On July 6, 2026, xAI announced 21 new flagship voices for Grok Voice, joining the original five (Ara, Eve, Leo, Rex, and Sal), which were retrained for more natural pacing, phrasing, and emphasis. The new voices are multilingual and cast for specific jobs like support, characters, commentary, advertising, and education, and they are available through the Realtime Voice Agent API, the Text to Speech API, and the new Voice Agent Builder. Grok Voice also supports Expressive Mode with inline speech tags and custom voice cloning from about a minute of audio. The confirmed part is the lineup and the delivery features. The exact size of the full catalog and the head-to-head quality claims are best read as vendor figures until you test them yourself. Voice is billed by API usage, and the official developer rate for realtime voice is 0.05 dollars per minute.
On July 6, 2026, xAI added 21 new flagship voices to Grok Voice and retrained the original five, so Grok now speaks in a much wider cast of voices across support, character, commentary, advertising, and education roles. The voices are multilingual, they are available through three xAI surfaces (the Realtime Voice Agent API, the Text to Speech API, and the new Voice Agent Builder), and they are billed by API usage rather than by a separate consumer price. That is the confirmed core of the news. Two things around it, the exact size of the full catalog and any head-to-head quality claim, are best treated as vendor figures until you test them yourself. This piece separates the two and shows you where the voices fit.
If you have followed Grok Voice, you already know the platform is not new. What changed on July 6 is breadth. Grok went from a small handful of general purpose voices to a lineup deliberately sorted by the job each voice is meant to do, and the five voices most builders already used got a quality pass at the same time. For anyone building a phone agent, a narration pipeline, or an in-app assistant, that is the difference between forcing one voice to cover every use and picking a voice that already sounds right for the task.
What xAI actually announced on July 6
The confirmed announcement is short and specific. xAI published a note titled "21 New Flagship Voices" and rolled the new voices out across its voice products the same day. Three facts hold up across the official page and independent trackers:
- There are 21 new flagship voices, joining the original five. They are multilingual and, in xAI's own framing, cast for specific jobs: support, characters, commentary, advertising, and education.
- The original five voices, Ara, Eve, Leo, Rex, and Sal, were retrained. xAI describes the retraining as an improved recipe that makes pacing, phrasing, and emphasis more natural.
- The voices are available immediately across the Realtime Voice Agent API, the Text to Speech API, and the Voice Agent Builder that xAI launched on July 1, 2026.
That timing matters. The Voice Agent Builder shipped on July 1 as a no-code way to build phone agents on Grok Voice, and the voice expansion five days later is the natural follow-on: give the new builder a real cast to choose from. If you want the full picture of the platform those voices run on, our companion piece on the Grok Voice Agent Builder covers what it does, what it costs, and which of its claims to read with caution.
Meet the voices: the original five, retrained
The five voices most developers already know are Ara, Eve, Leo, Rex, and Sal. They shipped with the Grok Text to Speech API earlier in 2026 and each carries a distinct tone. Reported descriptions across independent write-ups line up on the broad character of each one:
- Eve reads as warm and energetic, a good default for friendly, upbeat delivery.
- Ara is calm and soft, which suits guided or wellness style interactions.
- Rex is bold and assured, closer to a confident presenter.
- Sal is crisp and neutral, the clean, no-drama choice for information heavy scripts.
- Leo is deep and cinematic, the voice you reach for when you want gravity.
The July 6 update did not retire these voices. It retrained them. xAI frames the change as better pacing, phrasing, and emphasis, which are exactly the qualities that separate a synthetic read from one that sounds like a person who understands the sentence. If you already built on one of the five, you get the improvement without changing your voice selection, though you should still listen to a few of your real scripts again, because a retrained voice can shift emphasis in places your copy was tuned for.
The 21 new voices, cast for a job
The more interesting part of the news is how xAI organized the new arrivals. Rather than dropping 21 generic voices and leaving you to audition all of them, xAI grouped them by the work they are meant to do. The stated categories are support, characters, commentary, advertising, and education.
That framing is a design decision, not just marketing. A support voice needs to sound patient and clear under interruption. A character voice can lean into personality that would be distracting on a billing call. A commentary voice wants energy and momentum. An advertising voice needs a confident, sellable read. An education voice has to stay calm and legible across long passages. Casting voices for these jobs up front means you spend less time coaxing one general voice into five different personalities and more time picking the one that already fits.
A word of caution on the total count. Independent trackers report the full flagship catalog now sits at more than 80 natural voices spanning over 25 languages, with automatic language detection for 20 or more of them. Treat the exact catalog size as a reported figure rather than a settled number, because these totals move as voices are added and the count you see in your own account is the one that governs what you can actually build with today.
Expressive Mode and voice cloning: how Grok voices sound human
Two capabilities do most of the work in making Grok voices sound less synthetic, and both carry over to the new lineup.
The first is Expressive Mode. Grok voices can dynamically control emotional delivery, and you can steer that delivery with inline speech tags for pauses, laughter, whispers, and emphasis. In practice this means you write a script and mark it up where you want the voice to breathe, drop to a whisper, land a beat of laughter, or hit a word harder. For character and advertising work this is the difference between a flat read and a performance. For support and education it is subtler, a well placed pause or a softened phrase that keeps a long explanation from sounding robotic.
The second is custom voice cloning. xAI introduced Custom Voice on April 30, 2026, and it lets you create a cloned voice from roughly a minute of natural speech. The cloned voice inherits the same feature set as the flagship voices, including the inline speech tag system and multilingual output, so a clone is a first class voice rather than a stripped down copy. The obvious caveat is consent: clone a voice you own or have clear permission to use, and keep the provenance clean, because a cloned voice used without rights is a legal and trust problem no feature can fix for you.
Where you can use these voices
The new voices are not locked to one product. xAI exposes them across three surfaces, each aimed at a different builder.
- The Realtime Voice Agent API is for live, back and forth conversational agents where latency and turn taking matter. This is the layer under a real time phone or in-app assistant.
- The Text to Speech API is for generated audio: narration, in-product prompts, audio versions of written content, and any case where you send text and get back speech.
- The Voice Agent Builder is the no-code path. You describe a call flow in plain language, attach documents, tools, and guardrails, and get a working phone agent with the new voices available to choose from.
Beyond xAI's own products, the voices are reaching partner platforms. Telnyx, a communications platform, published a release note making Grok voices selectable inside its Voice AI Assistants, complete with Expressive Mode and inline speech tag control. That partner availability is worth knowing about for two reasons. It confirms the voices are being distributed beyond xAI's first party stack, and it surfaces an honest tradeoff: Telnyx notes that Grok voices with Expressive Mode carry higher latency than its own lowest latency option, which makes them a better fit for expressive, quality first agents than for the most latency sensitive routing. That is a partner observation about their own stack, not an xAI benchmark, so read it as a useful field note rather than a universal verdict.
What Grok Voice costs
Voice on Grok is priced by usage through the API, which is the honest way to think about cost here. There is no separate consumer sticker price for a voice, and you should be wary of any article that quotes one. The official developer rates published by xAI are the numbers to plan against:
- Realtime voice is 0.05 dollars per minute of audio.
- Text to speech is 15.00 dollars per one million characters.
- Speech to text is 0.10 dollars per hour for REST and 0.20 dollars per hour for streaming.
The Voice Agent Builder uses the same 0.05 dollars per minute rate for agent audio, plus a small reported add-on of about 0.01 dollars per minute for telephony on the free numbers it hands out. The realtime, text to speech, and speech to text rates are official developer figures from xAI's documentation; the telephony add-on is reported launch pricing, so confirm it against your own account. Whatever your use case, model the cost in minutes or characters of real usage before you commit, and re-check the live rates at docs.x.ai, because API pricing is the kind of number that moves quietly.
Confirmed facts versus claims to read with caution
For any fast moving Grok story, it helps to draw a clean line between what is confirmed and what is a claim. Here is that line for the July 6 voice update.
Confirmed, from xAI's own announcement and corroborating trackers:
- 21 new flagship voices were added on July 6, 2026, joining the original five.
- The original five (Ara, Eve, Leo, Rex, Sal) were retrained for more natural delivery.
- The voices are multilingual and available across the Realtime Voice Agent API, the Text to Speech API, and the Voice Agent Builder.
- Expressive Mode with inline speech tags, and custom voice cloning from about a minute of audio, apply to the lineup.
Read with caution, because these are vendor or third party figures rather than independently reproduced results:
- The exact size of the full catalog. More than 80 voices across 25 or more languages is a reported total, not a fixed count, and the number that matters is what appears in your account.
- The per voice character descriptions. Tone labels like warm, bold, or cinematic come from xAI and independent reviewers, not from a shared measurement, so audition before you decide.
- Any head-to-head quality claim. xAI has previously cited a self-scored voice benchmark ahead of named competitors. A vendor run test is a starting hypothesis, not a settled result, and your own scripts on your own hardware are the only test that counts for your build.
None of that undercuts the news. The lineup is real and the delivery features are real. The caution is simply about not treating a marketing tally or a self-scored chart as if it were an independent audit.
How to choose a voice for your project
With a bigger cast comes the small burden of choosing well. A practical order of operations:
Start from the job, not the voice. Decide whether you are building support, a character, commentary, an ad, or an education flow, then audition inside that category first. xAI already did the sorting; use it.
Test with your real copy. A voice that sounds great on a demo sentence can stumble on your actual scripts, your product names, and your longest paragraph. Feed it the hardest thing it will have to say.
Use Expressive Mode deliberately. Mark up pauses and emphasis where the meaning needs them, not everywhere. Over-tagging makes a read sound as artificial as no tagging at all.
Mind latency if the agent is live. For a real time phone agent, weigh the expressive quality against response speed, and remember that the most expressive setting is not always the right one for a fast routing call.
Only then think about a custom clone. A cloned voice is powerful for brand consistency, but it adds a consent and rights obligation. Reach for it when a stock voice genuinely cannot carry your brand, and keep the paperwork clean.
The through line of this update is simple. xAI did not just add voices, it organized them by purpose and improved the ones people already relied on. For builders that is the more useful kind of release, because it turns voice selection from a guessing game into a short, well signposted decision. Verify the live pricing and the catalog you actually see in your account, treat the benchmark talk as a claim, and the rest is a genuinely better set of voices to build on.
Questions readers ask
What did xAI announce for Grok Voice on July 6, 2026?
xAI added 21 new flagship voices to Grok Voice and retrained the original five (Ara, Eve, Leo, Rex, and Sal) for more natural pacing, phrasing, and emphasis. The new voices are multilingual and were described as cast for specific jobs such as support, characters, commentary, advertising, and education. They are available across the Realtime Voice Agent API, the Text to Speech API, and the Voice Agent Builder.
What are the original five Grok voices?
Ara, Eve, Leo, Rex, and Sal. They shipped with the Grok Text to Speech API earlier in 2026 with distinct tones, and xAI says it retrained them alongside the July 6 update so their delivery sounds more natural. Reported descriptions put Eve as warm and energetic, Ara as calm and soft, Rex as bold and assured, Sal as crisp and neutral, and Leo as deep and cinematic.
Can I clone my own voice on Grok?
Yes. Grok Voice includes custom voice cloning, which xAI introduced on April 30, 2026, letting you create a cloned voice from roughly a minute of natural speech. The cloned voice inherits the same delivery features as the flagship voices, including the inline speech tags and multilingual output. Only clone a voice you have the right to use.
How much does Grok Voice cost?
Voice is billed by API usage, not by a consumer subscription figure. The official developer rate for realtime voice is 0.05 dollars per minute of audio, text to speech is 15.00 dollars per one million characters, and speech to text is 0.10 dollars per hour for REST and 0.20 dollars per hour for streaming. The Voice Agent Builder uses the same 0.05 dollars per minute rate plus a small reported telephony add-on. Confirm live rates against docs.x.ai before you budget.
Where can I use the new Grok voices?
Three surfaces from xAI: the Realtime Voice Agent API for live conversational agents, the Text to Speech API for generated narration and audio, and the no-code Voice Agent Builder for phone agents. The voices are also being made available through partner platforms such as Telnyx for building voice AI assistants.