AgentNoiseBench (tau2-bench, User Noise): leaderboard
Metric: Success on a stratified 25 percent sample of tau2-bench tasks (airline and retail) with user noise injected into the simulated user's turns (ambiguous, inconsistent, redundant, topic-drifting or boundary-probing instructions); noise produced by a fixed generator whose prompt was optimized to degrade a reference agent while keeping every task solvable; task success rate (fraction times 100), mean of four trials; the paper's stability-gated criterion counts a task solved only with a correct final outcome and no noise-induced deviation in the trajectory; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 Max (Thinking) | 78 | #201 (Qwen 3 Max) |
| 2 | GPT-5.2 (Thinking) | 72 | #105 (GPT-5.2) |
| 3 | O3 | 68 | #121 |
| 4 | GLM-4.6 (Thinking) | 68 | #246 (GLM-4.6) |
| 5 | Claude Sonnet 4.5 (Thinking) | 65 | #138 (Claude Sonnet 4.5) |
| 6 | Claude Sonnet 4.5 | 64 | #138 |
| 7 | GLM-4.6 (Non-reasoning) | 64 | #246 (GLM-4.6) |
| 8 | GLM-4.5 (Thinking) | 63 | #265 (GLM-4.5) |
| 9 | LongCat-Flash-Chat | 62 | #322 |
| 10 | Doubao-Seed-1.6 (Thinking) | 61 | |
| 11 | Claude Sonnet 4 (Thinking) | 58 | #194 (Claude Sonnet 4) |
| 12 | Gemini 2.5 Pro | 55 | #145 |
| 13 | Claude Sonnet 4 | 55 | #194 |
| 14 | GLM-4.5 (Non-reasoning) | 55 | #265 (GLM-4.5) |
| 15 | Qwen 3 Max | 51 | #201 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=agentnoisebench-tau2-bench-user-noise · How It Works · Data refreshed daily, snapshot 2026-10-11.