xAI makes Grok Voice 3.0 available for voice agents

The company behind Grok announced Grok Voice 3.0 on October 14, 2026, as a speech-to-speech model for voice agents. It is available immediately through the Voice Agent API, and the `grok-voice-latest` alias has already been moved to the new version. That makes the release relevant to developers who want to test or deploy conversational systems without changing their existing model name.

The model is designed for conversations in which an agent has to listen, reason, speak and handle interruptions. The company says it offers first audio in under one second and can keep the same price as Grok Voice 2.0: $0.08 per minute, or $4.80 per hour. It also describes use cases such as customer support, sales calls and hands-free tasks in vehicles. These are availability and pricing details from the provider, not a guarantee that every application will have the same latency or cost.

There is independent context for the performance claims. Artificial Analysis placed Grok Voice 3.0 first on its Speech-to-Speech Index with a score of 86.9, ahead of OpenAI’s GPT-Live-1 and Google’s Gemini 3.8 Live in that comparison. It also scored 97% on the Full Duplex Bench for turn-taking and interruptions. SiliconANGLE notes a more mixed result: OpenAI’s model led the τ-Voice support-call test, while Grok reached first audio about 40% faster. ServiceNow’s EVA-Bench gave Grok a 74% accuracy score across more than 200 enterprise scenarios.

The announcement also points to current deployments. The company says Tesla vehicles handle millions of Grok Voice conversations per day, while Starlink call centers use the system for about 20,000 calls daily in several languages. It reports that roughly 70% of those calls end without a human joining and that one in four sales calls results in a purchase. Those operational figures remain company claims rather than an independently audited benchmark.

For makers and businesses, the practical shift is from voice demos toward measurable service workflows. A faster, API-accessible model can reduce waiting time and make interruption handling more natural, but teams still need to test transcription, accents, permissions, escalation and error recovery. A benchmark lead does not remove the need for human oversight when an agent can issue credits, change orders or communicate with customers.