Calibration
AA-IFBench → IFBench
IFBench is estimated from AA-IFBench with a Hill curve fitted on 11 models measured on both: y = 0.5529 + (1.2000 − 0.5529)·x^6.00 / (0.87644^6.00 + x^6.00), R² = 0.71, cross-validated error 6.9 pp. It is used for 119 estimates.
| Estimated model | AA-IFBench | IFBench | Source |
|---|---|---|---|
| Claude 3 Haiku | 36.1% | 55.6% | estimated ± 6.9 pp, low confidence |
| Claude 4.1 Opus Thinking | 55.4% | 59.2% | estimated ± 6.9 pp, medium confidence |
| Claude 4 Sonnet | 45.4% | 56.5% | estimated ± 6.9 pp, medium confidence |
| Claude Fable 5 | 63.5% | 63.5% | estimated ± 6.9 pp, medium confidence |
| Claude Opus 4.5 Thinking | 58.0% | 60.3% | estimated ± 6.9 pp, medium confidence |
| Claude Opus 4.6 | 44.6% | 56.4% | estimated ± 6.9 pp, medium confidence |
| Claude Opus 4.6 (Adaptive) | 53.1% | 58.3% | estimated ± 6.9 pp, medium confidence |
| Claude Opus 4.7 | 43.6% | 56.3% | estimated ± 6.9 pp, medium confidence |
| Claude Opus 4.7 (Adaptive) | 58.6% | 60.6% | estimated ± 6.9 pp, medium confidence |
| Claude Opus 4.8 | 62.2% | 62.6% | estimated ± 6.9 pp, medium confidence |
| Claude Sonnet 4.6 | 41.2% | 56.0% | estimated ± 6.9 pp, low confidence |
| Command A+ | 73.9% | 72.4% | estimated ± 6.9 pp, medium confidence |
| DeepSeek-R1 | 39.6% | 55.8% | estimated ± 6.9 pp, low confidence |
| DeepSeek R1 Distill Qwen 32B | 22.9% | 55.3% | estimated ± 6.9 pp, low confidence |
| DeepSeek V3 | 34.8% | 55.5% | estimated ± 6.9 pp, low confidence |
| DeepSeek V3 0324 | 41.0% | 56.0% | estimated ± 6.9 pp, low confidence |
| DeepSeek V3.1 | 37.8% | 55.7% | estimated ± 6.9 pp, low confidence |
| DeepSeek V3.1 (Reasoning) | 41.5% | 56.0% | estimated ± 6.9 pp, low confidence |
| DeepSeek V3.2 | 49.0% | 57.2% | estimated ± 6.9 pp, medium confidence |
| DeepSeek V4 Pro 0813 | 76.5% | 75.1% | estimated ± 6.9 pp, medium confidence |
| Exaone 4.0 1.2B | 25.3% | 55.3% | estimated ± 6.9 pp, low confidence |
| Exaone 4.0 32B | 33.5% | 55.5% | estimated ± 6.9 pp, low confidence |
| Gemini 2.5 Flash | 39.0% | 55.8% | estimated ± 6.9 pp, low confidence |
| Gemini 2.5 Pro | 48.7% | 57.1% | estimated ± 6.9 pp, medium confidence |
| Gemini 3.1 Pro | 77.1% | 75.8% | estimated ± 6.9 pp, medium confidence |
| Gemini 3 Flash | 55.1% | 59.1% | estimated ± 6.9 pp, medium confidence |
| Gemini 3 Pro | 70.4% | 69.0% | estimated ± 6.9 pp, medium confidence |
| Gemma 3 27B | 31.8% | 55.4% | estimated ± 6.9 pp, low confidence |
| Gemma 4 12B | 73.5% | 72.0% | estimated ± 6.9 pp, medium confidence |
| Gemma 4 26B A4B | 72.4% | 70.9% | estimated ± 6.9 pp, medium confidence |
| Gemma 4 31B | 75.6% | 74.2% | estimated ± 6.9 pp, medium confidence |
| Gemma 4 E2B | 38.0% | 55.7% | estimated ± 6.9 pp, low confidence |
| Gemma 4 E4B | 44.2% | 56.3% | estimated ± 6.9 pp, medium confidence |
| GLM-4.5-Air | 37.6% | 55.7% | estimated ± 6.9 pp, low confidence |
| GLM-4.6 | 36.7% | 55.6% | estimated ± 6.9 pp, low confidence |
| GLM-4.7 | 67.9% | 66.8% | estimated ± 6.9 pp, medium confidence |
| GLM-5 | 72.3% | 70.8% | estimated ± 6.9 pp, medium confidence |
| GLM-5.1 | 76.3% | 74.9% | estimated ± 6.9 pp, medium confidence |
| GLM-5.2 | 73.3% | 71.8% | estimated ± 6.9 pp, medium confidence |
| GLM-5-Turbo | 73.2% | 71.7% | estimated ± 6.9 pp, medium confidence |
| GLM-5V-Turbo | 61.1% | 62.0% | estimated ± 6.9 pp, medium confidence |
| GPT-4.1 | 43.0% | 56.2% | estimated ± 6.9 pp, medium confidence |
| GPT-4.1 mini | 38.3% | 55.7% | estimated ± 6.9 pp, low confidence |
| GPT-4.1 nano | 32.0% | 55.4% | estimated ± 6.9 pp, low confidence |
| GPT-4o | 34.3% | 55.5% | estimated ± 6.9 pp, low confidence |
| GPT-4o mini | 31.0% | 55.4% | estimated ± 6.9 pp, low confidence |
| GPT-5.1 | 72.9% | 71.4% | estimated ± 6.9 pp, medium confidence |
| GPT-5.1-Codex | 70.0% | 68.6% | estimated ± 6.9 pp, medium confidence |
| GPT-5.1-Codex-Max | 70.0% | 68.6% | estimated ± 6.9 pp, medium confidence |
| GPT-5.2 | 75.4% | 74.0% | estimated ± 6.9 pp, medium confidence |
| GPT-5.2-Codex | 77.6% | 76.3% | estimated ± 6.9 pp, medium confidence |
| GPT-5.3 Codex | 75.4% | 74.0% | estimated ± 6.9 pp, medium confidence |
| GPT-5.4 | 73.9% | 72.4% | estimated ± 6.9 pp, medium confidence |
| GPT-5.4 mini | 73.3% | 71.8% | estimated ± 6.9 pp, medium confidence |
| GPT-5.4 nano | 75.9% | 74.5% | estimated ± 6.9 pp, medium confidence |
| GPT-5.5 | 75.9% | 74.5% | estimated ± 6.9 pp, medium confidence |
| GPT-5.6 Sol | 72.7% | 71.2% | estimated ± 6.9 pp, medium confidence |
| GPT-5.6 Terra | 71.2% | 69.7% | estimated ± 6.9 pp, medium confidence |
| GPT-5 (high) | 73.1% | 71.6% | estimated ± 6.9 pp, medium confidence |
| GPT-5 (medium) | 70.6% | 69.2% | estimated ± 6.9 pp, medium confidence |
| GPT-OSS 120B | 69.0% | 67.7% | estimated ± 6.9 pp, medium confidence |
| GPT-OSS 20B | 65.1% | 64.6% | estimated ± 6.9 pp, medium confidence |
| Granite-4.0-350M | 15.9% | 55.3% | estimated ± 6.9 pp, low confidence |
| Granite-4.0-H-1B | 26.2% | 55.3% | estimated ± 6.9 pp, low confidence |
| Granite-4.0-H-350M | 17.6% | 55.3% | estimated ± 6.9 pp, low confidence |
| Grok 4 | 53.7% | 58.5% | estimated ± 6.9 pp, medium confidence |
| Grok 4.1 Fast | 36.5% | 55.6% | estimated ± 6.9 pp, low confidence |
| Grok 4.1 Fast (Reasoning) | 52.7% | 58.2% | estimated ± 6.9 pp, medium confidence |
| Grok 4 Fast (Reasoning) | 50.5% | 57.6% | estimated ± 6.9 pp, medium confidence |
| Grok Code Fast 1 | 41.4% | 56.0% | estimated ± 6.9 pp, low confidence |
| K-Exaone | 64.7% | 64.3% | estimated ± 6.9 pp, medium confidence |
| Kimi K2.6 | 76.0% | 74.6% | estimated ± 6.9 pp, medium confidence |
| Kimi K2 | 41.5% | 56.0% | estimated ± 6.9 pp, low confidence |
| Kimi K2.5 | 70.2% | 68.8% | estimated ± 6.9 pp, medium confidence |
| Kimi K2.5 (Reasoning) | 70.2% | 68.8% | estimated ± 6.9 pp, medium confidence |
| Kimi K2.7 Code | 63.1% | 63.2% | estimated ± 6.9 pp, medium confidence |
| LFM2.5-VL-1.6B-Extract | 33.1% | 55.5% | estimated ± 6.9 pp, low confidence |
| Llama 3.1 405B | 39.0% | 55.8% | estimated ± 6.9 pp, low confidence |
| Llama 4 Maverick | 43.0% | 56.2% | estimated ± 6.9 pp, medium confidence |
| Llama 4 Scout | 39.5% | 55.8% | estimated ± 6.9 pp, low confidence |
| MiMo-V2.5-Pro | 79.9% | 78.9% | estimated ± 6.9 pp, medium confidence |
| MiMo-V2-Flash | 39.9% | 55.9% | estimated ± 6.9 pp, low confidence |
| MiMo-V2-Omni | 53.5% | 58.5% | estimated ± 6.9 pp, medium confidence |
| MiMo-V2-Pro | 68.8% | 67.6% | estimated ± 6.9 pp, medium confidence |
| MiniMax M2.7 | 75.7% | 74.3% | estimated ± 6.9 pp, medium confidence |
| MiniMax M3 | 82.9% | 82.3% | estimated ± 6.9 pp, low confidence |
| Mistral Large 2 | 31.2% | 55.4% | estimated ± 6.9 pp, low confidence |
| Mistral Large 3 | 36.2% | 55.6% | estimated ± 6.9 pp, low confidence |
| Mistral Medium 3 | 39.3% | 55.8% | estimated ± 6.9 pp, low confidence |
| Mistral Medium 3.5 128B | 68.8% | 67.6% | estimated ± 6.9 pp, medium confidence |
| Mistral Small 4 | 48.2% | 57.0% | estimated ± 6.9 pp, medium confidence |
| Mistral Small 4 (Reasoning) | 48.2% | 57.0% | estimated ± 6.9 pp, medium confidence |
| Muse Spark | 75.9% | 74.5% | estimated ± 6.9 pp, medium confidence |
| Nemotron 3 Nano 30B | 71.1% | 69.6% | estimated ± 6.9 pp, medium confidence |
| Nemotron 3 Super 100B | 71.5% | 70.0% | estimated ± 6.9 pp, medium confidence |
| Nemotron Ultra 253B | 38.2% | 55.7% | estimated ± 6.9 pp, low confidence |
| North Mini Code | 57.6% | 60.1% | estimated ± 6.9 pp, medium confidence |
| Nova Pro | 38.1% | 55.7% | estimated ± 6.9 pp, low confidence |
| o1 | 70.3% | 68.9% | estimated ± 6.9 pp, medium confidence |
| o3 | 71.4% | 69.9% | estimated ± 6.9 pp, medium confidence |
| Phi-4 | 23.5% | 55.3% | estimated ± 6.9 pp, low confidence |
| Qwen3.5-122B-A10B | 75.7% | 74.3% | estimated ± 6.9 pp, medium confidence |
| Qwen3.5-27B | 75.6% | 74.2% | estimated ± 6.9 pp, medium confidence |
| Qwen3.5-35B-A3B | 72.5% | 71.0% | estimated ± 6.9 pp, medium confidence |
| Qwen3.5 397B | 51.6% | 57.9% | estimated ± 6.9 pp, medium confidence |
| Qwen3.5 397B (Reasoning) | 51.6% | 57.9% | estimated ± 6.9 pp, medium confidence |
| Qwen3.6-27B | 67.6% | 66.5% | estimated ± 6.9 pp, medium confidence |
| Qwen3.6-35B-A3B | 64.4% | 64.1% | estimated ± 6.9 pp, medium confidence |
| Qwen 3.6 Max (preview) | 76.6% | 75.2% | estimated ± 6.9 pp, medium confidence |
| Qwen3 Max | 44.1% | 56.3% | estimated ± 6.9 pp, medium confidence |
| Qwen3-Omni-30B-A3B-Instruct | 31.2% | 55.4% | estimated ± 6.9 pp, low confidence |
| Qwen3-Omni-30B-A3B-Thinking | 43.4% | 56.2% | estimated ± 6.9 pp, medium confidence |
| Sarvam 105B | 34.4% | 55.5% | estimated ± 6.9 pp, low confidence |
| Sarvam 30B | 26.5% | 55.3% | estimated ± 6.9 pp, low confidence |
| Solar Pro 2 | 33.7% | 55.5% | estimated ± 6.9 pp, low confidence |
| Step 3.7 Flash | 67.3% | 66.3% | estimated ± 6.9 pp, medium confidence |
| Trinity-Large-Preview | 56.3% | 59.5% | estimated ± 6.9 pp, medium confidence |
| Trinity-Large-Thinking | 56.3% | 59.5% | estimated ± 6.9 pp, medium confidence |
| Ultravox v0.6 Llama 3.3 70B | 47.1% | 56.8% | estimated ± 6.9 pp, medium confidence |