HealthBench raw score (HealthBench (raw))
We show this table for reference; we do not rank on it.
Raw score on realistic multi-turn healthcare conversations graded against expert-written rubrics.
Benchmark score on HealthBench (raw) — November 13, 2026
We compile the HealthBench (raw) rows from provider self-reports. Claude Sonnet 5.5 leads the table at 69.4%, followed by Claude Opus 5.5 (68.1%) and Claude Opus 5 (67.1%). We do not use these results to rank models overall.
Claude Sonnet 5.5
Anthropic
Reported by AnthropicClaude Opus 5.5
Anthropic
Reported by AnthropicClaude Opus 5
Anthropic
Reported by Anthropic8 modelsKnowledgeCurrentDisplay onlyUpdated November 13, 2026
Benchmark score table (8 models)
All 8 results on this page were reported by the labs that built the models.
| Rank | Model / configuration | Score | Who ran it | Parameters (B) | Open / closed |
|---|---|---|---|---|---|
| 1 | Claude Sonnet 5.5AnthropicReported by Anthropic | 69.4% | Reported by Anthropic | Not reported | Closed |
| 2 | Claude Opus 5.5AnthropicReported by Anthropic | 68.1% | Reported by Anthropic | Not reported | Closed |
| 3 | Claude Opus 5AnthropicReported by Anthropic | 67.1% | Reported by Anthropic | Not reported | Closed |
| 4 | GPT-6 AstraOpenAIReported by OpenAI | 56.9% | Reported by OpenAI | Not reported | Closed |
| 5 | GPT-6.1 SolOpenAIReported by OpenAI | 56.7% | Reported by OpenAI | Not reported | Closed |
| 6 | GPT-6.1 TerraOpenAIReported by OpenAI | 51.1% | Reported by OpenAI | Not reported | Closed |
| 7 | GPT-6 LunaOpenAIReported by OpenAI | 50.0% | Reported by OpenAI | Not reported | Closed |
| 8 | GPT-6 SolOpenAIReported by OpenAI | 47.1% | Reported by OpenAI | Not reported | Closed |
- Lab-reported
- Benchmark-run
- Independent
HealthBench (raw, Artificial Analysis)
Artificial Analysis evaluates 5,000 conversations with GPT-4.1 as judge and 1 run per task. Raw rubric score is reported in percentage points. Length adjustment uses 2.99 points per 500 answer characters above or below 2,000; reasoning is excluded. Effort and Default Fallback labels identify the evaluated configuration. These rows remain separate from provider runs and outside overall and category rankings.
| Source rank | Source model or configuration | Raw rubric score | Source setup | Parameters (B) | Open / closed |
|---|---|---|---|---|---|
| 1 | Claude Sonnet 5.5 (Max, Default Fallback)Anthropic | 65.9% | Max, Default FallbackOriginal result | Not reported | Closed |
| 2 | Claude Opus 5.5 (Max, Default Fallback)Anthropic | 64.3% | Max, Default FallbackOriginal result | Not reported | Closed |
| 3 | Claude Fable 5.1 (Max, Default Fallback)Anthropic | 63.9% | Max, Default FallbackOriginal result | Not reported | Closed |
| 4 | DeepSeek V4.1 Flash (Max)DeepSeek | 61.9% | MaxOriginal result | Not reported | Open |
| 5 | Claude Haiku 5.5 (Max)Anthropic | 61.2% | MaxOriginal result | Not reported | Closed |
| 6 | Gemini 4 Argon (High)Google | 60.0% | HighOriginal result | Not reported | Closed |
| 7 | Claude Fable 5.2 (Max, Default Fallback)Anthropic | 58.8% | Max, Default FallbackOriginal result | Not reported | Closed |
| 8 | Gemini 3.8 Flash (High)Google | 57.1% | HighOriginal result | Not reported | Closed |
| 9 | GPT-6.1 Sol (Max)OpenAI | 56.9% | MaxOriginal result | Not reported | Closed |
| 10 | GPT-6 Astra (Max)OpenAI | 56.9% | MaxOriginal result | Not reported | Closed |
| 11 | GLM-5.3 (Max)Z.AI | 54.7% | MaxOriginal result | Not reported | Open |
| 12 | Kimi K3 (Max)Moonshot AI | 54.2% | MaxOriginal result | Not reported | Open |
| 13 | Grok 4.7 (Xhigh)xAI | 54.2% | XhighOriginal result | Not reported | Closed |
| 14 | GPT-6 Luna (Max)OpenAI | 50.5% | MaxOriginal result | Not reported | Closed |
| 15 | GPT-6 Sol (Max)OpenAI | 48.2% | MaxOriginal result | Not reported | Closed |
Among the reported HealthBench (raw) rows, Claude Sonnet 5.5 is first at 69.4%. The third row is 2.3 points behind. The broader top-10 range is 22.3 points, so the table still separates the published systems.
8 models have been evaluated on HealthBench (raw). The benchmark falls in the Knowledge category. HealthBench (raw) is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About HealthBench (raw)
Year
2025
Tasks
5,000 multi-turn patient conversations
Format
Raw rubric score
Difficulty
Realistic healthcare conversations
OpenAI created HealthBench with physician-written rubrics for 5,000 health conversations. This table retains source-backed raw scores from provider reports. The separate Artificial Analysis table below shows its independently run raw scores with the original effort and fallback labels.
Freshness and provenance
Version
HealthBench (raw) 2025
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does HealthBench (raw) measure?
Raw score on realistic multi-turn healthcare conversations graded against expert-written rubrics.
Which model scores highest on HealthBench (raw)?
Claude Sonnet 5.5 by Anthropic currently leads with a score of 69.4% on HealthBench (raw).
How many models are evaluated on HealthBench (raw)?
8 AI models have published results on HealthBench (raw) in the BenchLM catalog.
Compare top models on HealthBench (raw)
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 7,000+ subscribers.
One email each week. Unsubscribe anytime.