x.ai Announces Grok Voice 3.0 With Sub-Second Latency

x.ai announced Grok Voice 3.0, a speech-to-speech model that reaches a latency of under one second. The model is available now in the Voice Agent API at $0.08 / min ($4.80 / hr), which is the same price as the previous Grok Voice 2.0.

In independent testing by Artificial Analysis on October 13, 2026, the model received a score of 86.9 on the Speech-to-Speech Index. This index evaluates performance across speech reasoning, preference, and task success. On the Full Duplex Bench, which measures the natural handling of interruptions, pauses, and turn-taking, Grok Voice 3.0 scored 97.0%. For comparison, Grok Voice 2.0 scored 95.1% and GPT-Live-1 (Astra, Medium) scored 94.9% on that same benchmark.

The model also scored 98.5% on Big Bench Audio for speech reasoning and 94.8% for task success. According to the announcement, the model's τ-Voice score of 71.6% is comparable to GPT-Live-1 (Astra), but with 40% lower latency. On ServiceNow’s EVA-Bench, the model achieved an accuracy score of 74% across more than 200 realistic enterprise scenarios. Its Speech Agent Arena Elo is 1070, compared to 1011 for Grok Voice 2.0 and 1048 for GPT-Live-1 (Astra, Medium).

The technology currently powers millions of daily conversations in Tesla vehicles. Drivers use the model to check calendars, answer emails, or perform tasks like ordering coffee through Grok Bot. At Starlink call centers, the model answers 20,000 customer calls per day across multiple languages. The company reports that 70% of these calls are resolved without a human in the loop, and one in four sales calls results in a purchase before the caller hangs up. A single agent built on this model can troubleshoot hardware, issue credits, and ship replacement equipment.

Grok Voice 3.0 is available today for developers. Those using the grok-voice-latest endpoint have been switched to the new model automatically with no changes needed to their existing code.