xAI released Grok Voice 3.0 on 14 October as a speech-to-speech model on the Voice Agent API, and kept the price where it was: $0.08 per minute, the same as Grok Voice 2.0. First audio arrives in under a second, and agents already pointed at the grok-voice-latest endpoint were migrated automatically rather than having to be re-pointed.
On the Artificial Analysis Speech-to-Speech Index it took first place at 86.9, ahead of GPT-Live-1 at 83.2, Gemini 3.8 Live at 82.6 and its own predecessor at 81.3. Component scores were 98.5% on Big Bench Audio, 97% on Full Duplex Bench and 94.8% task success. It does not sweep everything: on the tau-Voice benchmark OpenAI still leads at 74.5% against Grok at 71.6%, though xAI reports roughly 40% lower latency. On ServiceNow EVA-Bench, Grok tops the table at 74% accuracy.
The deployment claims are xAI's own and unpublished, so treat them as marketing until someone outside the company reproduces them: millions of conversations a day in Tesla vehicles, and about 20,000 multilingual calls a day at Starlink call centres with roughly 70% closed without a human. The verifiable part is the price and the index position, and on both counts this is the cheapest way to get a top-ranked voice agent into production today.
What changed for existing users
Nothing to do if you are on grok-voice-latest — you are already on 3.0. If you pinned a version, you keep 2.0 until you change the endpoint. Pricing did not move, so there is no migration cost beyond regression-testing your own prompts.
Sources:
xAI ·
Unite.AI