SpaceXSI’s launch art for Grok Voice 3.0. Image: SpaceXSI, launch art from the announcement
Elon Musk’s AI company says its new voice model already answers 20,000 Starlink support calls a day and fixes about 70% of them without a human. SpaceXSI, the company formerly known as xAI, announced Grok Voice 3.0 on October 14, calling it “the world’s most intelligent speech-to-speech model”. It is live now in the company’s Voice Agent API for developers building phone and app agents, at the same price as the version it replaces.
What Grok Voice 3.0 is
Grok Voice is a speech-to-speech model: it listens to a caller and talks back directly, rather than turning speech into text, running it through a chatbot and reading the answer aloud. Developers connect to it over a live WebSocket link, give it tools such as web search, and pick from a set of built-in voices or a cloned custom voice, according to the developer docs.
It is the third version this year. Grok Voice 1.0 came to the API in April and Grok Voice 2.0 followed in July, when the company said its voice models “reason through queries while speaking”. The new model promises replies in under a second and better handling of pauses, turn-taking and callers who talk over it.
The benchmark claims
SpaceXSI leans on independent tester Artificial Analysis. By its figures, Grok Voice 3.0 scores 86.9 on Artificial Analysis’s Speech-to-Speech Index, ahead of OpenAI’s GPT-Live-1 (83.2), Google’s Gemini 3.8 Live (82.6) and its own Grok Voice 2.0 (81.3). Other numbers in the announcement include:
- Full Duplex Bench (pauses and interruptions): 97%, up from 95.1%
- Big Bench Audio (spoken reasoning): 98.5%
- τ-Voice (support calls with accents and background noise): 71.6%, which the company says matches GPT-Live-1 with 40% lower latency
- EVA-Bench, ServiceNow’s test for business voice agents: 74% accuracy, the highest score
The gains over Grok Voice 2.0 are modest on some tests: task success moves from 94.6% to 94.8%. The table SpaceXSI published also picks its rivals’ settings, so these are the company’s chosen comparisons rather than a full league table.
ServiceNow, whose benchmark is quoted, gave the launch its backing:
Grok Voice 3.0 sets a new Pareto frontier on EVA-Bench, ServiceNow’s open-source benchmark scoring voice agents on accuracy and experience across 200+ realistic enterprise scenarios.
Nitin Aggarwal, Senior Director of Product Management, ServiceNow
Already running in Teslas and at Starlink
The more striking claims are about real use. SpaceXSI says Grok Voice 3.0 handles millions of conversations a day in Tesla cars, where drivers already use Grok Bot to order coffee or book dinner by voice, and it now powers Grok Bot’s voice mode too.
At Starlink, the company says Grok Voice takes 20,000 customer calls a day in several languages. Roughly 70% are resolved with no human involved, and one in four sales calls ends in a purchase before the caller hangs up. A single agent checks who the caller is, troubleshoots hardware, issues credits and ships replacement kit. These figures come from SpaceXSI and haven’t been independently checked.
Price and who can use it
Grok Voice 3.0 costs $0.08 a minute ($4.80 an hour, about €4.10 or £3.60), unchanged from Grok Voice 2.0, according to the pricing page. Developers whose agents point at grok-voice-latest have been moved over automatically. Anyone with an API account can try it from the company’s console; the announcement lists no regional limits, and the built-in voices can speak every language the model supports. It is aimed at businesses building agents, not a new feature in the Grok app, which has just been folded into The X Subscription.
Why it matters
Voice is where AI agents meet ordinary customers, and Musk’s companies are using their own call centres and cars as the test bed. If the Starlink numbers hold up, a 70% no-human resolution rate at under five dollars an hour puts real pressure on OpenAI and Google, and on the people who staff support lines.
Sources: SpaceXSI, “Introducing Grok Voice 3.0”; Grok Voice developer docs; xAI API pricing; SpaceXSI, “Introducing Grok Voice 2.0”.
Latest SpaceXAI news
