Robustness on τ-bench (airline, original)
How well does the agent maintain accuracy when inputs are perturbed (faults, structural changes, prompt rewording)? See full definition →
Sub-metric Comparison
Baseline vs Perturbed Accuracy
Agent Leaderboard — Robustness
| # | Agent | Acc | Robustness Agg | Fault | Struct | Prompt | Overall |
|---|---|---|---|---|---|---|---|
| 1 |  Claude Sonnet 4.5 |
54.4% | 0.98 | 0.97 | 1.00 | 0.97 | 0.79 |
| 2 |  GPT-4 Turbo |
35.6% | 0.98 | 0.98 | 0.96 | 1.00 | 0.69 |
| 3 | GPT-5.2 |
42.0% | 0.97 | 1.00 | 1.00 | 0.92 | 0.75 |
| 4 | Claude 3.7 Sonnet |
43.6% | 0.96 | 1.00 | 1.00 | 0.89 | 0.72 |
| 5 |  Gemini 2.5 Pro |
52.8% | 0.96 | 0.97 | 0.98 | 0.92 | 0.74 |
| 6 | Gemini 2.5 Flash |
47.2% | 0.95 | 1.00 | 0.97 | 0.88 | 0.70 |
| 7 | GPT-5.2 (xhigh) |
51.6% | 0.95 | 1.00 | 1.00 | 0.84 | 0.76 |
| 8 | Gemini 2.0 Flash |
32.0% | 0.94 | 0.98 | 1.00 | 0.85 | 0.67 |
| 9 | Gemini 3.0 Pro |
58.8% | 0.94 | 1.00 | 0.92 | 0.91 | 0.77 |
| 10 | O1 |
49.6% | 0.94 | 0.95 | 1.00 | 0.86 | 0.76 |
| 11 | Claude Opus 4.5 |
58.4% | 0.93 | 0.95 | 0.96 | 0.89 | 0.80 |
| 12 | GPT-4o Mini |
21.3% | 0.92 | 1.00 | 0.84 | 0.91 | 0.67 |
| 13 | Claude 3.5 Haiku |
29.6% | 0.88 | 0.86 | 1.00 | 0.77 | 0.68 |