The short answer: Grok Voice 3.0 is xAI's new speech-to-speech model for building voice agents, released on October 14, 2026. It listens and answers in audio with one model instead of three separate steps, costs $0.08 per minute in xAI's Voice Agent API (the same price as version 2.0), and xAI says it now ranks first on Artificial Analysis's speech-to-speech index with a score of 86.9.
Last updated: October 2026. This is a developer-facing release, and the performance numbers below come mostly from xAI itself.
What Is Grok Voice 3.0?
Most voice assistants still work as a three-stage chain: speech recognition turns your voice into text, a language model writes a text reply, and a text-to-speech engine reads it out. Each hand-off adds delay and can lose tone, hesitation and interruptions.
A speech-to-speech model skips the chain. Grok Voice 3.0 takes in audio, reasons about it, can call tools, and answers in audio, all inside one model. It is sold through xAI's API under the model name grok-voice-3.0, and it streams audio and text in both directions over a WebSocket connection, which is what lets a reply start in under a second according to xAI.
What Changed From Grok Voice 2.0
| Grok Voice 2.0 | Grok Voice 3.0 | |
|---|---|---|
| Released | July 29, 2026 | October 14, 2026 |
| Status | Deprecated | Current flagship |
| Price | $0.08 per minute | $0.08 per minute |
| Model alias | Replaced | grok-voice-latest now points here |
| API changes | None | None, per xAI's release notes |
The upgrade is designed to be painless. Agents already using grok-voice-latest were switched over automatically, with no code changes. If you pinned an agent to the 2.0 model name, you will need to move it. The two versions shipped about eleven weeks apart, so builders should expect fast model turnover in this category and pin a specific version when they need stable behavior.
Pricing
| Item | Price |
|---|---|
| Grok Voice 3.0 (speech-to-speech) | $0.08 per minute ($4.80 per hour), plus $0.004 per text input |
| Speech-to-text (separate API) | $0.10 per hour (REST) or $0.20 per hour (streaming) |
| Text-to-speech (separate API) | $15 per 1M characters |
| OpenAI GPT-Live-1, for comparison | $0.05 per minute, voice only; your backend model is billed on top |
Compare these carefully. The $0.08 per-minute rate is for the full speech-to-speech model, while the OpenAI figure covers the voice layer and your reasoning model is billed separately, so the two numbers are not directly comparable. In practice a real support call also pays for the phone carrier, any tools the agent calls and silent time on the line, so measure cost per resolved call, not only cost per minute.
What xAI Claims, and What Is Verified
| Claim (as reported by xAI) | Result |
|---|---|
| Artificial Analysis Speech-to-Speech Index | 86.9, ranked first |
| Big Bench Audio reasoning test | 98.5%, against 90.1% for GPT-Live-1 |
| Average cost per task on that test | $0.025, against $0.030 for GPT-Live-1 |
| Response latency | Under 1 second |
| Starlink support calls | About 20,000 a day, roughly 70% resolved without a human |
Treat these as xAI's numbers. The company's release notes confirm the availability, the unchanged price and the automatic migration, but in the coverage we reviewed, the benchmark and call-resolution figures were not independently confirmed. Index rankings in a category this new also move quickly, so re-check them before you build on a lead.
What You Can Build With It
- Customer support lines that answer and resolve routine calls, which is the use xAI emphasizes most.
- Phone agents that take calls over standard telephony, with developer guides describing support for call transfers and keypad input.
- In-app voice assistants that talk with users instead of showing a text box.
- Tool-using agents that look things up or trigger actions mid-conversation, since the model can call tools while it speaks.
Who Should Care, and Who Can Skip It
- Care if you build voice products, run a support line, or are choosing between voice APIs from xAI, OpenAI and Google.
- Care if you already use Grok Voice 2.0, because migration is now part of your roadmap.
- Skip it if you only need narration, voiceovers or cloned voices. Text-to-speech tools are built for that, and our guide to the best AI voice and audio generators covers them.
How to Try It
- Create an xAI API key and read the Voice Agent API documentation.
- Connect to the real-time WebSocket endpoint using
grok-voice-latestto follow the newest model. - Once your agent works, pin
grok-voice-3.0so future upgrades happen on your schedule. - Test with recordings of your real calls, including interruptions and background noise.
- Track cost per resolved call, not just the per-minute rate.
Our Take
The interesting part of this launch is not the benchmark crown. It is that a better model arrived at an unchanged price, which keeps pushing the cost of an automated call down while the shift from stitched pipelines to single speech-to-speech models continues. The risks are practical: version churn every few months, and benchmark claims that come from the vendor. Pilot it on a narrow, measurable job first, then expand.
Quick FAQ
What is Grok Voice 3.0?
xAI's speech-to-speech model for real-time voice agents, available in its Voice Agent API since October 14, 2026.
How much does Grok Voice 3.0 cost?
$0.08 per minute ($4.80 per hour), plus $0.004 per text input, the same price as Grok Voice 2.0.
What is a speech-to-speech model?
A single model that takes audio in and produces audio out, instead of chaining speech recognition, a language model and text-to-speech.
Is Grok Voice 2.0 still available?
It is deprecated. Agents using the grok-voice-latest alias moved to 3.0 automatically, and pinned 2.0 agents should migrate.
Is Grok Voice 3.0 better than OpenAI's voice model?
On xAI's reported results, it beats GPT-Live-1 on the Artificial Analysis index and the Big Bench Audio test. Independent confirmation is limited so far, so test it on your own calls.
Sources: xAI's October 14, 2026 announcement and API release notes, as reported by Unite.AI and Blockchain.News. More voice and audio tools are listed in Vinggle's AI Audio Generators directory. If you are building one, submit it to Vinggle.