# Scale Labs DistressBench (DistressBench)

> Measures crisis-support replies against clinician-written rubrics. These published results are excluded from overall and category rankings.

Canonical page: https://benchlm.ai/benchmarks/scale-distressbench

- Category: [external](/external)
- Last updated: October 15, 2026 snapshot

## About DistressBench

- Year: 2026
- Tasks: 718 conversations (359 public, 359 private)
- Format: Weighted rubric score (%)
- Difficulty: Crisis-support response evaluation
- Paper: [DistressBench: Evaluating Multi-Dimensional Crisis Support in Large Language Models](https://labs.scale.com/papers/distressbench)
- Source: [DistressBench leaderboard](https://labs.scale.com/leaderboard/distressbench)
- Source: [DistressBench paper](https://labs.scale.com/papers/distressbench)

The weighted rubric score measures how much of each clinician-written rubric a reply satisfies. Scores average three runs on 718 conversations, including 359 public and 359 private cases. A three-model judge panel grades the replies; the published intervals use task-level bootstrap resampling.

Compare the published evaluation configurations. A higher score means more rubric weight satisfied; it does not measure a clinical outcome.

The conversations are simulated, English-only, and use U.S. crisis resources. Judge choices and reply length affect the result. Scores do not establish clinical suitability. Dataset and transcript permissions remain separate from these published aggregates.

These published aggregates are display only and excluded from overall and category rankings. We do not import conversations, rubrics, reference replies, or the separately reported long-horizon transcripts.

DistressBench is currently displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Leaderboard (33 models)

| Rank | Model | Creator | Score |
|------|-------|---------|-------|
| 1 | [muse-spark-1.2](https://labs.scale.com/leaderboard/distressbench) | Meta | 88.39% ±0.93 points |
| 2 | [muse-spark-1.3](https://labs.scale.com/leaderboard/distressbench) | Meta | 84.88% ±0.93 points |
| 3 | [muse-glimmer-30b](https://labs.scale.com/leaderboard/distressbench) | Meta | 80.57% ±0.96 points |
| 4 | [deepseek-v4.1-flash](https://labs.scale.com/leaderboard/distressbench) | Deepseek | 79.78% ±1.01 points |
| 5 | [inkling](https://labs.scale.com/leaderboard/distressbench) | Thinking-Machines | 78.47% ±1.06 points |
| 6 | [qwen3.8-max](https://labs.scale.com/leaderboard/distressbench) | Alibaba | 77.64% ±1.03 points |
| 7 | [mistral-large-4](https://labs.scale.com/leaderboard/distressbench) | Mistral | 77.02% ±1.01 points |
| 8 | [gpt-5.6-terra](https://labs.scale.com/leaderboard/distressbench) | OpenAI | 76.44% ±1.08 points |
| 9 | [claude-opus-5.5](https://labs.scale.com/leaderboard/distressbench) | Anthropic | 75.58% ±1.20 points |
| 10 | [kimi-k2.6](https://labs.scale.com/leaderboard/distressbench) | Moonshot AI | 75.18% ±1.09 points |
| 11 | [gpt-6-astra](https://labs.scale.com/leaderboard/distressbench) | OpenAI | 74.31% ±1.10 points |
| 12 | [gpt-5.6-sol](https://labs.scale.com/leaderboard/distressbench) | OpenAI | 73.57% ±1.10 points |
| 13 | [kimi-k3](https://labs.scale.com/leaderboard/distressbench) | Moonshot AI | 73.45% ±1.11 points |
| 14 | [qwen3.7-plus](https://labs.scale.com/leaderboard/distressbench) | Alibaba | 72.87% ±1.19 points |
| 15 | [gpt-5.6-luna](https://labs.scale.com/leaderboard/distressbench) | OpenAI | 72.68% ±1.09 points |
| 16 | [claude-fable-5.1](https://labs.scale.com/leaderboard/distressbench) | Anthropic | 71.55% ±1.32 points |
| 17 | [claude-opus-4.8](https://labs.scale.com/leaderboard/distressbench) | Anthropic | 71.43% ±1.13 points |
| 18 | [glm-5](https://labs.scale.com/leaderboard/distressbench) | Zhipu | 71.17% ±1.16 points |
| 19 | [gemini-3.8-flash](https://labs.scale.com/leaderboard/distressbench) | Google | 70.74% ±1.23 points |
| 20 | [deepseek-v4-flash](https://labs.scale.com/leaderboard/distressbench) | Deepseek | 69.66% ±1.22 points |
| 21 | [claude-fable-5](https://labs.scale.com/leaderboard/distressbench) | Anthropic | 69.43% ±1.75 points |
| 22 | [deepseek-v4-pro](https://labs.scale.com/leaderboard/distressbench) | Deepseek | 69.19% ±1.24 points |
| 23 | [gemini-3.7-flash](https://labs.scale.com/leaderboard/distressbench) | Google | 68.95% ±1.28 points |
| 24 | [gemini-3.1-pro](https://labs.scale.com/leaderboard/distressbench) | Google | 68.76% ±1.30 points |
| 25 | [claude-sonnet-5](https://labs.scale.com/leaderboard/distressbench) | Anthropic | 67.98% ±1.16 points |
| 26 | [gpt-6-sol](https://labs.scale.com/leaderboard/distressbench) | OpenAI | 67.96% ±1.17 points |
| 27 | [gemini-3.5-flash](https://labs.scale.com/leaderboard/distressbench) | Google | 66.89% ±1.28 points |
| 28 | [gpt-6-luna](https://labs.scale.com/leaderboard/distressbench) | OpenAI | 66.66% ±1.14 points |
| 29 | [gpt-oss-120b](https://labs.scale.com/leaderboard/distressbench) | OpenAI | 63.68% ±1.50 points |
| 30 | [mistral-large](https://labs.scale.com/leaderboard/distressbench) | Mistral | 59.08% ±1.42 points |
| 31 | [mistral-large-3](https://labs.scale.com/leaderboard/distressbench) | Mistral | 58.77% ±1.37 points |
| 32 | [grok-4.7](https://labs.scale.com/leaderboard/distressbench) | xAI | 52.39% ±1.32 points |
| 33 | [llama-4-maverick](https://labs.scale.com/leaderboard/distressbench) | Meta | 41.25% ±1.20 points |

## FAQ

### What does DistressBench measure?

Measures crisis-support replies against clinician-written rubrics. These published results are excluded from overall and category rankings.

### Which model leads the published DistressBench snapshot?

muse-spark-1.2 currently leads the published DistressBench snapshot with a score of 88.39%.

### How many models are evaluated on DistressBench?

The October 15, 2026 snapshot contains 33 AI models.

### Does DistressBench affect BenchLM's overall score?

Not directly. DistressBench is still displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.