Calibration
τ²-bench results → GDPval-AA
GDPval-AA is estimated from τ²-bench results with a Hill curve fitted on 66 models measured on both: y = 0.0000 + (1.2000 − 0.0000)·x^5.55 / (1.16845^5.55 + x^5.55), R² = 0.56, cross-validated error 12.5 pp. It is used for 38 estimates.
| Estimated model | τ²-bench results | GDPval-AA | Source |
|---|---|---|---|
| Claude 3 Haiku | 21.1% | 0.0% | estimated ± 12.5 pp, low confidence |
| Claude 4.1 Opus Thinking | 71.4% | 7.3% | estimated ± 12.5 pp, low confidence |
| Claude Opus 4.5 Thinking | 89.5% | 22.3% | estimated ± 12.5 pp, low confidence |
| DeepSeek-R1 | 36.5% | 0.2% | estimated ± 12.5 pp, low confidence |
| DeepSeek V3.1 | 34.8% | 0.1% | estimated ± 12.5 pp, low confidence |
| DeepSeek V3.1 (Reasoning) | 37.4% | 0.2% | estimated ± 12.5 pp, low confidence |
| Exaone 4.0 1.2B | 20.5% | 0.0% | estimated ± 12.5 pp, low confidence |
| Exaone 4.0 32B | 4.1% | 0.0% | estimated ± 12.5 pp, low confidence |
| Gemini 2.5 Flash | 14.9% | 0.0% | estimated ± 12.5 pp, low confidence |
| GLM-4.5-Air | 46.5% | 0.7% | estimated ± 12.5 pp, low confidence |
| GLM-4.6 | 76.9% | 10.7% | estimated ± 12.5 pp, low confidence |
| GPT-5.1-Codex-Max | 83.0% | 15.6% | estimated ± 12.5 pp, low confidence |
| GPT-5 (medium) | 86.5% | 19.0% | estimated ± 12.5 pp, low confidence |
| Granite-4.0-350M | 13.2% | 0.0% | estimated ± 12.5 pp, low confidence |
| Granite-4.0-H-1B | 19.6% | 0.0% | estimated ± 12.5 pp, low confidence |
| Granite-4.0-H-350M | 14.6% | 0.0% | estimated ± 12.5 pp, low confidence |
| Grok 4.1 Fast (Reasoning) | 93.3% | 26.8% | estimated ± 12.5 pp, low confidence |
| Grok 4 Fast (Reasoning) | 65.8% | 4.8% | estimated ± 12.5 pp, low confidence |
| Grok Code Fast 1 | 75.7% | 9.9% | estimated ± 12.5 pp, low confidence |
| Kimi K2 | 61.1% | 3.2% | estimated ± 12.5 pp, low confidence |
| LFM2.5-VL-1.6B-Extract | 8.5% | 0.0% | estimated ± 12.5 pp, low confidence |
| Llama 3.1 405B | 19.0% | 0.0% | estimated ± 12.5 pp, low confidence |
| Mistral Large 2 | 30.7% | 0.1% | estimated ± 12.5 pp, low confidence |
| Mistral Medium 3 | 24.3% | 0.0% | estimated ± 12.5 pp, low confidence |
| Nemotron Ultra 253B | 11.4% | 0.0% | estimated ± 12.5 pp, low confidence |
| Nova Pro | 14.0% | 0.0% | estimated ± 12.5 pp, low confidence |
| o1 | 62.6% | 3.6% | estimated ± 12.5 pp, low confidence |
| o3 | 80.7% | 13.6% | estimated ± 12.5 pp, low confidence |
| o3-mini | 28.7% | 0.0% | estimated ± 12.5 pp, low confidence |
| Phi-4 | 0.0% | 0.0% | estimated ± 12.5 pp, low confidence |
| Qwen3.5 397B (Reasoning) | 83.9% | 16.5% | estimated ± 12.5 pp, low confidence |
| Qwen 3.6 Max (preview) | 95.9% | 30.1% | estimated ± 12.5 pp, low confidence |
| Qwen3-Omni-30B-A3B-Instruct | 16.4% | 0.0% | estimated ± 12.5 pp, low confidence |
| Qwen3-Omni-30B-A3B-Thinking | 21.3% | 0.0% | estimated ± 12.5 pp, low confidence |
| Sarvam 105B | 46.8% | 0.7% | estimated ± 12.5 pp, low confidence |
| Sarvam 30B | 34.5% | 0.1% | estimated ± 12.5 pp, low confidence |
| Solar Pro 2 | 31.9% | 0.1% | estimated ± 12.5 pp, low confidence |
| Ternary Bonsai 2 27B | 80.2% | 13.2% | estimated ± 12.5 pp, low confidence |