Artificial Analysis launches HealthBench leaderboards

On October 20, 2026, Artificial Analysis launched two leaderboards, HealthBench and HealthBench Professional, to evaluate how AI models respond in realistic health conversations. Both benchmarks come from OpenAI and grade model replies.

Artificial Analysis also published several findings. On HealthBench Professional, OpenAI models occupy the entire score-versus-cost-per-task Pareto frontier: GPT-6 Luna (max) scores 61.9% at $0.004 per task, 4 points below GPT-6 Astra (max) while costing under 2% of the latter's $0.27. On HealthBench, GPT-6 Luna (max) costs $0.002 per task, and DeepSeek V4.1 Flash also appears in the comparison.

On length adjustment, Artificial Analysis reports that OpenAI's adjustment moves each task's score by 1.47 points for every 500 characters above or below 2,000 on HealthBench Professional, and by 2.99 points on HealthBench. Claude Opus 5.5 (max with fallback) holds the highest raw HealthBench Professional score at 71.4%.

By conversation type, raw scores on HealthBench Professional red-teaming conversations range from 30% for DeepSeek V4.1 Flash (max) to 65% for GPT-6 Astra (max), compared with 63% to 79% on good-faith conversations; Claude Fable 5.1 (max with fallback) scores highest on good-faith conversations at 79%. On HealthBench, GPT-6 Sol (max) writes the shortest answers at roughly 1,000 characters and has the highest raw Communication Quality score at 81%, but the lowest Completeness score at 34%. Context Seeking or Global Health is the lowest-scoring theme for 13 of the 14 models.