA → ResearchClawBench
ResearchClawBench is estimated from GDPval-AA with a Michaelis–Menten curve fitted on 13 models measured on both: y = 0.2430·x / (0.12373 + x), R² = 0.45, cross-validated error 2.2 pp. It is used for 40 estimates.
| Estimated model | GDPval-AA | ResearchClawBench | Source |
|---|---|---|---|
| A.X K2 | 23.2% | 15.8% | estimated ± 2.2 pp, medium confidence |
| Apodex 1.1 | 34.8% | 17.9% | estimated ± 2.2 pp, medium confidence |
| Apodex 1.1 Mini | 34.8% | 17.9% | estimated ± 2.2 pp, medium confidence |
| Claude Opus 5.5 | 68.3% | 20.6% | estimated ± 2.2 pp, low confidence |
| Claude Sonnet 5.5 | 67.0% | 20.5% | estimated ± 2.2 pp, low confidence |
| DeepSeek V3 0324 | 0.0% | 0.0% | estimated ± 2.2 pp, low confidence |
| DeepSeek V4.1 Flash | 55.0% | 19.8% | estimated ± 2.2 pp, low confidence |
| Gemini 4 Argon | 56.3% | 19.9% | estimated ± 2.2 pp, low confidence |
| Gemma 4 12B | 0.0% | 0.0% | estimated ± 2.2 pp, low confidence |
| Gemma 4 26B A4B | 3.4% | 5.2% | estimated ± 2.2 pp, low confidence |
| Gemma 4 E2B | 0.0% | 0.0% | estimated ± 2.2 pp, low confidence |
| Gemma 4 E4B | 0.0% | 0.0% | estimated ± 2.2 pp, low confidence |
| GPT-4.1 mini | 0.0% | 0.0% | estimated ± 2.2 pp, low confidence |
| GPT-4.1 nano | 0.0% | 0.0% | estimated ± 2.2 pp, low confidence |
| GPT-4o | 0.0% | 0.0% | estimated ± 2.2 pp, low confidence |
| GPT-4o mini | 0.0% | 0.0% | estimated ± 2.2 pp, low confidence |
| GPT-5.1 | 16.5% | 13.9% | estimated ± 2.2 pp, medium confidence |
| GPT-5 (high) | 21.0% | 15.3% | estimated ± 2.2 pp, medium confidence |
| GPT-6.1 Sol | 53.8% | 19.8% | estimated ± 2.2 pp, low confidence |
| GPT-6 Luna | 46.9% | 19.2% | estimated ± 2.2 pp, medium confidence |
| GPT-6 Sol | 50.5% | 19.5% | estimated ± 2.2 pp, low confidence |
| Granite 4.2 30B | 3.9% | 5.8% | estimated ± 2.2 pp, low confidence |
| Granite 4.2 3B | 0.0% | 0.0% | estimated ± 2.2 pp, low confidence |
| Grok 4.7 | 60.8% | 20.2% | estimated ± 2.2 pp, low confidence |
| K-Exaone | 0.0% | 0.0% | estimated ± 2.2 pp, low confidence |
| Ling 2.6 Flash | 0.0% | 0.0% | estimated ± 2.2 pp, low confidence |
| Ling 3.0 Flash VL | 33.2% | 17.7% | estimated ± 2.2 pp, medium confidence |
| Ling 3.0 Tiny | 3.2% | 5.0% | estimated ± 2.2 pp, low confidence |
| Ling 3.1 Flash | 56.1% | 19.9% | estimated ± 2.2 pp, low confidence |
| Mercury 2.5 | 0.0% | 0.0% | estimated ± 2.2 pp, low confidence |
| MiMo-V2.6-Flash | 55.5% | 19.9% | estimated ± 2.2 pp, low confidence |
| MiMo-V2.6-Pro | 59.3% | 20.1% | estimated ± 2.2 pp, low confidence |
| MiMo-V2-Flash | 6.6% | 8.5% | estimated ± 2.2 pp, low confidence |
| MiniCPM5-2B | 10.5% | 11.2% | estimated ± 2.2 pp, low confidence |
| Mistral Large 4 | 46.2% | 19.2% | estimated ± 2.2 pp, medium confidence |
| Nemotron 3 Nano Omni 30B A3B | 0.0% | 0.0% | estimated ± 2.2 pp, low confidence |
| North Mini Code | 0.0% | 0.0% | estimated ± 2.2 pp, low confidence |
| Qwen3.8-Flash-Next | 56.6% | 19.9% | estimated ± 2.2 pp, low confidence |
| Solar Pro 3 | 0.0% | 0.0% | estimated ± 2.2 pp, low confidence |
| Ultravox v0.6 Llama 3.3 70B | 0.0% | 0.0% | estimated ± 2.2 pp, low confidence |