HealthBench Professional Benchmark | Artificial Analysis
This is Artificial Analysis's leaderboard page for HealthBench Professional, OpenAI's set of 525 clinician conversations graded against physician-written rubrics. It ranks models on length-adjusted score, cost, token use, and speed.
- HealthBench Professional covers consult, writing, and research use cases, including red-teaming and difficult conversations, and is graded with GPT-5.4 (low) across 4 repeats per task.
- The length-adjusted score penalizes answers by 1.47 points per 500 characters above 2,000 and rewards the same amount for shorter ones, matching OpenAI's reported method.
- Claude Fable 5.2 (Xhigh, Default Fallback) currently tops the board at a 68.1% length-adjusted score among models with published results.
HealthBench Professional is OpenAI's clinician-focused health benchmark, and this Artificial Analysis page is a leaderboard for how models score on it. The benchmark has 525 conversations with clinicians, split across consult, writing and research use cases, and it includes red-teaming and harder cases. Each conversation is graded against a rubric written by physicians. Artificial Analysis grades responses with GPT-5.4 (low), runs four repeats per task, and reports the mean. The headline number is a length-adjusted score: for every 500 characters above 2,000, 1.47 percentage points come off, and for every 500 below, the same amount is added back before averaging. That's the figure OpenAI reports, and it means a model can't pad its way to a higher score. Reasoning tokens don't count toward length.
The page's main claim is that Claude Fable 5.2 (Xhigh, Default Fallback) currently tops the published results with a length-adjusted score of 68.1%. The rest of the page is a set of charts that let you compare models along different axes: length-adjusted score against cost per task, raw score against answer length, performance broken down by use case, type and difficulty, output tokens used, dollar cost, time per task, and score against release date. That last view is useful if you want to see whether newer models are really improving or just getting more expensive. One detail worth noticing: a task that fails after retries counts as zero everywhere, including the length-adjusted and breakdown scores, and it's excluded from the average answer length. So a model that often crashes or times out gets punished, not quietly skipped.
The page also lists the rest of Artificial Analysis's evaluation suite, which runs well beyond health. It includes agentic benchmarks for spreadsheets, legal work from Harvey's dataset of 120 private tasks, Kubernetes incident analysis, enterprise workflow simulations, terminal and coding tasks, and scientific reasoning. Most of these are described in one line each, with notes on how grading works, such as the legal benchmark's hallucination check against source documents. For a read-it-later save, the main reason to keep this is the HealthBench section and its length-adjustment method, which is a practical way to avoid rewarding verbose answers. The rest is useful as a reference if you're trying to find an evaluation for a specific kind of model work.