OpenAI MentalHealthBench - Behavior - Harm Avoidance: leaderboard
Behavior axis of OpenAI's MentalHealthBench (launch post and paper, 2026-09-23): not enabling harmful or coercive behavior while staying helpful through proportionate refusal and redirection, scored on the 230 rubric criteria tagged with this axis. The post's axis score is the signed rubric score on this axis, normalized to its possible range (0-100, higher is better); GPT-5.6 Sol grades, models run at API default reasoning effort. The post's 15 models only: the paper's behavior figure uses a different scale.
Metric: Axis score, normalized to its possible range (%). Source: openai.com. Saturation forecast: Around August 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Muse Spark 1.3 | 82.29 | #31 |
| 2 | Grok 4.7 | 79.35 | #26 |
| 3 | GPT-6 Sol | 72.37 | #17 |
| 4 | GPT-6 Luna | 71.61 | #62 |
| 5 | GPT-6 Astra | 70.84 | #5 |
| 6 | Claude Opus 5.5 | 66.22 | #3 |
| 7 | Gemini 3.8 Flash | 61.55 | #30 |
| 8 | GPT-5.6 Luna (August 2026) | 60.91 | |
| 9 | GPT-5.6 Sol (August 2026) | 59.27 | |
| 10 | Claude Fable 5.1 | 59.19 | #10 |
| 11 | Claude Haiku 4.5 | 55.16 | #291 |
| 12 | GPT-4o (Mar 2025) | 50.56 | #339 |
| 13 | Claude Sonnet 5 | 48.58 | #67 |
| 14 | Gemini 3.1 Pro (Preview) | 45.27 | #66 |
| 15 | Gemini 2.5 Pro | 37.07 | #161 |
No result here: #6 GPT-6.1 Sol, #9 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=openai-mentalhealthbench-behavior-harm-avoidance · How It Works · Data refreshed daily, snapshot 2026-11-07.