DistressBench
Safety
Clinician-authored mental-health response evaluation over 718 tasks: 299 single-turn and 419 multi-turn, split evenly between public and private sets. Scores average weighted rubric compliance over tasks and three runs, grading only the final reply. Majority votes of Claude Opus 4.8, Llama 4 Maverick and Gemini 3.1 Pro judge each rubric. Published uncertainty is a paired bootstrap 95% interval, not an independent-item standard error.
Top models (higher is better)
| Model | Score |
|---|---|
| Muse Spark 1.2 | 88.4 |
| Muse Spark 1.3 | 84.9 |
| Muse Glimmer 30B | 80.6 |
| DeepSeek V4.1 Flash | 79.8 |
| Inkling | 78.5 |
| Mistral Large 4 Preview | 77.0 |
| GPT-5.6 Terra | 76.4 |
| Claude Opus 5.5 | 75.6 |
| Kimi K2.6 | 75.2 |
| GPT-6 Astra | 74.3 |
| GPT-5.6 Sol | 73.6 |
| Kimi K3 | 73.5 |
| Qwen3.7-Plus (source unspecified) | 72.9 |
| GPT-5.6 Luna | 72.7 |
| Claude Fable 5.1 | 71.5 |
| Opus 4.8 | 71.4 |
| GLM-5 | 71.2 |
| Gemini 3.8 Flash | 70.7 |
| DeepSeek-V4-Flash | 69.7 |
| Claude Fable 5 | 69.4 |
| DeepSeek-V4-Pro | 69.2 |
| Gemini 3.7 Flash | 69.0 |
| Gemini 3.1 Pro Preview | 68.8 |
| Sonnet 5 | 68.0 |
| GPT-6 Sol | 68.0 |
| Gemini 3.5 Flash | 66.9 |
| GPT-6 Luna | 66.7 |
| gpt-oss-120b | 63.7 |
| Mistral Large 3 675B Instruct 2512 | 58.8 |
| Grok 4.7 | 52.4 |
| Llama 4 Maverick Instruct | 41.3 |