Scale Labs DistressBench (DistressBench)
We show this table for reference; we do not rank on it.
Measures crisis-support replies against clinician-written rubrics. These published results are excluded from overall and category rankings.
Weighted rubric (%) on DistressBench — October 15, 2026 snapshot
We mirror the published weighted rubric (%) view for DistressBench. muse-spark-1.2 leads the public snapshot at 88.4%, followed by muse-spark-1.3 (84.9%) and muse-glimmer-30b (80.6%). We do not use these results to rank models overall.
33 modelsCross-domainCurrentDisplay onlyUpdated October 15, 2026 snapshot
Weighted rubric (%) table (33 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | muse-spark-1.2Meta | 88.4% | Not reported | Unknown |
| 2 | muse-spark-1.3Meta | 84.9% | Not reported | Unknown |
| 3 | muse-glimmer-30bMeta | 80.6% | Not reported | Unknown |
| 4 | deepseek-v4.1-flashDeepseek | 79.8% | Not reported | Unknown |
| 5 | inklingThinking-Machines | 78.5% | Not reported | Unknown |
| 6 | qwen3.8-maxAlibaba | 77.6% | Not reported | Unknown |
| 7 | mistral-large-4Mistral | 77.0% | Not reported | Unknown |
| 8 | gpt-5.6-terraOpenAI | 76.4% | Not reported | Unknown |
| 9 | claude-opus-5.5Anthropic | 75.6% | Not reported | Unknown |
| 10 | kimi-k2.6Moonshot AI | 75.2% | Not reported | Unknown |
| 11 | gpt-6-astraOpenAI | 74.3% | Not reported | Unknown |
| 12 | gpt-5.6-solOpenAI | 73.6% | Not reported | Unknown |
| 13 | kimi-k3Moonshot AI | 73.5% | Not reported | Unknown |
| 14 | qwen3.7-plusAlibaba | 72.9% | Not reported | Unknown |
| 15 | gpt-5.6-lunaOpenAI | 72.7% | Not reported | Unknown |
| 16 | claude-fable-5.1Anthropic | 71.5% | Not reported | Unknown |
| 17 | claude-opus-4.8Anthropic | 71.4% | Not reported | Unknown |
| 18 | glm-5Zhipu | 71.2% | Not reported | Unknown |
| 19 | gemini-3.8-flashGoogle | 70.7% | Not reported | Unknown |
| 20 | deepseek-v4-flashDeepseek | 69.7% | Not reported | Unknown |
| 21 | claude-fable-5Anthropic | 69.4% | Not reported | Unknown |
| 22 | deepseek-v4-proDeepseek | 69.2% | Not reported | Unknown |
| 23 | gemini-3.7-flashGoogle | 69.0% | Not reported | Unknown |
| 24 | gemini-3.1-proGoogle | 68.8% | Not reported | Unknown |
| 25 | claude-sonnet-5Anthropic | 68.0% | Not reported | Unknown |
| 26 | gpt-6-solOpenAI | 68.0% | Not reported | Unknown |
| 27 | gemini-3.5-flashGoogle | 66.9% | Not reported | Unknown |
| 28 | gpt-6-lunaOpenAI | 66.7% | Not reported | Unknown |
| 29 | gpt-oss-120bOpenAI | 63.7% | Not reported | Unknown |
| 30 | mistral-largeMistral | 59.1% | Not reported | Unknown |
| 31 | mistral-large-3Mistral | 58.8% | Not reported | Unknown |
| 32 | grok-4.7xAI | 52.4% | Not reported | Unknown |
| 33 | llama-4-maverickMeta | 41.3% | Not reported | Unknown |
How to read this leaderboard
Compare the published evaluation configurations. A higher score means more rubric weight satisfied; it does not measure a clinical outcome.
Operator receipt: 33 sourced rows are currently displayable on this page; the leading published row is muse-spark-1.2 at 88.4%.
Honest limit: 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.
How to read DistressBench results
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.
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.
Snapshot
The published DistressBench snapshot places muse-spark-1.2 first at 88.4%. The third row is 7.8 points behind. The broader top-10 range is 13.2 points, so the table still separates the published systems.
33 models have been evaluated on DistressBench. The benchmark falls in the Cross-domain category. We keep external benchmark mirrors separate from the weighted global scoring system, so these results remain source-specific evidence. DistressBench is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About DistressBench
Year
2026
Tasks
718 conversations (359 public, 359 private)
Format
Weighted rubric score (%)
Difficulty
Crisis-support response evaluation
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.
Freshness and provenance
Version
DistressBench 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
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 88.4% weighted rubric (%). BenchLM shows this benchmark for display only and does not use it in overall rankings.
How many models are evaluated on DistressBench?
The October 15, 2026 snapshot snapshot contains 33 AI models.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 5,600+ readers.
One email each week. Unsubscribe anytime.