Is τ-voice still meaningful?

Known problems

• The paper found the best voice agent at 51% with clean audio. Artificial Analysis reported 79.9% on its own harness, and the two may not compare.^{[1][1] τ-voice, arXiv 2603.13686 Primary[3][3] Artificial Analysis, speech-to-speech Primary}
• Voice agents reach 31–51% with clean audio and 26–38% with realistic noise and accents.^{[1][1] τ-voice, arXiv 2603.13686 Primary}
• The text-mode reference of 85% pass@1 comes from one model, GPT-5 with reasoning.^{[1][1] τ-voice, arXiv 2603.13686 Primary}

What to look at instead

τ-voice extends τ²-bench to full-duplex voice calls.^{[1][1] τ-voice, arXiv 2603.13686 Primary}

• τ²-bench Status:Near ceiling Fading. Top scores are high, and the newer τ³-Banking spreads models more.

Where it appears in the report

Chapter 7, Multimodal, voice and embodied

• Figure 1. Tests Beyond Text Fill Up in Months, Too
• Table 1. Each modality has one number that tells its story
• Speech reasoning is solved, but voice agents lose most of their text-mode success
• Figure 6. The same tasks drop from 85% in text to as low as 26% in voice
• Table 10. Single-turn speech tests have saturated; multi-turn voice agents remain the frontier in “Go deeper: Speech-to-speech and voice-agent benchmarks”
• Figure 11. Robotics, GUI agents and deepfake detection have the weakest independent checks