ToolPrivacyBench - Synthetic-private - Task Success: leaderboard
Successful completion of the intended workflow. Executed multi-tool workflows with purpose-bound privacy policies. Paper v1 Tables 7-8 separately report 1,000 public-derived cases and 1,150 synthetic-private cases. Task success is reported alongside leakage; these are not a new composite score. This series reports TaskSuccess (%). Source: https://arxiv.org/html/2606.28061v1.
Metric: TaskSuccess (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K2.5 | 97.7 | #152 |
| 2 | Qwen 3.6 Plus | 97.7 | #124 |
| 3 | MiniMax-M2.7 | 97.42 | #270 |
| 4 | DeepSeek V4 Flash | 97.2 | #155 |
| 5 | GPT-5.5 | 96.8 | #33 |
| 6 | Claude Opus 4.7 | 95.9 | #51 |
| 7 | Gemini 3.5 Flash | 92.45 | #66 |
| 8 | GLM-5.1 | 92.27 | #109 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=toolprivacybench-synthetic-private-task-success · How It Works · Data refreshed daily, snapshot 2026-11-05.