Calibration
τ²-bench results → AA ITBench
AA ITBench is estimated from τ²-bench results with a Hill curve fitted on 5 models measured on both: y = 1.0000 + (0.3440 − 1.0000)·x^6.00 / (0.72970^6.00 + x^6.00), R² = 0.77, cross-validated error 3.6 pp. It is used for 107 estimates.
| Estimated model | τ²-bench results | AA ITBench | Source |
|---|---|---|---|
| Agents-A1-4B | 78.2% | 60.5% | estimated ± 3.6 pp, low confidence |
| Claude 3 Haiku | 21.1% | 100.0% | estimated ± 3.6 pp, low confidence |
| Claude 4.1 Opus Thinking | 71.4% | 69.3% | estimated ± 3.6 pp, low confidence |
| Claude 4 Sonnet | 52.3% | 92.2% | estimated ± 3.6 pp, low confidence |
| Claude Fable 5 | 98.5% | 43.7% | estimated ± 3.6 pp, medium confidence |
| Claude Opus 4.5 Thinking | 89.5% | 49.3% | estimated ± 3.6 pp, medium confidence |
| Claude Opus 4.6 | 84.8% | 53.3% | estimated ± 3.6 pp, low confidence |
| Claude Opus 4.6 (Adaptive) | 92.1% | 47.4% | estimated ± 3.6 pp, medium confidence |
| Claude Opus 4.7 | 74.0% | 65.8% | estimated ± 3.6 pp, low confidence |
| Claude Sonnet 4.6 | 79.5% | 58.9% | estimated ± 3.6 pp, low confidence |
| Command A+ | 85.0% | 53.1% | estimated ± 3.6 pp, low confidence |
| DeepSeek-R1 | 36.5% | 99.0% | estimated ± 3.6 pp, low confidence |
| DeepSeek V3 | 22.8% | 99.9% | estimated ± 3.6 pp, low confidence |
| DeepSeek V3 0324 | 47.1% | 95.6% | estimated ± 3.6 pp, low confidence |
| DeepSeek V3.1 | 34.8% | 99.2% | estimated ± 3.6 pp, low confidence |
| DeepSeek V3.1 (Reasoning) | 37.4% | 98.8% | estimated ± 3.6 pp, low confidence |
| DeepSeek V3.2 | 78.9% | 59.6% | estimated ± 3.6 pp, low confidence |
| Exaone 4.0 1.2B | 20.5% | 100.0% | estimated ± 3.6 pp, low confidence |
| Exaone 4.0 32B | 4.1% | 100.0% | estimated ± 3.6 pp, low confidence |
| Gemini 2.5 Flash | 14.9% | 100.0% | estimated ± 3.6 pp, low confidence |
| Gemini 2.5 Pro | 54.1% | 90.7% | estimated ± 3.6 pp, low confidence |
| Gemini 3.1 Pro | 95.6% | 45.2% | estimated ± 3.6 pp, medium confidence |
| Gemini 3 Flash | 43.3% | 97.3% | estimated ± 3.6 pp, low confidence |
| Gemini 3 Pro | 87.1% | 51.3% | estimated ± 3.6 pp, medium confidence |
| Gemma 3 27B | 10.5% | 100.0% | estimated ± 3.6 pp, low confidence |
| Gemma 4 12B | 36.3% | 99.0% | estimated ± 3.6 pp, low confidence |
| Gemma 4 26B A4B | 43.6% | 97.1% | estimated ± 3.6 pp, low confidence |
| Gemma 4 31B | 59.9% | 84.6% | estimated ± 3.6 pp, low confidence |
| Gemma 4 E2B | 20.8% | 100.0% | estimated ± 3.6 pp, low confidence |
| Gemma 4 E4B | 20.8% | 100.0% | estimated ± 3.6 pp, low confidence |
| GLM-4.5-Air | 46.5% | 95.9% | estimated ± 3.6 pp, low confidence |
| GLM-4.6 | 76.9% | 62.1% | estimated ± 3.6 pp, low confidence |
| GLM-4.7 | 95.9% | 45.1% | estimated ± 3.6 pp, medium confidence |
| GLM-5-Turbo | 98.5% | 43.7% | estimated ± 3.6 pp, medium confidence |
| GLM-5V-Turbo | 98.5% | 43.7% | estimated ± 3.6 pp, medium confidence |
| GPT-4.1 | 47.1% | 95.6% | estimated ± 3.6 pp, low confidence |
| GPT-4.1 mini | 52.9% | 91.7% | estimated ± 3.6 pp, low confidence |
| GPT-4.1 nano | 17.3% | 100.0% | estimated ± 3.6 pp, low confidence |
| GPT-4o | 25.1% | 99.9% | estimated ± 3.6 pp, low confidence |
| GPT-5.1 | 81.9% | 56.3% | estimated ± 3.6 pp, low confidence |
| GPT-5.1-Codex | 83.0% | 55.1% | estimated ± 3.6 pp, low confidence |
| GPT-5.1-Codex-Max | 83.0% | 55.1% | estimated ± 3.6 pp, low confidence |
| GPT-5.2 | 84.8% | 53.3% | estimated ± 3.6 pp, low confidence |
| GPT-5.2-Codex | 92.1% | 47.4% | estimated ± 3.6 pp, medium confidence |
| GPT-5.3 Codex | 86.0% | 52.2% | estimated ± 3.6 pp, medium confidence |
| GPT-5 (high) | 84.8% | 53.3% | estimated ± 3.6 pp, low confidence |
| GPT-5 (medium) | 86.5% | 51.8% | estimated ± 3.6 pp, medium confidence |
| GPT-OSS 120B | 65.8% | 77.1% | estimated ± 3.6 pp, low confidence |
| GPT-OSS 20B | 60.2% | 84.3% | estimated ± 3.6 pp, low confidence |
| Granite-4.0-350M | 13.2% | 100.0% | estimated ± 3.6 pp, low confidence |
| Granite-4.0-H-1B | 19.6% | 100.0% | estimated ± 3.6 pp, low confidence |
| Granite-4.0-H-350M | 14.6% | 100.0% | estimated ± 3.6 pp, low confidence |
| Grok 4 | 74.9% | 64.6% | estimated ± 3.6 pp, low confidence |
| Grok 4.1 Fast | 63.7% | 79.9% | estimated ± 3.6 pp, low confidence |
| Grok 4.1 Fast (Reasoning) | 93.3% | 46.6% | estimated ± 3.6 pp, medium confidence |
| Grok 4.3 | 97.7% | 44.1% | estimated ± 3.6 pp, medium confidence |
| Grok 4 Fast (Reasoning) | 65.8% | 77.1% | estimated ± 3.6 pp, low confidence |
| Grok Code Fast 1 | 75.7% | 63.6% | estimated ± 3.6 pp, low confidence |
| K-Exaone | 74.3% | 65.4% | estimated ± 3.6 pp, low confidence |
| Kimi K2 | 61.1% | 83.2% | estimated ± 3.6 pp, low confidence |
| Kimi K2.5 (Reasoning) | 95.9% | 45.1% | estimated ± 3.6 pp, medium confidence |
| LFM2.5-8B-A1B | 88.1% | 50.4% | estimated ± 3.6 pp, medium confidence |
| LFM2.5-VL-1.6B-Extract | 8.5% | 100.0% | estimated ± 3.6 pp, low confidence |
| Ling 2.6 Flash | 86.0% | 52.2% | estimated ± 3.6 pp, medium confidence |
| LLaDA2.2-mini | 57.5% | 87.3% | estimated ± 3.6 pp, low confidence |
| Llama 3.1 405B | 19.0% | 100.0% | estimated ± 3.6 pp, low confidence |
| Llama 4 Maverick | 17.8% | 100.0% | estimated ± 3.6 pp, low confidence |
| Llama 4 Scout | 15.5% | 100.0% | estimated ± 3.6 pp, low confidence |
| MiMo-V2.5-Pro | 94.2% | 46.1% | estimated ± 3.6 pp, medium confidence |
| MiMo-V2-Flash | 83.9% | 54.2% | estimated ± 3.6 pp, low confidence |
| MiMo-V2-Omni | 91.2% | 48.0% | estimated ± 3.6 pp, medium confidence |
| MiMo-V2-Pro | 95.0% | 45.6% | estimated ± 3.6 pp, medium confidence |
| MiniMax M2.7 | 84.8% | 53.3% | estimated ± 3.6 pp, low confidence |
| Mistral Large 2 | 30.7% | 99.6% | estimated ± 3.6 pp, low confidence |
| Mistral Large 3 | 24.6% | 99.9% | estimated ± 3.6 pp, low confidence |
| Mistral Medium 3 | 24.3% | 99.9% | estimated ± 3.6 pp, low confidence |
| Mistral Medium 3.5 128B | 94.2% | 46.1% | estimated ± 3.6 pp, medium confidence |
| Mistral Small 4 | 41.2% | 97.9% | estimated ± 3.6 pp, low confidence |
| Mistral Small 4 (Reasoning) | 41.2% | 97.9% | estimated ± 3.6 pp, low confidence |
| Muse Spark | 91.5% | 47.8% | estimated ± 3.6 pp, medium confidence |
| Nemotron 3 Nano 30B | 40.9% | 98.0% | estimated ± 3.6 pp, low confidence |
| Nemotron 3 Nano Omni 30B A3B | 45.3% | 96.4% | estimated ± 3.6 pp, low confidence |
| Nemotron 3 Super 100B | 67.8% | 74.3% | estimated ± 3.6 pp, low confidence |
| Nemotron Ultra 253B | 11.4% | 100.0% | estimated ± 3.6 pp, low confidence |
| North Mini Code | 37.4% | 98.8% | estimated ± 3.6 pp, low confidence |
| Nova Pro | 14.0% | 100.0% | estimated ± 3.6 pp, low confidence |
| o1 | 62.6% | 81.3% | estimated ± 3.6 pp, low confidence |
| o3 | 80.7% | 57.6% | estimated ± 3.6 pp, low confidence |
| o3-mini | 28.7% | 99.8% | estimated ± 3.6 pp, low confidence |
| Phi-4 | 0.0% | 100.0% | estimated ± 3.6 pp, low confidence |
| Qwen3.5-122B-A10B | 93.6% | 46.4% | estimated ± 3.6 pp, medium confidence |
| Qwen3.5-27B | 93.9% | 46.2% | estimated ± 3.6 pp, medium confidence |
| Qwen3.5-35B-A3B | 89.2% | 49.5% | estimated ± 3.6 pp, medium confidence |
| Qwen3.5 397B (Reasoning) | 83.9% | 54.2% | estimated ± 3.6 pp, low confidence |
| Qwen3.6-27B | 94.2% | 46.1% | estimated ± 3.6 pp, medium confidence |
| Qwen 3.6 Max (preview) | 95.9% | 45.1% | estimated ± 3.6 pp, medium confidence |
| Qwen3 Max | 74.3% | 65.4% | estimated ± 3.6 pp, low confidence |
| Qwen3-Omni-30B-A3B-Instruct | 16.4% | 100.0% | estimated ± 3.6 pp, low confidence |
| Qwen3-Omni-30B-A3B-Thinking | 21.3% | 100.0% | estimated ± 3.6 pp, low confidence |
| Sarvam 105B | 46.8% | 95.7% | estimated ± 3.6 pp, low confidence |
| Sarvam 30B | 34.5% | 99.3% | estimated ± 3.6 pp, low confidence |
| Solar Pro 2 | 31.9% | 99.5% | estimated ± 3.6 pp, low confidence |
| Solar Pro 3 | 86.3% | 52.0% | estimated ± 3.6 pp, medium confidence |
| Ternary Bonsai 2 27B | 80.2% | 58.1% | estimated ± 3.6 pp, low confidence |
| Trinity-Large-Preview | 90.1% | 48.8% | estimated ± 3.6 pp, medium confidence |
| Trinity-Large-Thinking | 90.1% | 48.8% | estimated ± 3.6 pp, medium confidence |
| Ultravox v0.6 Llama 3.3 70B | 26.6% | 99.8% | estimated ± 3.6 pp, low confidence |