The case for putting an AI model on the phone has always been arithmetic, and xAI arrived with the numbers attached to customers that already use it.
The company released Grok Voice 3.0 on October 14, a speech model sold through a voice agent API. The price stays at 8 cents per minute. xAI said the software runs in Tesla vehicles for in-car voice interaction and in Starlink’s customer service centers, where it answers questions from subscribers about hardware, billing and connectivity.
The figures the company offered for that deployment are the ones a sales team would want on a slide. The Starlink centers handle about 20,000 multilingual calls a day, and xAI said roughly 70 percent are resolved without a person joining the conversation. The company put response latency at about 0.81 seconds.
Latency is the measure that decides whether a voice agent is bearable. Above about a second, callers begin talking over the system, and the exchange starts to feel like radio protocol rather than conversation. The engineering that gets a model under that threshold is largely about removing steps: speech recognition, a language model and speech synthesis are fast on their own, and slower when bolted together with network round trips between them.
Eight cents a minute is the unit of account in this market because that is how the buyer budgets. It works out to $4.80 an hour of talking, which a call center can compare directly with the loaded cost of an agent and the time customers spend on hold. Holding the price while releasing a new version is a claim about serving costs as much as it is a product launch.
The competitive field is crowded at that price. OpenAI sells realtime speech models to developers, and a group of specialist vendors sells voice synthesis and conversation tools that plug into the same workflows. What separates them for a buyer is rarely intelligence in the abstract. It is whether the agent sounds natural, whether it handles interruptions, whether it understands accented speech and phone-line noise, and whether it knows when to hand a caller to a person.
The internal customers carry the launch. Tesla and Starlink are both controlled by Elon Musk, whose other companies give xAI a place to run the model at volume without winning a competitive bid. That volume is useful for testing and for reference calls with prospects. It is not the same as third-party demand, because a call answered inside the same corporate family can be reviewed, tuned and forgiven in ways an outside customer will not forgive.
Multilingual handling matters more for Starlink than for most call centers. Subscribers are spread across many countries, and a support organization that answers 20,000 calls a day in a range of languages cannot staff every one of them with someone who speaks the caller’s language. Models do that part more cheaply, which is why satellite and telecom operators have been early buyers of voice agents.
The unresolved part of the pitch is the 30 percent that still reaches a human. Those calls tend to be the harder ones: the customers with complicated problems, the ones who have already called twice, the ones who are already angry. Containing the easy majority reduces call volume without reducing the staffing needed for the difficult remainder, which is why call center managers measure deflection and escalation separately.
Voice agents also bring a set of obligations that text agents do not. Calls are recorded in many jurisdictions, voice can function as a biometric identifier, and consent rules differ by country. A support center that automates most of its volume inherits those questions with the volume.
The economics underneath the per-minute price are the reason the market is competitive. Serving speech costs money for graphics processors, network capacity and the inference that keeps a conversation going while the caller is still talking, and a model that spends more compute per minute must charge more or lose money. A price held steady across two generations of the product implies the company expects the serving cost per conversation to fall.
Audio is also where AI systems meet the parts of language that text hides. Tone, hesitation, background noise and interruptions all carry meaning, and a support agent that misreads a caller’s frustration makes a problem worse. That is why the deployments that matter most for a vendor are the ones with consequences attached, a complaint line or a dispatch desk, rather than a demo that books a table.
What the company did not say is how many external customers have signed for the API, how the model performs in each language, or how the resolution rate holds up over a longer window than a single reference period. Those are the numbers enterprise buyers will ask for in pilots, and the ones xAI has not published.
For now the arithmetic is the argument. An agent that answers a call for less than a dime a minute, at a latency close to human turn-taking, is priced to replace a category of labor that companies have been trying to automate since the first phone tree. Whether that price holds depends on what it costs xAI to serve each minute, and on whether competitors decide to undercut it.
