Psi-Bench: leaderboard
Psi-Bench (2026) tests persona-sensitive influence: a model talks with a simulated client who role-plays a hidden persona and tries to change the client's view, mindset or behaviour over three rounds. This board is the average effect over the three scenarios on a 1 to 9 scale, rated by a DeepSeek-V3.2 judge that also plays the client; higher is better.
Metric: Average persuasion effect over the three scenarios (1-9; DeepSeek-v3.2 judge, which also plays the persona-profiled client; three rounds after the client's opening message, client profile hidden from the model). Source: arxiv.org. Saturation forecast: Around October 2028. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.1 | 5.79 | #229 |
| 2 | Qwen 3 Next 80B A3B | 5.38 | #757 |
| 3 | DeepSeek V4 Pro | 5.22 | #188 |
| 4 | Gemini 3.1 Pro (Preview) | 5.11 | #94 |
| 5 | GPT-5 Mini | 4.92 | #348 |
| 6 | Gemini 3 Flash | 4.89 | #164 |
| 7 | Qwen 3 32B | 4.54 | #960 |
| 8 | DeepSeek V3.2 | 4.47 | #406 |
| 9 | Grok 4 Fast | 4.19 | #441 |
| 10 | Qwen 3 8B | 4.05 | #1830 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=psi-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.