Next upDPTV HighlightsEpisode 1 · From SF Tech Week to TechCrunch Disrupt
News

xAI launches Grok Voice 3.0 in its Voice Agent API

xAI released Grok Voice 3.0 at the existing $0.08-per-minute audio price, moving its latest-model API alias without requiring client changes.

D
Oct 14, 2026 · 2 min read

xAI released Grok Voice 3.0 for speech-to-speech applications in its Voice Agent API on October 14. The model is available under the grok-voice-3.0 identifier.

The migration is automatic for developers using grok-voice-latest. xAI’s release notes say the alias now points to the new model, existing sessions using it were upgraded without client changes, and the API surface is unchanged. Teams that pin model versions can select grok-voice-3.0 directly.

Audio remains $0.08 per minute, or $4.80 per hour, matching the listed rate for Grok Voice 2.0. xAI’s speech-to-speech documentation lists a separate $0.004 charge for each billable text-input conversation.item.create event. The real-time WebSocket service accepts audio and text, returns both formats and supports function calling.

xAI also deprecated grok-voice-2.0 on October 14, directing developers to grok-voice-3.0 or grok-voice-latest. Its release notes do not give an end-of-life date for the deprecated model.

An independent speech-to-speech benchmark table lists Grok Voice 3.0 with an 86.9 index score, the highest entry in the table reviewed. The table also lists 97.0% for conversational dynamics, 71.6% for agentic performance, 1,070 Arena Preference Elo and 94.8% task success.

xAI described the model as delivering class-leading latency under one second. The independent table measured 0.81 seconds to first audio, but it includes faster results: 0.70 seconds for Grok Voice 2.0 and 0.78 seconds for Grok Voice Fast 1.0. That evidence does not support an unqualified class-leading latency claim.

The company said the model handles millions of conversations a day in Tesla vehicles. It also said Starlink call centers use Grok Voice for 20,000 calls daily across multiple languages, roughly 70% of those calls are resolved without a person, and one in four sales calls results in a purchase before hang-up. The opened evidence did not include independent operational data supporting those volume, resolution or conversion figures.

The release follows xAI’s Team Bots launch in September, another addition to the company’s developer-facing product lineup.

More news