HealthBench Consensus (length-adjusted)
Professional Work
Length-adjusted HealthBench Consensus value from OpenAI's updated implementation. Responses of 2,000 characters receive no adjustment; each additional 500 characters incurs a 0.20-point penalty, while shorter responses receive a credit. Reported as a diagnostic to avoid counting the same evaluation twice; the exact item count is not specified by the source tables.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.6 Sol (ChatGPT, August 2026) | 95.5 |
| GPT-6 Luna (ChatGPT, October 2026) | 95.0 |
| GPT-5.6 Luna (ChatGPT, August 2026) | 94.8 |
| GPT-6 Sol (ChatGPT, October 2026) | 93.8 |