Skip to content

Grok Voice 3.0: how to read the figures announced for an AI voice agent

On 14 October 2026, xAI unveiled Grok Voice 3.0, a model built for voice agents that listens and replies directly by voice. The announcement lines up flattering test results and one striking example: about 70% of calls resolved without a human in Starlink's call centres. For a French business considering an automated phone reception service, these figures deserve a careful reading: they come from the vendor, and several points that are essential for use in France are missing from the announcement.

What xAI announced

Grok Voice 3.0 is a so-called "speech-to-speech" model: it receives the caller's voice and directly produces a spoken reply. xAI makes it available in its voice agent API, that is, a building block for developers who build their own service, not a ready-to-use switchboard.

The vendor announces a price of $0.08 per minute, or $4.80 per hour, the same as the previous version. On its pricing grid, Grok Voice 2.0 is now marked as deprecated. xAI also claims latency "under one second", with no more precise figure in the text of the announcement.

Test results that need putting in context

xAI highlights first place in the Artificial Analysis speech-to-speech ranking, with a score of 86.9. It also cites a test called τ-Voice, which, according to its description, measures an agent's ability to handle support calls with varied accents, background noise and interruptions. On this test, Grok Voice 3.0 scores 71.6%, against 74.5% for the competing model GPT-Live-1: xAI describes this as "comparable", stressing a latency that is 40% lower.

On τ-Voice, the model is therefore not in the lead; xAI highlights another customer service test, ServiceNow's EVA-Bench, where it achieves the best accuracy. Results vary depending on the test chosen, and these comparisons are presented by the vendor itself: for phone reception, what matters is quality on your own calls, in your language, with your customers.

70% of calls resolved without a human: an example, not a promise

The most striking figure concerns Starlink. According to xAI, its call centres use Grok Voice to answer 20,000 customer calls a day, in multiple languages; about 70% are resolved without human intervention, and one sales call in four ends in a purchase before the caller hangs up.

This is data reported by the vendor about one customer, with no detailed method: we do not know exactly what "resolved" covers, or which types of request are involved. The calls of a satellite internet provider (customer verification, hardware troubleshooting, issuing credits, sending replacement equipment) are not the same in nature as those of a practice, a tradesperson or an agency. This rate therefore cannot be carried over as it stands to a small business.

French and hosting: what the documentation does not specify

The announcement mentions "multiple languages" without naming them. xAI's technical voice documentation states that transcription supports "38+" languages, meaning at least 38, but it does not publish the list on that page: French is not named there. For a French business, this is the first thing to get confirmed, through a real trial with callers who speak fast, with an accent or in a noisy environment.

On data, the same documentation mentions EU data residency options and regional processing, without naming the regions concerned. It also states that audio is processed in real time and is never stored or used for training; what remains unknown is what happens to the transcripts and summaries produced by the provider that builds the service. Before entrusting customer conversations to a service built on this model, it is reasonable to ask in writing where calls are transcribed, processed and kept.

A model that changes without notice, and cloned voices

xAI states that agents using the "grok-voice-latest" name were switched automatically to the new version, with no action on their part. This is convenient, but it means a voice agent can change behaviour overnight. A business should find out whether its provider pins the version in use and tests each change before applying it to its calls.

The documentation also describes creating a custom voice from an audio sample of at most 120 seconds. According to this documentation, the feature is for now open only in the United States, except Illinois: it therefore does not apply to a French business today. The technical pages consulted also do not detail any condition tied to the consent of the person whose voice is reproduced. Whatever the tool, giving a switchboard the voice of a manager or an employee therefore means, on your side, obtaining their explicit agreement and checking the approach with legal counsel.

Questions to ask before choosing a voice agent

This announcement confirms that voice models are progressing fast. Above all, it provides a framework for reading such claims. When faced with an AI phone reception offer, five simple questions: has French been tested on real calls? Where are conversations processed and stored? Is the model version pinned, or does it change automatically? Is a cloned voice used, and with what consent? Is the caller told at the start that they are speaking to an AI?

For its part, Alyas announces the AI and the recording at the start of the call, answers from the information supplied by the business, transcribes and summarises conversations, transfers the call to the chosen number when necessary, and its hosting is located in the European Union.

Frequently asked questions

Does Grok Voice 3.0 speak French?

xAI's announcement mentions "multiple languages" and the documentation indicates "38+" languages for transcription, without publishing the list on the page consulted. French is not named there: it needs to be confirmed by the vendor or through a real trial.

How much does Grok Voice 3.0 cost?

xAI lists $0.08 per minute of conversation, or $4.80 per hour, in its voice agent API. The grid adds $0.004 per text input and does not mention telephony. This price covers the model: integrating and running a complete phone reception service remains to be budgeted.

Can a voice agent really handle 70% of calls without a human?

That is the figure xAI attributes to Starlink's call centres, with no detailed method. The real rate depends heavily on the nature of the calls: repetitive requests lend themselves better to automation than unusual or urgent situations.

Sources

Further reading

Does your phone ring while you're busy working?

Try an Alyas chat agent free for 10 days, no card required. The voice agent is included from Pro.