The newest speech-to-speech model behind the Voice Agent API, Grok Voice 3.0, is out from SpaceXSI, and it is being pitched at organizations that put voice agents on customer support and sales calls. Artificial Analysis, an independent testing outfit, places it at the top of its class with a score of 86.9 on the Speech-to-Speech Index.
GPT-Live-1 from OpenAI, with reasoning handled by GPT-6 Astra, came in at 83.2. Gemini 3.8 Live from Google recorded 82.6.
Where it wins and where it does not
A grade of 97 percent went to Grok Voice 3.0 on Full Duplex Bench, which measures how naturally a system copes with pauses and interruptions. Tau-Voice, a benchmark centered on accented speech and background noise, stayed with OpenAI by a slim margin: 74.5 percent to 71.6 percent.
On price, Grok holds the advantage. Completing a task in Artificial Analysis’s Big Bench Audio run cost roughly 2.5 cents on Grok Voice 3.0, versus 3 cents for GPT-Live-1 and 5.5 cents for OpenAI’s GPT-Realtime-2.1. Accuracy leadership on ServiceNow’s open-source EVA-Bench, spanning more than 200 enterprise scenarios, also went to the Grok model.
Already at scale inside the family
Deployment figures make the strongest case. Millions of Grok Voice conversations happen in Tesla vehicles each day, while Starlink handles about 20,000 customer calls daily in multiple languages — roughly 70 percent resolved with no human involved, and a purchase at the end of one in four sales calls.
Pricing sits at eight cents a minute, the same as the prior release. xAI was acquired by SpaceX in February, and the unit has been renamed twice since, most recently to SpaceXSI this month.
