HealthBench (length-adjusted, Artificial Analysis)
Professional Work
Artificial Analysis' length-adjusted HealthBench headline. AA runs the 5,000 conversations published with openai/simple-evals (commit 652c89d) once each, with no system prompt or tools, grades every rubric item with GPT-4.1 at temperature 0.5 as in OpenAI's original implementation, scores failed replies zero, and applies OpenAI's published length adjustment (2.99 points per 500 characters away from 2,000) before averaging. OpenAI's own length-adjusted figures come from its updated implementation, so this original-grader reading is not pooled with them. The same run's unadjusted rubric score is imported to HealthBench, so this second reading is diagnostic.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Sonnet 5.5 | 62.1 |
| DeepSeek V4.1 Flash | 60.9 |
| GPT-6.1 Sol | 58.5 |
| Claude Haiku 5.5 | 58.5 |
| GPT-6 Astra | 58.2 |
| Claude Fable 5.1 | 57.2 |
| Claude Opus 5.5 | 57.0 |
| Grok 4.7 | 56.0 |
| Kimi K3 | 54.7 |
| GPT-6 Luna | 54.6 |
| GLM-5.3 | 54.5 |
| GPT-6 Sol | 54.2 |
| Gemini 4 Argon | 49.1 |
| Gemini 3.8 Flash | 48.4 |