xAI has launched Grok Voice 3.0 at $0.08 per minute, with improved reported voice-agent scores and no price increase. For existing developers, the immediate detail is operational: xAI says agents using grok-voice-latest have already switched to the new model automatically. The launch gives voice-agent builders a substantial new option, and gives existing customers a reason to run their tests rather than assume an unchanged bill means unchanged behaviour.
xAI Watch analysis
What happened
The speech-to-speech model became available through xAI’s Voice Agent API on 14 October, according to Unite.AI’s account of the announcement. Its $0.08-per-minute rate, equivalent to $4.80 an hour, matches Grok Voice 2.0.
xAI cites Artificial Analysis testing dated 13 October that puts Grok Voice 3.0 first on the Speech-to-Speech Index at 86.9, compared with 81.3 for its predecessor. That index combines speech reasoning, simulated customer-support performance, listener preference and task success. It is a composite score, not a promise that the model wins every individual test.
What the results show
- A stronger overall voice score
xAI’s cited comparison puts Grok Voice 3.0 ahead of GPT-Live-1 at 83.2 and Gemini 3.8 Live at 82.6 on the composite index. - A sizeable customer-support improvement
The reported tau-Voice score rises from 56.5% to 71.6%, although GPT-Live-1 remains ahead at 74.5% in that comparison. - Better conversational timing
xAI reports 97.0% on Full Duplex Bench, up from 95.1%, testing behaviour around pauses, turn-taking and interruptions. - Enterprise evaluation adds another signal
xAI reports 74% accuracy on ServiceNow’s EVA-Bench, which evaluates complete spoken conversations across customer service, IT and HR scenarios.
Why it matters
The useful proposition is better reported performance for the same per-minute charge. Customer-support agents must do more than sound pleasant: they need to use tools, follow policies and leave the right account or booking state behind. Tau-Voice tests that final state, making its improvement more informative than a polished sample conversation alone.
The automatic switch matters just as much as the leaderboard. Existing integrations can receive a different model without a code change. Teams should check task completion, interruption handling and tool behaviour against their own scenarios. The invoice may look familiar; the conversation need not.
Our read
This earns a place on a voice-agent shortlist. The reported support-task improvement is substantial, and holding the price makes the upgrade commercially interesting rather than merely another leaderboard ceremony.
But choose the metric that matches the job. Grok Voice 3.0 leads the reported composite index while trailing GPT-Live-1 on the cited customer-support test. That is not a contradiction. It is a reminder that “best voice model” becomes a much more useful claim once someone specifies what the caller needs done.
For existing users, prioritise regression testing. For prospective customers, compare successful task completion and conversational quality alongside the per-minute price.
What to watch
- Whether developer evaluations reproduce the reported gains on real support workflows and varied accents.
- How automatically migrated agents behave on interruptions, tool calls and policy-sensitive requests.
- Whether the benchmark improvements reduce repeat calls and escalations, rather than simply producing smoother conversations.
Discussion spark: Should voice-agent providers automatically move production aliases to better-scoring models, or should behavioural changes require explicit customer approval even when the price stays the same?
Sources and evidence
- xAI Introduces Grok Voice 3.0 Speech-to-Speech Model for Voice Agents (14 October 2026, 17:00 UTC)
not affiliated with or endorsed by xAI, SpaceXAI, SpaceX, X Corp. or Cursor