AgentAbstain: leaderboard
Tool-using agents get 263 task pairs: one version they should carry out and a minimally changed version where they should stop, ask or decline (missing parameter, ambiguous request, irreversible high-stakes action, failed tool, risk found mid-run); the score is the share of pairs where both are handled right. Unlike over-refusal tests, acting is also required, so blanket refusal scores zero.
Metric: Paired accuracy (0-1, higher is better). Source: arxiv.org. Saturation forecast: Around October 2027. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 0.6 | #94 |
| 2 | Claude Opus 4.7 | 0.59 | #79 |
| 3 | Claude Sonnet 4.6 | 0.53 | #161 |
| 4 | GPT-5.5 | 0.53 | #49 |
| 5 | Claude Haiku 4.5 | 0.5 | #568 |
| 6 | GPT-5 | 0.5 | #167 |
| 7 | GPT-5.4 | 0.49 | #123 |
| 8 | GLM-5 | 0.48 | #260 |
| 9 | GPT-OSS-120B | 0.46 | #700 |
| 10 | GPT-5.2 | 0.43 | #176 |
| 11 | MiniMax-M2.5 | 0.42 | #582 |
| 12 | DeepSeek V3.2 | 0.41 | #406 |
| 13 | GPT-5.1 | 0.41 | #229 |
| 14 | Gemini 3 Flash | 0.4 | #164 |
| 15 | DeepSeek V4 Pro | 0.37 | #188 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=agentabstain · How It Works · Data refreshed daily, snapshot 2026-11-13.