Grok Voice Gets 21 New Voices. The Price Is the Point
xAI pushed a big Grok Voice update over the weekend. Twenty-one new multilingual voices, voice cloning from a short reference clip, inline speech tags for pacing control, and a no-code agent builder that launched a few days earlier. All of it at $0.05 per minute of audio.
That price is the part I keep coming back to.
What they shipped
The headline number is 21 new voices joining the original five (Ara, Eve, Leo, Rex, Sal). Each one is natively multilingual across 25+ languages, so a single voice can switch between English, Hindi, Spanish, or Japanese mid-conversation without dropping character. The originals got retrained too, with better pacing and phrasing.
Voice cloning works from about a minute of reference audio. Upload a clip to the xAI console, get a voice ID back, and use it anywhere a built-in voice is accepted across the TTS API, the Voice Agent API, and the agent builder. No per-clone markup.
The speech tags are worth a separate mention. Inline directives like [pause] and <whisper> let a script control delivery at the word level without swapping voices. That is the kind of control that turns a single narrator into a product with range.
$0.05 per minute
That rate covers everything. Built-in voices, cloned voices, TTS, the realtime voice agent API, the no-code builder. There is no premium tier for custom voices. You pay five cents per minute of generated audio at any volume tier.
Compare that to what else is on the market. ElevenLabs’ API charges $50 per million characters on Flash ($0.0375/min at average speaking rate) and $100 on Multilingual v2/v3. Google Chirp 3 HD lists at $30 per million characters ($0.0225/min). But those are TTS-only prices for a single pipeline step. Grok Voice at $0.05/min includes the entire chain from speech in to speech out across APIs and the no-code builder, so the comparison only makes sense as a total cost of a voice interaction, not isolated TTS.
I expected the agent builder to carry a surcharge. It does not.
What the agent builder actually does
The Grok Voice Agent Builder launched July 1. You write a plain-language description of a call flow, attach a knowledge base and tools, and it produces a working voice agent that speaks through one of the 26 voices. No ASR model to configure, no TTS model to wire in, no VAD thresholds to tune. One speech-to-speech model end to end.
Priced at the same $0.05 per minute.
The builder is the part that should worry competitors more than the voices do, because voices are a commodity problem now and the builder is a distribution problem. Anyone can train a good voice model. Building the infrastructure that collapses a three-API stack into one call and puts a no-code UI in front of it takes a different kind of investment, and getting it to $0.05 on day one means they are playing a volume game, not a margin game.
Where this lands in the market
The voice AI pricing map for mid-2026 is getting interesting at the low end. Top-quality TTS from the market leaders runs $10 to $100 per million characters. Grok Voice at $0.05 per minute is priced differently (per minute, not per character) because it is a speech-to-speech pipeline, not isolated TTS. A direct per-character comparison would be misleading. What matters is that $0.05 gets you the full interaction: the model listens, thinks, and speaks back without three separate API bills.
The catch is that it is one model doing everything. You give it text, it generates speech with its own prosody and phrasing. If you need exact verbatim rendering of a pre-written script with no interpretation (think audiobooks, IVR prompts, dubbing where the translated text is locked), a dedicated TTS model that reads what you wrote is still the right tool. Grok Voice is built for the conversational case, not the narration case.
But for voice agents, customer support, and any product where the user talks to an AI and the AI talks back in real time, $0.05 per minute with voice cloning included is going to reset the floor. Every competitor’s pricing conversation now starts with “well, we offer more control over X” rather than “we cost less.”
One thing to watch
xAI did not announce any watermarking, consent verification, or safety framework for the voice cloning. ElevenLabs has its own cloning with consent verification baked in, and OpenAI’s GPT-Live models (which launch tomorrow, from the look of the press invites) will almost certainly ship with SynthID or equivalent. xAI cloning from one minute of audio with no disclosed safety guardrails is the part of this launch that is going to get regulatory attention, especially under the EU AI Act.
I work on the SpeechifyAI TTS API, so I am biased toward models that render your text verbatim and cost per character. But bias acknowledged: this Grok Voice update is the most interesting voice AI pricing move of the year so far, and it happened on a Sunday with no press embargo. The industry noticed.