What xAI launched on October 14
xAI's Grok Voice 3.0 went live on October 14, 2026, priced at $0.08 per minute. Developers whose calls ran through the grok-voice-latest alias were moved to the new model automatically, with no code change on their side. Some production voice agents changed behavior before their owners had made any decision about it.
The model is a speech-to-speech system. It takes audio in and answers in audio directly, rather than chaining speech-to-text, a language model and text-to-speech, as many voice stacks do, according to DevX. xAI pitches it at phone support lines, contact centers and in-app assistants.
That pitch lands in a fast-growing enterprise category. Companies that already run AI phone agents, or are weighing them, now have a new price to measure against. Developers who built on the alias face a different problem: they must decide whether to keep following a moving target or lock in a version they have tested.
Grok Voice 3.0 pricing and what a business would pay
The headline rate is $0.08 per minute, or $4.80 per hour. Blockchain.News reports that pricing stayed flat from Grok Voice 2.0, so the upgrade came without a price increase. For a small company, the per-minute rate sets the monthly bill more directly than any benchmark does.
Scale is where the numbers get concrete. One analyst's arithmetic, cited by Vizaca, puts a business handling 2,000 calls a month at an average of five minutes at about $800 a month in model usage. That is 10,000 minutes multiplied by $0.08. The figure covers model usage only. Phone numbers, telephony charges and the engineering time to build the agent sit outside it, and the sources reviewed do not price them.
The missing piece is competitor pricing. The sources reviewed name OpenAI's GPT-Live-1 and Google's Gemini 3.8 Live as the rivals in xAI's comparison, but they do not list per-minute rates for either. Until those numbers are public, a buyer cannot run a clean cost comparison. Enterprise buyers already cap AI spending. Accenture's limit on central Claude spending at $29 a month per employee is one recent example of how that works in practice.
The grok-voice-latest switch and how to pin your model
The alias change is the part most likely to surprise a production team. According to a DEV Community guide, calls that used grok-voice-latest moved to Grok Voice 3.0 on October 14 with no redeploy. xAI deprecated Grok Voice 2.0 in the same release notes, and the sources reviewed do not give an end date for that retirement.
An alias works like a pointer. When the vendor moves it, every caller that references it gets the new model on the next request. A voice agent tuned around the old model's behavior may respond differently after the switch, and a normal deploy log will not flag that.
Developers can avoid the surprise with a few steps:
- Search repositories, environment files and infrastructure configs for the string grok-voice-latest.
- Pin the specific model name grok-voice-3.0 in production if you need stable behavior, as the DEV Community guide recommends.
- Run your own call scripts against the pinned model before any switch, covering interruptions, accents, hold music and background noise.
- Log the exact model name on every session, so a caller complaint can be traced to a model version.
- Put the 2.0 retirement on your calendar and watch xAI's release notes for a date.
Benchmark claims that still need checking
xAI says Grok Voice 3.0 ranks first on the Artificial Analysis Speech-to-Speech Index with a score of 86.9. Unite.AI reports that xAI cited Artificial Analysis testing dated October 13. Blockchain.News, by contrast, says the verified release notes confirm pricing and migration but not the performance claims. A Basenor post published two days ago says no third-party evaluation had been published when it was written.
Those accounts do not agree. Until the score appears on the Artificial Analysis site, treat it as a company-reported figure. The table below lists the performance numbers in the launch materials and how each one stands.
| Claim | Figure as reported | Status |
|---|---|---|
| Speech-to-Speech Index | 86.9, versus 83.2 for GPT-Live-1 and 82.6 for Gemini 3.8 Live | Company-reported by xAI |
| Full Duplex Bench | 97% | Company-reported by xAI |
| Response latency | Under one second, described as 40% lower than before | Company-reported in xAI's launch post |
Phrases such as "world's most intelligent" and "world's leading" are promotional and unverified. They show how xAI wants the product read, but they do not tell a buyer how the model performs on their calls. A practical test is to run a sample of real call recordings through the new model and your current stack, then measure task completion, transfer rate and caller satisfaction.
Where Tesla and Starlink fit in
xAI says Grok Voice already runs millions of daily conversations across Tesla and SpaceX products. The company also says it answers about 20,000 Starlink support calls a day, and that roughly 70% of them involve no human. At those rates, about 14,000 calls a day would be handled without staff. These are xAI's own numbers and have not been independently audited.
Musk praised the model on X the same day. A Basenor blog post quotes him calling Grok Voice 3.0 "the world's leading conversational voice model." Because the quote comes from a secondary blog, check the wording against Musk's original post before repeating it.
For consumers the effects are less clear. Tesla owners may hear the upgraded voice assistant in their vehicles, and Starlink customers may reach it on support lines, but xAI has not published a consumer rollout schedule in the sources reviewed. Naming is shifting too. eesel.ai reports that xAI now signs its developer documentation as SpaceXSI, which is why launch coverage uses both names.
What to watch next
Three things will show whether this launch holds up. The first is independent testing. If Artificial Analysis confirms the 86.9 score, Grok Voice 3.0 gains a credible claim to the top of the speech-to-speech category. If the score never appears, xAI will face questions about the comparison it chose to publish.
The second is pricing. A flat $0.08 per minute gives xAI an easy pitch to small businesses, and rivals will have to answer with their own per-minute rates or bundles. The third is the 2.0 retirement date, the one deadline in the sources that could force engineering work on a fixed schedule.
Regulators are also moving. EU officials have opened their first AI Act probes into OpenAI and Anthropic, and any voice product handling customer calls at scale may face similar questions about disclosure and accuracy. The next few weeks will show whether xAI's numbers hold up on ordinary production calls, and whether OpenAI and Google respond on price.
