Claude Opus 5.5 Has the Best Raw Score on HealthBench Pro but Drops to Sixth After Length Penalties
Artificial Analysis has added HealthBench and HealthBench Professional to its model leaderboard suite, running a field of frontier models through both benchmarks and publishing results on October 19, 2026.
The benchmarks originate from OpenAI. HealthBench covers 5,000 health conversations with everyday users and health workers across seven themes: emergency referrals, medication questions, disease information, mental health, nutrition, global health, and end-of-life care. HealthBench Professional uses a smaller set of 525 conversations targeting clinicians, covering consultation, writing, and research tasks, including adversarial scenarios.
Both benchmarks report a raw score and a length-adjusted score. The adjustment penalises replies longer than 2,000 characters and rewards shorter ones. The intent is to discourage models from padding answers to maximise rubric coverage at the cost of practical usability. Grading is handled by GPT-5.4 (low) for HealthBench Professional and GPT-4.1 for general HealthBench.
HealthBench Professional: Length Penalty Reshuffles the Rankings
Claude Opus 5.5 has the highest raw score on HealthBench Professional at 71.4%. That is more than 4 percentage points above the next model. But Claude Opus 5.5 averages roughly 5,900 characters per reply — nearly three times the 2,000-character adjustment threshold — and loses over 11 points to the penalty, landing at 59.9% and falling from first to sixth on the adjusted leaderboard.
The adjusted standings:
| Model | Raw | Adjusted | Cost/Task |
|---|---|---|---|
| GPT-6 Astra (max) | — | 66.0% | $0.27 |
| GPT-6.1 Sol (max) | — | 65.1% | $0.047 |
| Claude Sonnet 5.5 (max with fallback) | — | 63.0% | — |
| GPT-6 Sol (max) | — | 62.3% | $0.036 |
| GPT-6 Luna (max) | — | 61.9% | $0.004 |
| Claude Opus 5.5 (max with fallback) | 71.4% | 59.9% | — |
GPT-6 Astra leads at 66.0% and GPT-6.1 Sol is within 0.9 points at 65.1% for roughly one-sixth of Astra’s per-task cost. Claude Fable 5.1 also incurs large length penalties by averaging around 5,900 characters, placing it below 60%.
GPT-6 Sol is the only model in the field averaging under 2,000 characters per reply — roughly 1,600 — and it is the only model whose adjusted score rises above its raw score.
OpenAI models occupy every position on the score-versus-cost Pareto frontier: GPT-6 Luna at $0.004 per task, GPT-6 Sol at $0.036, GPT-6.1 Sol at $0.047, and GPT-6 Astra at $0.27.
General HealthBench: Sonnet Over Astra, DeepSeek Within Reach
On general HealthBench — the 5,000-conversation consumer and health worker set — Claude Sonnet 5.5 leads the adjusted leaderboard at 62.1%. DeepSeek V4.1 Flash (max) follows at 60.9% and GPT-6.1 Sol at 58.5%.
The cost split is stark. DeepSeek V4.1 Flash reaches 60.9% at $0.005 per task. Claude Sonnet 5.5 reaches 62.1% at $0.08 per task — roughly 16 times the cost for a 1.2-point gain. GPT-6 Luna reaches 54.6% at $0.002 per task and forms the lowest-cost anchor on the general HealthBench Pareto frontier alongside DeepSeek V4.1 Flash.
What the Length Penalty Reveals
The rubric axes on HealthBench split into completeness (does the answer include all important information) and communication quality (is it clear, well-structured, and appropriately concise). GPT-6 Sol, writing the shortest replies at around 1,000 characters, scores the highest raw communication quality mark at 81% but the lowest completeness score at 34%. Claude Sonnet 5.5 leads on completeness at 63%.
The length-adjustment mechanism is consequential. A model that maximises completeness by writing exhaustively will hit the penalty floor. Claude Opus 5.5’s 11-point drop is the largest in the field, and it pulls what would be a dominant result into the bottom half of the adjusted leaderboard. The benchmark design explicitly favours concise, targeted replies — a different optimisation target from the extended reasoning outputs that Opus 5.5 is built around.