DTap-Bench (Benign Tasks): leaderboard
Utility of tool-using AI agents on DTap-Bench (Chen et al., 2026), a red-teaming platform of simulated MCP environments in 14 domains; each row is a model run inside one agent framework. Benign task success rate in percent over 2,806 tasks with no attack applied, where a rule-based check of the final environment state must find the instruction fully completed, averaged over the 14 domains; higher is better.
Metric: Benign task success rate (%; 2,806 benign agent tasks, no attack applied; a task succeeds when a rule-based check of the final environment state finds the instruction fully completed; mean of the 14 domain rates). Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 87 | #135 |
| 2 | GPT-5.5 | 86.3 | #48 |
| 3 | GPT-5.4 | 85.3 | #121 |
| 4 | DeepSeek V4 Pro | 83.3 | #181 |
| 5 | Claude Sonnet 4.5 | 80.8 | #244 |
| 6 | GPT-5.2 | 80.5 | #172 |
| 7 | GPT-5.1 | 78.7 | #225 |
| 8 | GPT-OSS-120B | 36.2 | #699 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=dtap-bench-benign-tasks · How It Works · Data refreshed daily, snapshot 2026-11-12.