Tau3-Voice English Retail: leaderboard
Sierra's τ³-Voice leaderboard (taubench.com), English edition: voice agents (speech-to-speech models, realtime APIs and cascaded speech-to-text/LLM/text-to-speech pipelines) handle τ³-bench customer-service tasks over simulated phone calls, using tools against a domain database, while a simulated customer speaks to them. Standard submissions with the default v1.0 voice user simulator. Retail domain: order, return and exchange requests. Score: pass^1 (%), the share of tasks whose final database state matches the goal. Separate from the text Tau3-Bench boards and from the tau-Voice paper's Clean/Realistic boards. Higher is better.
Metric: Pass^1 (%). Source: taubench.com. Saturation forecast: Rough model projection: around 2026. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | gpt-live-1-diamond-alpha | 78.95 | |
| 2 | Pine Voice 1.0 | 78.76 | |
| 3 | Phonic 1 (High) | 76.32 | |
| 4 | Ava Zero | 75.44 | |
| 5 | Pine Voice Preview | 70.18 | |
| 6 | Grok Voice 1.0 | 62.28 | |
| 7 | Grok Voice 2.0 (High) | 59.65 | |
| 8 | gpt-realtime-2 (xhigh) | 47.37 | |
| 9 | qwen3.5-omni-plus-realtime | 46.49 | |
| 10 | Gemini 3.1 Flash Live Preview (Thinking, High) | 45.61 |
Interactive version: theaggregate.ai/benchmark?slug=tau3-voice-english-retail · How It Works · Data refreshed daily, snapshot 2026-11-05.