AgentAbstain - Abstain Accuracy: leaderboard
Share of AgentAbstain's abstain variants (underspecified, conflicting, irreversible or newly risky tasks) where the agent stops, asks or declines instead of acting. This ignores the act variants that the headline also requires.
Metric: Abstain accuracy (0-1, higher is better). Source: arxiv.org. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 0.79 | #79 |
| 2 | GPT-5 | 0.7 | #167 |
| 3 | GPT-5.4 | 0.68 | #123 |
| 4 | Claude Sonnet 4.6 | 0.66 | #161 |
| 5 | Claude Haiku 4.5 | 0.66 | #568 |
| 6 | Gemini 3.1 Pro (Preview) | 0.65 | #94 |
| 7 | GPT-5.2 | 0.63 | #176 |
| 8 | GLM-5 | 0.62 | #260 |
| 9 | GPT-5.5 | 0.61 | #49 |
| 10 | GPT-5.1 | 0.61 | #229 |
| 11 | GPT-OSS-120B | 0.6 | #700 |
| 12 | DeepSeek V3.2 | 0.52 | #406 |
| 13 | Kimi K2.5 | 0.52 | #256 |
| 14 | MiniMax-M2.5 | 0.5 | #582 |
| 15 | GPT-4o | 0.44 | #741 |
Interactive version: theaggregate.ai/benchmark?slug=agentabstain-abstain-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-13.