The short answer
● Cheapest top-ranked option: Grok Voice 3.0 at $0.08/min, first on the Artificial Analysis index at 86.9.
● Best for multi-turn task completion: GPT-Live-1.1 — OpenAI still leads tau-Voice at 74.5% vs 71.6%.
● Best if you are already on Google Cloud: Gemini 3.8 Live at 82.6, close enough that migration cost decides it.
How they score
| Model | Speech Index | Price | Released | Notes |
|---|
| Grok Voice 3.0 | 86.9 | $0.08/min | 14 Oct 2026 | Sub-second first audio; auto-migrated on grok-voice-latest |
| GPT-Live-1.1 | 83.2 (1.0) | Not published | 15 Oct 2026 | 13 new voices, six languages improved, Quip mini variant |
| Gemini 3.8 Live | 82.6 | Not published | 2026 | Closest to Grok on index score |
| Grok Voice 2.0 | 81.3 | $0.08/min | 2026 | Still available on a pinned version |
Where each one actually wins
Grok Voice 3.0 is first on the headline index at 86.9, with 98.5% on Big Bench Audio, 97% on Full Duplex Bench and 94.8% task success. It also did not raise its price, which matters more than it sounds: a voice agent handling a thousand ten-minute calls a month costs $800 either way, so the upgrade is free. xAI reports roughly 40% lower latency than OpenAI.
GPT-Live-1.1 began rolling out across ChatGPT and Codex on 15 October with 13 new voices, better accents in six languages and a mini variant called GPT-Live-1.1 Quip. API access was still pending at publication, so API pricing is not yet comparable. The 1.0 version leads tau-Voice at 74.5% against Grok at 71.6%, and tau-Voice is the benchmark closest to multi-step task completion rather than conversational fluency.
Gemini 3.8 Live sits at 82.6, four points behind Grok. That gap is real but small, and for a team already inside Google Cloud the integration saving usually exceeds the quality difference.
How to choose
If you are deploying a voice agent this quarter and price is fixed → Grok Voice 3.0. Top index score at a published, unchanged $0.08/min is the only combination here where you can budget before you build.
If your agent has to complete multi-step tasks rather than converse → wait for GPT-Live-1.1 API pricing. tau-Voice is the benchmark that correlates with that work, and OpenAI leads it.
If latency is your constraint → Grok, on xAI's reported 40% advantage, but test it on your own traffic before committing.
Sources:
xAI ·
Unite.AI