WellnessBench - Disordered Eating: leaderboard
WellnessBench - Disordered Eating: Primary pass rate (%). WellnessBench native-API evaluation: 60 clinician-authored primary/control task pairs (25 single-turn, 35 dialogue), provider-default reasoning, no system prompt, 4,096 output tokens per turn, three planned repetitions. Gemini 3.7 Flash simulates the user, gates response windows and grades equal-weight rubric rows. Provider blocks and judge failures are excluded; judge audit found optimistic single-turn passes. The Muse models' October 15 run is partial (about two repetitions, three ungraded tasks each). The live overall scoreboard includes later top-ups; older report breakdowns retain their original ten- or twelve-model populations. Consumer-app comparisons are separate and not admitted here.
Metric: Primary pass rate (%). Source: www.micro1.ai. Saturation forecast: Rough model projection: around 2026. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (WellnessBench Run: provider-default; native-API; no-system-prompt) | 84 | |
| 2 | GPT-6.1 Sol (WellnessBench Run: provider-default; native-API; no-system-prompt) | 82 | |
| 3 | GPT-6 Sol (WellnessBench Run: provider-default; native-API; no-system-prompt) | 73 | |
| 4 | Gemini 3.8 Flash (WellnessBench Run: provider-default; native-API; no-system-prompt) | 65 | |
| 5 | GPT-5.6 Sol (WellnessBench Run: provider-default; native-API; no-system-prompt) | 61 | |
| 6 | Claude Sonnet 5.5 (WellnessBench Run: provider-default; native-API; no-system-prompt) | 61 | |
| 7 | Grok 4.6 (WellnessBench Run: provider-default; native-API; no-system-prompt) | 57 | |
| 8 | Claude Fable 5.1 (WellnessBench Run: provider-default; native-API; no-system-prompt) | 56 | |
| 9 | Grok 4.7 (WellnessBench Run: provider-default; native-API; no-system-prompt) | 55 | |
| 10 | Claude Opus 5.5 (WellnessBench Run: provider-default; native-API; no-system-prompt) | 53 |
Interactive version: theaggregate.ai/benchmark?slug=wellnessbench-disordered-eating · How It Works · Data refreshed daily, snapshot 2026-11-06.