Calibration
AA-HLE → HealthBench Professional
HealthBench Professional is estimated from AA-HLE with a offset logistic curve fitted on 14 models measured on both: y = 0.5766 + (0.6490 − 0.5766) / (1 + exp(−72.63·(x − 0.4956))), R² = 0.40, cross-validated error 4.8 pp. It is used for 90 estimates.
| Estimated model | AA-HLE | HealthBench Professional | Source |
|---|---|---|---|
| Apodex 1.1 | 34.1% | 57.7% | estimated ± 4.8 pp, low confidence |
| Apodex 1.1 Mini | 34.1% | 57.7% | estimated ± 4.8 pp, low confidence |
| Celeris-1 | 6.8% | 57.7% | estimated ± 4.8 pp, low confidence |
| Claude 3 Haiku | 4.1% | 57.7% | estimated ± 4.8 pp, low confidence |
| Claude 3 Opus | 2.8% | 57.7% | estimated ± 4.8 pp, low confidence |
| Claude 4.1 Opus Thinking | 12.5% | 57.7% | estimated ± 4.8 pp, low confidence |
| Claude 4 Sonnet | 4.3% | 57.7% | estimated ± 4.8 pp, low confidence |
| Command A+ | 12.0% | 57.7% | estimated ± 4.8 pp, low confidence |
| DeepSeek-R1 | 15.8% | 57.7% | estimated ± 4.8 pp, low confidence |
| DeepSeek R1 Distill Qwen 32B | 4.6% | 57.7% | estimated ± 4.8 pp, low confidence |
| DeepSeek V3 0324 | 4.7% | 57.7% | estimated ± 4.8 pp, low confidence |
| DeepSeek V3.1 | 6.7% | 57.7% | estimated ± 4.8 pp, low confidence |
| DeepSeek V3.1 (Reasoning) | 14.3% | 57.7% | estimated ± 4.8 pp, low confidence |
| DeepSeek V3.2 | 11.2% | 57.7% | estimated ± 4.8 pp, low confidence |
| Exaone 4.0 1.2B | 5.7% | 57.7% | estimated ± 4.8 pp, low confidence |
| Exaone 4.0 32B | 5.0% | 57.7% | estimated ± 4.8 pp, low confidence |
| Gemini 1.0 Pro | 4.2% | 57.7% | estimated ± 4.8 pp, low confidence |
| Gemini 1.5 Pro | 4.6% | 57.7% | estimated ± 4.8 pp, low confidence |
| Gemini 2.5 Flash | 4.7% | 57.7% | estimated ± 4.8 pp, low confidence |
| Gemini 4 Argon | 57.1% | 64.9% | estimated ± 4.8 pp, medium confidence |
| Gemma 3 27B | 4.4% | 57.7% | estimated ± 4.8 pp, low confidence |
| GLM-4.5-Air | 7.0% | 57.7% | estimated ± 4.8 pp, low confidence |
| GLM-5-Turbo | 27.8% | 57.7% | estimated ± 4.8 pp, low confidence |
| GLM-5V-Turbo | 17.1% | 57.7% | estimated ± 4.8 pp, low confidence |
| GPT-4 Turbo | 3.1% | 57.7% | estimated ± 4.8 pp, low confidence |
| GPT-4o | 2.4% | 57.7% | estimated ± 4.8 pp, low confidence |
| GPT-4o mini | 4.2% | 57.7% | estimated ± 4.8 pp, low confidence |
| GPT-5.1 | 28.5% | 57.7% | estimated ± 4.8 pp, low confidence |
| GPT-5.1-Codex | 25.7% | 57.7% | estimated ± 4.8 pp, low confidence |
| GPT-5.1-Codex-Max | 25.7% | 57.7% | estimated ± 4.8 pp, low confidence |
| GPT-5.2-Codex | 35.7% | 57.7% | estimated ± 4.8 pp, low confidence |
| GPT-5.3 Codex | 42.5% | 57.7% | estimated ± 4.8 pp, medium confidence |
| GPT-5 (high) | 28.5% | 57.7% | estimated ± 4.8 pp, low confidence |
| GPT-5 (medium) | 25.4% | 57.7% | estimated ± 4.8 pp, low confidence |
| GPT-OSS 120B | 19.6% | 57.7% | estimated ± 4.8 pp, low confidence |
| GPT-OSS 20B | 11.0% | 57.7% | estimated ± 4.8 pp, low confidence |
| Granite-4.0-350M | 5.5% | 57.7% | estimated ± 4.8 pp, low confidence |
| Granite-4.0-H-1B | 5.0% | 57.7% | estimated ± 4.8 pp, low confidence |
| Granite-4.0-H-350M | 6.4% | 57.7% | estimated ± 4.8 pp, low confidence |
| Grok 4 | 26.7% | 57.7% | estimated ± 4.8 pp, low confidence |
| Grok 4.1 Fast | 5.1% | 57.7% | estimated ± 4.8 pp, low confidence |
| Grok 4.1 Fast (Reasoning) | 19.3% | 57.7% | estimated ± 4.8 pp, low confidence |
| Grok 4 Fast (Reasoning) | 19.1% | 57.7% | estimated ± 4.8 pp, low confidence |
| Grok Code Fast 1 | 8.0% | 57.7% | estimated ± 4.8 pp, low confidence |
| Hy3 | 33.5% | 57.7% | estimated ± 4.8 pp, low confidence |
| K-Exaone | 13.9% | 57.7% | estimated ± 4.8 pp, low confidence |
| Kimi K2 | 7.4% | 57.7% | estimated ± 4.8 pp, low confidence |
| Kimi K2.7 Code | 35.0% | 57.7% | estimated ± 4.8 pp, low confidence |
| LFM2.5-2.6B | 6.2% | 57.7% | estimated ± 4.8 pp, low confidence |
| LFM2.5-8B-A1B | 6.9% | 57.7% | estimated ± 4.8 pp, low confidence |
| LFM2.5-VL-1.6B-Extract | 5.1% | 57.7% | estimated ± 4.8 pp, low confidence |
| Ling 3.0 Flash VL | 22.0% | 57.7% | estimated ± 4.8 pp, low confidence |
| Ling 3.0 Tiny | 9.3% | 57.7% | estimated ± 4.8 pp, low confidence |
| Llama 3.1 405B | 4.0% | 57.7% | estimated ± 4.8 pp, low confidence |
| Llama 4 Maverick | 4.9% | 57.7% | estimated ± 4.8 pp, low confidence |
| Llama 4 Scout | 3.8% | 57.7% | estimated ± 4.8 pp, low confidence |
| MiMo-V2.6-Flash | 35.1% | 57.7% | estimated ± 4.8 pp, low confidence |
| MiMo-V2.6-Pro | 49.4% | 61.1% | estimated ± 4.8 pp, medium confidence |
| MiMo-V2-Omni | 22.1% | 57.7% | estimated ± 4.8 pp, low confidence |
| MiMo-V2-Pro | 30.4% | 57.7% | estimated ± 4.8 pp, low confidence |
| Mistral Large 2 | 3.3% | 57.7% | estimated ± 4.8 pp, low confidence |
| Mistral Large 3 | 4.2% | 57.7% | estimated ± 4.8 pp, low confidence |
| Mistral Large 4 | 35.0% | 57.7% | estimated ± 4.8 pp, low confidence |
| Mistral Medium 3 | 4.1% | 57.7% | estimated ± 4.8 pp, low confidence |
| Mistral Small 4 | 9.9% | 57.7% | estimated ± 4.8 pp, low confidence |
| Mistral Small 4 (Reasoning) | 9.9% | 57.7% | estimated ± 4.8 pp, low confidence |
| Muse Glimmer 30B | 22.0% | 57.7% | estimated ± 4.8 pp, low confidence |
| Muse Spark 1.3 | 48.7% | 60.2% | estimated ± 4.8 pp, medium confidence |
| Nemotron 3 Nano 30B | 11.4% | 57.7% | estimated ± 4.8 pp, low confidence |
| Nemotron 3 Super 100B | 20.8% | 57.7% | estimated ± 4.8 pp, low confidence |
| Nemotron Ultra 253B | 7.4% | 57.7% | estimated ± 4.8 pp, low confidence |
| North Mini Code | 11.1% | 57.7% | estimated ± 4.8 pp, low confidence |
| Nova Pro | 3.2% | 57.7% | estimated ± 4.8 pp, low confidence |
| o3 | 20.1% | 57.7% | estimated ± 4.8 pp, low confidence |
| Phi-4 | 3.8% | 57.7% | estimated ± 4.8 pp, low confidence |
| Phi-4 Multimodal Instruct | 5.0% | 57.7% | estimated ± 4.8 pp, low confidence |
| Quasar 438B | 18.7% | 57.7% | estimated ± 4.8 pp, low confidence |
| Qwen2.5 Coder 32B Instruct | 3.5% | 57.7% | estimated ± 4.8 pp, low confidence |
| Qwen3.5 397B (Reasoning) | 19.8% | 57.7% | estimated ± 4.8 pp, low confidence |
| Qwen 3.6 Max (preview) | 30.8% | 57.7% | estimated ± 4.8 pp, low confidence |
| Qwen3.8 Max Preview | 43.1% | 57.7% | estimated ± 4.8 pp, medium confidence |
| Qwen3 Max | 11.9% | 57.7% | estimated ± 4.8 pp, low confidence |
| Qwen3-Omni-30B-A3B-Instruct | 4.6% | 57.7% | estimated ± 4.8 pp, low confidence |
| Qwen3-Omni-30B-A3B-Thinking | 7.5% | 57.7% | estimated ± 4.8 pp, low confidence |
| Sarvam 105B | 11.0% | 57.7% | estimated ± 4.8 pp, low confidence |
| Sarvam 30B | 7.5% | 57.7% | estimated ± 4.8 pp, low confidence |
| Solar Pro 2 | 3.7% | 57.7% | estimated ± 4.8 pp, low confidence |
| Solar Pro 3 | 10.3% | 57.7% | estimated ± 4.8 pp, low confidence |
| Step 3.7 Flash | 21.4% | 57.7% | estimated ± 4.8 pp, low confidence |
| Ultravox v0.6 Llama 3.3 70B | 3.6% | 57.7% | estimated ± 4.8 pp, low confidence |