What shipped
xAI released Grok Voice 3.0 on Tuesday, calling it the world’s most intelligent speech-to-speech model. It is available now in the company’s Voice Agent API at $0.08 a minute, or $4.80 an hour — the same price as Grok Voice 2.0 — and agents pointed at grok-voice-latest were switched over automatically with no changes required.
The headline capability is latency. xAI says time to first audio is under one second, and about 40% lower than GPT-Live-1 on its own latency chart.
The benchmark table, read closely
xAI reports a score of 86.9 on the Artificial Analysis Speech-to-Speech Index, which it says is first place, against 83.2 for GPT-Live-1, 82.6 for Gemini 3.8 Live and 81.3 for its own previous version. It also lists 98.5% on Big Bench Audio for speech reasoning, an Elo of 1070 on Speech Agent Arena, a 94.8% task success rate, 97.0% on Full Duplex Bench, and 74% on ServiceNow’s EVA-Bench. All of these figures are published by xAI; they are the vendor’s account of its own model.
The interesting row is τ-Voice, which tests support calls. xAI reports 71.6% and describes it as comparable to GPT-Live-1 — whose score on the same chart is 74.5%. On the single benchmark closest to the commercial use case xAI is selling, its model is behind, and the company published that rather than leaving it out.
On cost per task on Big Bench Audio, xAI puts Grok Voice 3.0 at $0.025 against $0.030 for GPT-Live-1 and $0.055 for GPT-Realtime-2.1.
Where it is already running
xAI says the model is used in Tesla vehicles for millions of conversations a day and powers Grok Bot’s voice mode, and that it handles about 20,000 customer calls daily across multiple languages in Starlink call centres, where roughly 70% are resolved without a human and about one in four sales calls ends in a purchase. Those are deployment figures reported by the company about businesses in the same corporate orbit, and no independent audit of them has been published.
What to watch
Artificial Analysis maintains the index xAI leads with, so the first independent confirmation of the 86.9 is a matter of waiting for the maintainer rather than guessing. The second thing to watch is whether holding the price flat across a generation becomes the pattern in voice, where per-minute billing makes a price cut harder to hide than a per-token one.
