xAI released Grok Voice 3.0, a new speech-to-speech model for its Voice Agent API, on October 14, 2026. It replaces Grok Voice 2.0 at the same price, and it now holds the top score on Artificial Analysis’s independent Speech-to-Speech Index. Developers who already use the grok-voice-latest alias were moved to the new model automatically.
Key details
- Model ID:
grok-voice-3.0. Thegrok-voice-latestalias now points to it. - Price: $0.08 per minute ($4.80 per hour), plus $0.004 per text input, unchanged from Grok Voice 2.0.
- Grok Voice 2.0: marked deprecated. Apps pinned to
grok-voice-2.0must change themodelsetting to upgrade; the API itself is unchanged. - Latency: xAI says responses start in under one second.
- Availability: hosted API only; there are no open weights.
Benchmarks: strong, with one caveat
On the Artificial Analysis Speech-to-Speech Index, Grok Voice 3.0 scores 86.9, ahead of OpenAI’s GPT-Live-1 (Astra, medium) at 83.2 and Google’s Gemini 3.8 Live with extended thinking (high) at 82.6. Its predecessor scored 81.3. This index is run by a third party, which makes it the most useful single number in the launch.
The model does not win everything. On τ-Voice, a test of simulated customer-support calls that xAI cites in its announcement, Grok Voice 3.0 scores 71.6% against 74.5% for GPT-Live-1 (Astra, medium). xAI says it matches that result at about 40% lower latency. Other figures in the announcement, such as 97.0% on Full Duplex Bench (turn-taking and interruptions) and 74% accuracy on ServiceNow’s EVA-Bench, are reported by xAI and have not been independently reproduced.
How it compares on price
| Model | S2S Index | Price |
|---|---|---|
| Grok Voice 3.0 (xAI) | 86.9 | $0.08 / min |
| GPT-Live-1, Astra medium (OpenAI) | 83.2 | $0.05 / min, billed per second |
| Gemini 3.8 Live, extended thinking high (Google) | 82.6 | Not compared here |
Per minute, GPT-Live 1 is cheaper on list price. Per task the gap can narrow: in Artificial Analysis’s Big Bench Audio run, as cited by xAI, a task cost about $0.025 on Grok Voice 3.0 versus $0.030 on GPT-Live-1 (Astra, medium) and $0.055 on GPT-Realtime-2.1 (high). Your own costs will depend on call length and how much backend work each call triggers.
Who should care
- Teams building phone and support agents: worth a side-by-side test against GPT-Live 1, especially on interruption handling and your own scripts, not just public benchmarks.
- Existing Grok Voice users: if you use the alias, you are already on 3.0; re-check prompts and call flows. If you pinned 2.0, plan the switch now that it is deprecated.
- Non-English deployments: test your languages before routing live calls; launch materials give few language-level results.
See more options in best AI for voice and audio and best AI for customer support, or read our Grok overview.
