xAI Voice Agent: a guide to Grok Voice Agent Builder

xAI launched Grok Voice Agent Builder in beta in July 2026, and it moved the voice AI conversation fast. Instead of assembling a stack, you describe the call in plain language and get a working agent in minutes. This guide explains what the grok voice agent is, how its architecture differs, what it costs, and what still needs testing. The facts here come from xAI's own pages, so confirm current details before you build.

Evalgent is platform-agnostic, so we cover this new entrant honestly, both its strengths and its gaps. First, what it is.

What is the xAI Voice Agent?

xAI Voice Agent: Grok Voice Agent Builder, a no-code platform that compiles a plain-language description of a call into a live voice agent running on xAI's Grok Voice model.

The grok voice agent is built for operators who want production phone agents without engineering the surrounding stack. You write a playbook in plain language, and the builder turns it into a working agent. Out of the box you get telephony, knowledge retrieval, tools, guardrails, MCP support, and observability in one place, per the xAI Voice page.

That bundling is the pitch. Most platforms make you wire the parts together yourself. You pick an STT provider, an LLM, a voice, and telephony, then glue them. Grok Voice skips that assembly. The grok voice agent builder ships them assembled, which is why it can go live in about two minutes.

How does the Grok Voice agent work?

Here is the architectural difference that matters. Most voice agents run three stitched-together APIs: speech-to-text, then an LLM, then text-to-speech. The Grok Voice agent runs on a single speech-to-speech model instead.

That unified design is where its sub-second latency comes from. There is no hand-off between three services on every turn. Audio goes in and audio comes out through one model, Grok Voice. It reached the top of the audio benchmarks xAI cites, ahead of comparable realtime models. On xAI's leaderboard, Grok Voice scores well above Gemini and GPT realtime models. Benchmarks are a starting signal, not proof for your calls, but the architecture is clearly built for speed. Single speech-to-speech is the same architecture we cover in our full-duplex voice agents guide, and it is increasingly common for latency-critical use cases.

How much does Grok Voice cost?

Pricing is one of the launch's headline points. The xAI speech-to-speech agent is billed at the API rate, currently $0.05 per minute of audio, with voices included and no separate platform fee.

Telephony on a free provisioned number adds about $0.01 per minute. So a rough all-in grok voice pricing figure is around $0.06 per minute. A ten-minute support call costs roughly $0.60 in total. That is one predictable number, not a stack of vendor invoices. That is cheaper and more predictable than a build-your-own stack, where costs come from several vendors. Our AI voice agent cost guide puts this in context against Vapi, Retell, and self-hosting. Confirm the current rate on the xAI pricing page, since beta pricing can change.

What features does it include?

The builder bundles what most platforms make you assemble. The feature set is broad for a launch.

• No-code builder: describe the call in plain language, launch in minutes.
• 80+ voices and cloning: built-in voices, or clone a brand voice from two minutes of audio.
• 25+ languages: with mid-conversation language switching.
• Telephony and SIP: a free phone number, or bring your own over SIP.
• Tools and MCP: wire your APIs and MCP servers for mid-call actions.
• Guardrails: define what the agent will and will not do, such as no PII.
• Observability: every call is recorded, transcribed, and replayable.

Enterprise coverage is there too: SOC 2, HIPAA eligible, and GDPR compliant, per the launch announcement. For developers, the xAI docs cover the API and WebSocket client.

Where the single-model approach helps and hurts

A single speech-to-speech model is a real advantage for latency and simplicity. Fewer moving parts means fewer hand-offs and a faster, more natural feel. For many phone use cases that is exactly right.

It also has trade-offs. You get less control over each stage than a pipeline gives. You cannot swap the LLM for a cheaper or specialised one, because the model is unified. And a newer architecture has a shorter production track record than the mature STT-LLM-TTS pipeline. For a breakdown of the pipeline alternative, see our voice agent stack guide, and for model trade-offs, best LLM for voice agents.

Who should consider it

The Grok Voice agent fits teams that want a fast, low-latency phone agent without building infrastructure. If you value speed to launch, predictable per-minute pricing, and a bundled stack, it is a strong option, especially for high-volume support, sales, and scheduling calls.

It fits less well if you need to pick your own LLM, run a heavily customised pipeline, or require a long production track record before you commit. Those teams may prefer a build-your-own platform or a more established provider. As with any platform, the right choice depends on your calls and your constraints, not on the launch buzz around a new release.

Setting one up in practice

Setup is genuinely fast. You open the xai voice agent builder in the browser. You describe the call in plain language. The builder drafts a playbook. You edit it, preview the agent live, and launch.

The knowledge base is simple too. You drop in PDFs, Markdown, or plain text. The agent uses them to answer questions on the call. Voice cloning takes about two minutes of audio. Your brand voice then joins the 80-plus built-in ones.

Tools wire in through your APIs or MCP servers. The agent can look up a record, book a slot, or transfer to a human mid-call. Guardrails keep it in bounds. You define what it will not do, such as reading back a card number.

None of this needs code. That is the point. But it also means the hard parts are hidden. The model, the turn-taking, and the latency are handled for you. You trade control for speed.

How it fits the market

The launch matters beyond one product. It signals where voice AI is heading. Single speech-to-speech models are getting good enough to bundle. Prices are falling toward a few cents a minute. Setup is collapsing from weeks to minutes.

That is good for buyers. It lowers the barrier to a working phone agent. It also raises a question. If anyone can ship an agent in two minutes, quality becomes the differentiator, not the build. The teams that win will not be the ones that launched fastest. They will be the ones whose agents actually handle real calls.

So treat the ease as a starting line, not a finish. The builder gives you an agent. Your callers decide if it is a good one. That verdict comes from testing, not from the demo.

Do you need to test a Grok Voice agent?

Yes, and arguably more than with a mature stack. A no-code builder makes an agent easy to ship, but easy to ship is not the same as ready for real callers. Accents, interruptions, noise, and edge cases break agents on every platform, and a brand-new speech-to-speech model has less production history behind it.

This is where Evalgent fits. Evalgent runs realistic calls against your Grok Voice agent before real callers do. Scenarios reproduce noisy, accented, and off-script calls. Profiles vary caller behaviour so results split per cohort. Metrics score task completion, latency, and adherence with thresholds. Evaluations run this as automated batches of synthetic callers, and Reviews let you hear where the agent struggled. The builder gets you live fast; testing tells you whether it is actually ready. See the ai voice agent testing pillar for the full method.

Frequently asked questions

What is the xAI Voice Agent?

The Grok Voice agent is Grok Voice Agent Builder, a no-code platform from xAI that turns a plain-language description of a phone call into a live voice agent in about two minutes. It runs on a single speech-to-speech model, Grok Voice, and bundles telephony, tools, guardrails, MCP support, and observability. It launched in beta in July 2026 for high-volume production voice agents.

How much does Grok Voice cost?

Grok Voice is billed at the API rate, currently $0.05 per minute of audio, with voices included and no separate platform fee. Telephony on a free provisioned number adds about $0.01 per minute, so an all-in cost of around $0.06 per minute is typical. A ten-minute support call costs roughly $0.60. Confirm current rates on the xAI pricing page, since beta pricing changes.

Is the Grok Voice agent no-code?

Yes. The Grok Voice agent builder is no-code: you describe how the call should go in plain language, and the platform compiles it into a live agent in about two minutes. Developers can also go deeper, wiring tools to their APIs and MCP servers or connecting a custom client over WebSocket, but no code is required to build and launch a working agent.

How does the xAI Voice Agent work?

The Grok Voice agent runs on a single speech-to-speech model rather than the usual three stitched-together APIs of speech-to-text, an LLM, and text-to-speech. Audio goes in and audio comes out through one model, Grok Voice, which is where its sub-second latency comes from. The builder wraps that model with telephony, tools, guardrails, and observability in one platform.

What languages does Grok Voice support?

Grok Voice supports 25 or more languages, with the ability to switch language mid-conversation without manual configuration. It offers more than 80 built-in voices and can clone a brand voice from about two minutes of audio. This makes it suitable for multilingual support and sales calls, though the exact language list should be confirmed on xAI's current documentation.

Does Grok Voice support telephony and SIP?

Yes. Each Grok Voice account includes a free provisioned phone number, ready for test calls or production traffic. You can also bring your own number over direct SIP, which connects an existing number from any major telephony provider. Telephony on the free number adds about $0.01 per minute on top of the per-minute audio rate.

How does Grok Voice compare to ElevenLabs?

Grok Voice runs a single speech-to-speech model with bundled all-in pricing near $0.05 to $0.06 per minute, while ElevenLabs uses a cascading pipeline with best-in-class text-to-speech and your choice of LLM, priced around $0.08 per minute plus the LLM. Grok Voice favours latency and simplicity; ElevenLabs favours voice quality and LLM flexibility. The right pick depends on your priorities.

Do you need to test a Grok Voice agent?

Yes. A no-code builder makes an agent easy to ship, but not automatically ready for real callers. Accents, interruptions, noise, and edge cases break agents on every platform, and a new speech-to-speech model has a short production track record. Platform-agnostic testing with synthetic callers, such as Evalgent, verifies the agent under real conditions before launch.

Conclusion

The Grok Voice agent is a genuinely new take: a single speech-to-speech model, wrapped in a no-code builder, at bundled pricing near $0.06 per minute all-in. It trades some control for speed, simplicity, and low latency. That suits many phone use cases very well.

New and easy to ship does not mean ready for production. A two-minute build is a starting point, not a finish line. Whatever platform you build on, test the agent against real accents, noise, and edge cases before callers do, because the launch demo is never the production call.