On 14 October 2026 xAI introduced Grok Voice 3.0, the new version of its voice model for voice agents. It takes in spoken language and answers in spoken language. It is used through the Voice Agent API at the same rate as the previous version: $0.08 per minute.
What happened
Several reports agree that xAI made Grok Voice 3.0 available on 14 October 2026 as a speech-to-speech model in the Voice Agent API, at $0.08 per minute. That equals $4.80 per hour and is the same rate as for Grok Voice 2.0. Applications that use the alias grok-voice-latest were switched to version 3.0 automatically, with no code changes needed.
What the model can do
According to the reports, Grok Voice 3.0 handles 20 languages with automatic detection. It can therefore switch language in the middle of a conversation without a language being set in advance. It comes with 28 built-in voices, and teams can clone a custom voice from a reference clip of up to 120 seconds. Interruptions are handled on the server side through voice activity detection.
On speed: the time to first audio is about 0.81 seconds. That is slower than the predecessor, where it was 0.70 seconds. According to one review, the competing model GPT-Live-1, by contrast, answers only after 1.34 seconds.
The ratings
According to xAI's announcement, an independent test by Artificial Analysis dated 13 October 2026 puts Grok Voice 3.0 first on the Speech-to-Speech Index, with 86.9 points. This index blends, as one report explains, spoken reasoning, simulated customer-support calls and blind preference tests. xAI calls the model the world's most intelligent speech-to-speech model.
On cost per task, xAI gives $0.025 for Grok Voice 3.0, $0.030 for GPT-Live-1 and $0.055 for GPT-Realtime-2.1 in one run of the Big Bench Audio test. These are figures from xAI that refer to Artificial Analysis.
One report contradicts the big wording: on the customer-support test τ-Voice, Grok scores 71.6%, whereas GPT-Live-1 scores 74.5%. On Full Duplex Bench, which measures how naturally a model handles pauses, turn-taking and interruptions, Grok Voice 3.0 reaches 97%. One review therefore calls the phrase "most intelligent model in the world" marketing.
Background
Grok Voice 3.0 comes in a line: the Voice Agent API already existed with Grok Voice 2.0 at the same rate of $0.08 per minute. In the service's beta phase the rate was $0.05 per minute. As comparison models, the reports name GPT-Live-1 and GPT-Realtime-2.1.
The figures at a glance, each according to xAI or reports about it:
| Measurement | Grok Voice 3.0 | Comparison |
|---|---|---|
| Speech-to-Speech Index (Artificial Analysis) | 86.9, first place | – |
| Time to first audio | 0.81 s | Predecessor 0.70 s; GPT-Live-1 1.34 s |
| Customer-support test τ-Voice | 71.6% | GPT-Live-1 74.5% |
| Full Duplex Bench | 97% | – |
| Cost per task (Big Bench Audio) | $0.025 | GPT-Live-1 $0.030; GPT-Realtime-2.1 $0.055 |
xAI gives a deployment example from its own environment. According to it, Starlink call centres answer 20,000 customer calls a day with Grok Voice in several languages, and about 70% of them are resolved without a human. These figures come from xAI and are not independently confirmed.
The sources are inconsistent on the company's name. Depending on the report, the provider is called xAI, SpaceXAI or SpaceXSI; the same company behind Grok is meant. This article keeps "xAI", as the original announcement does.
On cost, the reports point to pitfalls. Per-minute rates cannot easily be compared between providers, because the billing units differ. Tool calls may be billed separately. And some third-party pages still quote $0.05 per minute, the rate from the service's beta phase; it is out of date.
- Voxtral 3: Mistral AI Bets on Real-Time Voice Agents
- OpenAI Decisions API: you only pay for the input
- “The X Subscription”: Grok, Cursor and X Premium in one plan
- Claude Duet: Anthropic Launches a Proactive AI Agent
What this means
For developers building voice agents, the per-minute rate stays the same, and the switch to version 3.0 happens by itself through the alias. Anyone choosing the model should measure their own application: the figures on response time and the customer-support test are mixed, and the big claims come from the provider.
For companies with customer service, the announcement is a sign that voice agents are becoming cheaper and broader on the market. Whether a particular use works depends on your own use case; the statement that about 70% of calls at Starlink are resolved without a human is a self-report by xAI.
For everyone else, the announcement mostly concerns technology in the background. It shows that providers offer their voice models for customer conversations in several languages; whether and where you yourself will talk to such an agent is not stated in the sources.
A note on interpretation: the sources evaluated are mostly announcements and reports about them. An independent test confirming all the figures named was not available.
Frequently asked questions
What does Grok Voice 3.0 cost?
$0.08 per minute in the Voice Agent API, which is $4.80 per hour. The rate is the same as for Grok Voice 2.0. Tool calls may be billed separately.
What is a speech-to-speech model?
A model that understands spoken language directly and answers directly in spoken language, instead of first converting speech into text and text back into speech. It is intended for voice agents, for example in customer support.
Do developers have to change anything?
Not necessarily: according to reports, applications using the grok-voice-latest alias were switched to version 3.0 automatically, with no code changes needed.
Is Grok Voice 3.0 really the best voice model?
It ranks first on Artificial Analysis's Speech-to-Speech Index with 86.9. On a customer-support test (τ-Voice), however, it scores 71.6% against 74.5% for GPT-Live-1, and its time to first audio rose from 0.70 to 0.81 seconds. Many of the figures come from xAI itself.
