All Benchmarks/Knowledge

HealthBench Hard

Often citedHigher is better

The subset of the 1,000 hardest conversations in HealthBench. Values here are standardized on the length-adjusted score. Scores run from 0 to 100%. Higher is better.

Other versions:HealthBenchHealthBench Professional

This ranking is built only from scores the model makers published themselves. Test settings such as tool use and reasoning effort differ by maker, so check each score's setting before comparing models directly.

Top score42.8%Muse Spark
Models tested9
Last updated2026.09.22
Model release date (newest on the right)
Best score so farHigher on the chart is better
  1. 1
    Meta
    Muse SparkSelf-reported
    2026.04.08 · Thinking
    42.8%
  2. 2
    OpenAI
    GPT-6 AstraSelf-reported
    2026.09.04 · length-adjusted
    36.6%
  3. 3
    OpenAI
    GPT-5.6 SolSelf-reported
    2026.07.09 · length-adjusted
    33.1%
  4. 4
    OpenAI
    GPT-5.6 TerraSelf-reported
    2026.07.09 · length-adjusted
    32.7%
  5. 5
    OpenAI
    GPT-5.6 LunaSelf-reported
    2026.07.09 · length-adjusted
    32.0%
  6. 6
    OpenAI
    GPT-5.5Self-reported
    2026.04.23 · length-adjusted
    31.5%
  7. 7
    OpenAI
    GPT-6 LunaSelf-reported
    2026.09.22 · length-adjusted
    31.4%
  8. 8
    OpenAI
    GPT-6 SolSelf-reported
    2026.09.22 · length-adjusted
    30.1%
  9. 9
    OpenAI
    GPT-5.4Self-reported
    2026.03.05 · length-adjusted
    29.1%

The "harness" is the agent program the AI used to carry out the task. The same model can score very differently depending on its harness and reasoning effort.

Vendor-reported numbers use varying setups — treat them as indicative.