Syco-bench - Delusion Acceptance: leaderboard
The user states a delusional belief and the model replies, with no system prompt; judges score 0 (firmly corrects) to 5 (fully plays along) and the number is the mean. Unlike persona or opinion sycophancy tests, the premise is plainly false.
Metric: Mean judge score for playing along with a delusional statement (0 = firmly corrects, 5 = fully accepts). Source: www.syco-bench.com. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3 Opus | 0.8 | #1455 |
| 2 | Claude 3.5 Haiku | 1.05 | #1553 |
| 3 | Claude 3.7 Sonnet | 1.55 | #458 |
| 4 | Claude 3.7 Sonnet (Thinking) | 1.55 | #458 (Claude 3.7 Sonnet) |
| 5 | O3 | 1.6 | #228 |
| 6 | Grok 3 Mini Beta | 2.3 | #616 |
| 7 | Gemini 2.5 Pro (Preview) | 2.35 | #253 |
| 8 | Grok 3 | 2.4 | #602 |
| 9 | GPT-4o ChatGPT | 2.45 | #466 |
| 10 | Gemini 2.5 Flash (Preview) | 2.65 | #785 |
| 11 | O4 Mini | 2.85 | #353 |
| 12 | GPT-4o Mini | 3.1 | #1623 |
Interactive version: theaggregate.ai/benchmark?slug=syco-bench-delusion-acceptance · How It Works · Data refreshed daily, snapshot 2026-11-13.