HealthBench Professional (Length-Adjusted)
Professional Work · 2026-04-30
HealthBench Professional evaluates real clinical work using 525 physician-authored conversations spanning care consultations, writing and documentation, and medical research. This metric applies the benchmark's published length adjustment to penalize unnecessarily verbose responses.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5 | 66.0 |
| GPT-6 Astra | 64.7 |
| Muse Wonder Full | 62.3 |
| GPT-5.6 Sol Pro | 61.4 |
| Muse Wonder | 60.5 |
| GPT-5.6 Terra | 57.7 |
| Opus 4.8 | 57.4 |
| Muse Spark Open | 57.3 |
| GPT-5.6 Luna | 55.9 |
| Claude Opus 5 | 55.5 |
| GPT-6 Sol (ChatGPT, October 2026) | 55.2 |
| GPT-5.6 Sol (ChatGPT, August 2026) | 54.0 |
| Ember-1 | 52.3 |
| DeepSeek V4.1 Flash 1017 | 51.7 |
| Grok 4.6 | 51.7 |
| GPT-5.5 | 51.2 |
| Gemini 3.8 Flash | 49.8 |
| Opus 4.7 | 49.7 |
| Kimi K3 | 48.9 |
| GLM-5.3 | 48.7 |
| Gemini 3.7 Flash | 48.4 |
| GPT-6 Luna (ChatGPT, October 2026) | 48.2 |
| GPT-5.4 | 48.1 |
| DeepSeek V4.1 Flash | 47.8 |
| Gemini 3.6 Flash | 46.4 |
| GPT-5 | 46.2 |
| GPT-5.2 | 45.9 |
| GPT-5.6 Luna (ChatGPT, August 2026) | 44.1 |
| Gemini 3.1 Pro Preview | 43.9 |
| Gemini 3.5 Flash | 43.4 |
| Grok 4.5 | 42.5 |
| Inkling | 41.2 |
| DeepSeek V4 Pro 0813 | 40.9 |
| GPT-5.1 | 39.6 |
| GPT-5.5 Instant | 38.4 |
| GPT-5.1 Instant | 37.6 |
| MiniMax M3 | 37.1 |
| GPT-5.2 Instant (source unspecified) | 35.7 |
| Gemini 3.5 Flash-Lite | 35.0 |
| Nemotron 3 Ultra 550B A55B (NVFP4) | 32.9 |