mheval
Submit results

Quality benchmark

HealthBench-Psych

The mental-health conversations from HealthBench, graded against conversation-specific rubrics written by physicians.

“Responses are evaluated using conversation-specific rubrics created by 262 physicians.”
Arora et al., 2025
Metric
Clipped mean score
Scale
0–1
Source
Paper ↗

Headline score

  1. GPT-5OpenAI · Public API
    0.662
  2. Claude Opus 5.5Anthropic · Public API
    0.635
  3. GPT-6 AstraOpenAI · Public API
    0.566
  4. Qwen3.5-122B-A10BQwen · Open weights
    0.546
  5. Qwen3.5-27BQwen · Open weights
    0.541
  6. GPT-6 LunaOpenAI · Public API
    0.509
  7. Gemini 2.5 ProGoogle · Public API
    0.484
  8. Claude Sonnet 4.5Anthropic · Public API
    0.433
  9. GPT-4oOpenAI · Public API
    0.276
Bars span the metric's full 0–1 scale; higher is better. Hover the info icon for settings a model ran with.

Axis

Model
Claude Opus 5.5Anthropic · Public API0.7370.7010.6370.4500.630
GPT-5OpenAI · Public API0.7130.6380.6590.5490.602
Qwen3.5-122B-A10BQwen · Open weights0.7020.6700.5710.2390.627
Qwen3.5-27BQwen · Open weights0.7010.6710.5520.2620.612
GPT-6 AstraOpenAI · Public API0.6910.7180.5110.4540.617
Gemini 2.5 ProGoogle · Public API0.6750.6230.4890.2340.632
GPT-6 LunaOpenAI · Public API0.6120.6910.4540.3670.657
Claude Sonnet 4.5Anthropic · Public API0.5950.6850.3730.2570.562
GPT-4oOpenAI · Public API0.4660.6020.1710.1290.509

Theme

Model
GPT-5OpenAI · Public API0.7300.7000.6190.7970.5910.5820.665
Claude Opus 5.5Anthropic · Public API0.6900.6560.5500.7640.5470.6510.648
Qwen3.5-122B-A10BQwen · Open weights0.6590.5340.4090.5920.5150.4640.589
Qwen3.5-27BQwen · Open weights0.6280.4700.4320.6130.5130.4230.600
GPT-6 AstraOpenAI · Public API0.6120.5690.5020.7360.4520.5120.609
Gemini 2.5 ProGoogle · Public API0.6010.5450.3570.5760.4070.3770.528
GPT-6 LunaOpenAI · Public API0.5810.4590.4680.6640.3960.4700.523
Claude Sonnet 4.5Anthropic · Public API0.4930.4910.3030.6180.3120.3740.502
GPT-4oOpenAI · Public API0.4070.3270.0950.4050.1880.1650.344