PersistBench - Memory Sycophancy: leaderboard
Metric: Failure Rate (%). Source: arxiv.org. 18 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5.2 (High) | 59 |
| 2 | Llama 3.3 70B Instruct | 82 |
| 3 | Claude Sonnet 4.5 | 83.5 |
| 4 | Claude Opus 4.5 | 87.5 |
| 5 | GPT-4o | 88 |
| 6 | MiniMax-M2.1 | 94 |
| 7 | Llama 4 Maverick | 96 |
| 8 | GPT-OSS-120B | 96.5 |
| 9 | Kimi K2 (Thinking) | 97 |
| 10 | DeepSeek V3.2 Speciale | 98.5 |
| 11 | GLM-4.7 | 99 |
| 12 | Gemini 3 Flash | 99 |
| 13 | Qwen 3 235B A22B | 99.5 |
| 14 | Kimi K2 0905 | 99.5 |
| 15 | Grok 4.1 Fast | 99.5 |
Interactive version: theaggregate.ai/benchmark?slug=persistbench-memory-sycophancy · How It Works · Data refreshed daily, snapshot 2026-09-19.