Artificial Analysis ranks models on HealthBench

Artificial Analysis has launched public leaderboards for HealthBench and HealthBench Professional. Both are OpenAI benchmarks that grade model replies to realistic health conversations against rubrics written by physicians.

HealthBench has 5,000 conversations with everyday users and health workers across seven themes. Its rubric criteria fall under five axes: accuracy, completeness, communication quality, context awareness and instruction following. HealthBench Professional is a smaller set of 525 conversations, all with clinicians. It covers consultation, writing and research tasks and includes adversarial cases.

Artificial Analysis reports both raw scores and OpenAI's length-adjusted score, which lowers scores for replies over 2,000 characters and raises them for shorter ones. Following OpenAI's original setup, GPT-5.4 (low) grades HealthBench Professional and GPT-4.1 grades HealthBench.

On HealthBench Professional, GPT-6 Astra (max) leads at 66.0% length-adjusted. GPT-6.1 Sol follows at 65.1% and Claude Sonnet 5.5 at 63.0%. OpenAI models make up the entire score-versus-cost Pareto frontier. GPT-6 Luna scores 61.9% at $0.004 per task, about 4 points below Astra at under 2% of the cost.

Length changes the rankings sharply. Claude Opus 5.5 and Claude Fable 5.1 average about 5,900 characters per answer and each lose more than 11 points after adjustment. Opus 5.5 has the highest raw score (71.4%) but falls to sixth (59.9%). GPT-6 Sol, which averages about 1,600 characters, is the only model whose score rises.

On HealthBench, Claude Sonnet 5.5 leads at 62.1%, ahead of DeepSeek V4.1 Flash (60.9%) and GPT-6.1 Sol (58.5%). Sonnet 5.5 costs about 16 times as much per task as DeepSeek V4.1 Flash ($0.08 vs $0.005) for a gain of about 1 point. The axis breakdown shows a tradeoff: GPT-6 Sol's short answers score highest on communication quality (81%) but lowest on completeness (34%), while Sonnet 5.5 leads on completeness (63%).