ImpactBench: leaderboard
The model holds simulated multi-turn conversations with teen and adult users in psychological, physical and social wellbeing scenarios; an LLM judge checks 809 expert-defined metrics, and the number is the pooled share of positive metrics passed (the model promotes a desirable behaviour). Unlike HumaneBench, which it absorbs as one of 26 component benchmarks, it pools many wellbeing suites over 48,540 conversations.
Metric: Desirable-behaviour metrics passed (%, pooled, higher is better). Source: arxiv.org. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Terra | 75.1 | #62 |
| 2 | GPT-5.6 Luna | 74.9 | #121 |
| 3 | Claude Sonnet 5 | 73 | #111 |
| 4 | Claude Opus 5 | 71.5 | #22 |
| 5 | Grok 4.6 | 70.5 | #45 |
| 6 | Gemini 3.7 Flash | 68.5 | #32 |
| 7 | Qwen 3.8 27B | 67.8 | #204 |
| 8 | Gemma 4 31B (IT) | 63.4 | #342 |
| 9 | GLM-5.2 | 62.7 | #138 |
| 10 | DeepSeek V4 Flash | 59.2 | #258 |
Interactive version: theaggregate.ai/benchmark?slug=impactbench · How It Works · Data refreshed daily, snapshot 2026-11-13.