HealthBench Hard
We show this table for reference; we do not rank on it.
A harder subset of OpenAI's HealthBench for evaluating open-ended medical and health reasoning with rubric-based grading.
Benchmark score on HealthBench Hard — November 13, 2026
We compile the HealthBench Hard rows from provider self-reports and secondary reports. Muse Spark leads the table at 42.8%, followed by GPT-5.4 (40.1%) and GPT-6 Astra (36.6%). We do not use these results to rank models overall.
Muse Spark
Meta
Reported by MetaGPT-5.4
OpenAI
Reported by Meta AI (comparison table)GPT-6 Astra
OpenAI
Reported by OpenAI13 modelsKnowledgeCurrentDisplay onlyUpdated November 13, 2026
Benchmark score table (13 models)
None of the 13 results on this page has a documented run by the benchmark or an independent evaluator: 9 reported by the labs that built the models and 4 reported in another publisher’s document.
| Rank | Model / configuration | Score | Who ran it | Parameters (B) | Open / closed |
|---|---|---|---|---|---|
| 1 | Muse SparkMetaReported by Meta | 42.8% | Reported by Meta | Not reported | Closed |
| 2 | GPT-5.4OpenAIReported by Meta AI (comparison table) | 40.1% | Reported by Meta AI (comparison table) | Not reported | Closed |
| 3 | GPT-6 AstraOpenAIReported by OpenAI | 36.6% | Reported by OpenAI | Not reported | Closed |
| 4 | GPT-6.1 SolOpenAIReported by OpenAI | 36.2% | Reported by OpenAI | Not reported | Closed |
| 5 | GPT-5.6 SolOpenAIReported by OpenAI | 33.1% | Reported by OpenAI | Not reported | Closed |
| 6 | GPT-5.6 TerraOpenAIReported by OpenAI | 32.7% | Reported by OpenAI | Not reported | Closed |
| 7 | GPT-5.6 LunaOpenAIReported by OpenAI | 32.0% | Reported by OpenAI | Not reported | Closed |
| 8 | GPT-6.1 TerraOpenAIReported by OpenAI | 31.4% | Reported by OpenAI | Not reported | Closed |
| 9 | GPT-6 LunaOpenAIReported by OpenAI | 31.4% | Reported by OpenAI | Not reported | Closed |
| 10 | GPT-6 SolOpenAIReported by OpenAI | 30.1% | Reported by OpenAI | Not reported | Closed |
| 11 | Gemini 3.1 ProGoogleReported by Meta AI (comparison table) | 20.6% | Reported by Meta AI (comparison table) | Not reported | Closed |
| 12 | Grok 4.20xAIReported by Meta AI (comparison table) | 20.3% | Reported by Meta AI (comparison table) | Not reported | Closed |
| 13 | Claude Opus 4.6AnthropicReported by Meta AI (comparison table) | 14.8% | Reported by Meta AI (comparison table) | Not reported | Closed |
- Lab-reported
- Benchmark-run
- Independent
Among the reported HealthBench Hard rows, Muse Spark is first at 42.8%. The third row is 6.2 points behind. The broader top-10 range is 12.7 points, so the table still separates the published systems.
13 models have been evaluated on HealthBench Hard. The benchmark falls in the Knowledge category. HealthBench Hard is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About HealthBench Hard
Year
2026
Tasks
1,000 health prompts
Format
Open-ended health evaluation
Difficulty
Advanced health reasoning
Meta describes HealthBench Hard as a 1,000-prompt subset of OpenAI's HealthBench, graded with the same simple-evals implementation and a GPT-4.1-based judge. BenchLM treats it as a display-only health benchmark reference.
Freshness and provenance
Version
HealthBench Hard 2026
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 Hard measure?
A harder subset of OpenAI's HealthBench for evaluating open-ended medical and health reasoning with rubric-based grading.
Which model scores highest on HealthBench Hard?
Muse Spark by Meta currently leads with a score of 42.8% on HealthBench Hard.
How many models are evaluated on HealthBench Hard?
13 AI models have published results on HealthBench Hard in the BenchLM catalog.
Compare top models on HealthBench Hard
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.