Grok Voice 3.0 Speech Model: Features and Pricing

Grok Voice 3.0 is presented by xAI as a speech-to-speech model for real-time voice agents, and xAI’s announcement says it became available in the Voice Agent API on October 14, 2026. The reported price is $0.08 per minute ($4.80 per hour), the same as Grok Voice 2.0; the announcement does not establish a free tier.

What Grok Voice 3.0 does

Speech-to-speech systems take spoken input and produce spoken responses, supporting live conversations rather than a workflow built around separately transcribing and voicing every turn. xAI positions Grok Voice 3.0 for voice agents, including customer support.

The model’s reported performance claims come with important context: they are figures cited by xAI from Artificial Analysis testing, not an independent evaluation in this article. xAI says the model scored 86.9 on the Speech-to-Speech Index, with a 97.0% score on Full Duplex Bench, which evaluates pauses, turn-taking, and interruptions.

Results vary by test. On τ-Voice, a support-call benchmark involving accents, background noise, and interruptions, Grok Voice 3.0 scored 71.6%. xAI’s comparison lists GPT-Live-1 at 74.5% on that test, while claiming Grok Voice 3.0 had about 40% lower latency. The overall index and an individual support-call score measure different things, so neither tells the whole story for a particular use case.

Is it an official model?

Yes. xAI published an announcement titled “Introducing Grok Voice 3.0” on its official site, describing the model and its Voice Agent API availability. That is stronger evidence of the name and launch than reposts, third-party tutorials, or search snippets.

The announcement is about a voice-agent model and API. It does not, by itself, confirm that every Grok app, device, or account has a user-facing setting called “Grok Voice 3.0.” If you want to use it, check xAI’s current API documentation and account console for access and supported features.

Availability and price

xAI’s announcement says Grok Voice 3.0 is available through its Voice Agent API at $0.08 per minute, or $4.80 per hour. It says this is the same price as Grok Voice 2.0. Actual charges depend on usage; the cited announcement does not describe a free allowance.

The announcement also says agents using the grok-voice-latest alias were switched automatically. An alias can point to a newer model over time, so developers who need stable behavior should check the current documentation and consider whether their integration should specify a versioned model name.

How to access it

For a developer, the supported starting point is the xAI Voice Agent API, not a third-party tutorial claiming to offer access.

1. Check the official API docs and console. Confirm that your account can create an API key and that the current documentation lists the model.
2. Review the connection instructions. The launch announcement confirms API availability, but it is not a complete integration guide. Use xAI’s current documentation for the endpoint, authentication, event formats, and audio requirements.
3. Build a small test agent. Start with a narrow task, such as answering a few common support questions, before connecting business systems or routing real calls.
4. Test real conversation conditions. Try interruptions, background noise, varied speaking styles, and requests that require a handoff. Measure response time, successful task completion, and whether the agent follows your instructions.
5. Estimate costs from expected usage. At the announced rate, multiply expected minutes by $0.08. Confirm billing details and any other charges in your account before production use.

A third-party tutorial in the search results describes a WebSocket workflow, but its examples use a non-xAI API gateway. Treat those connection details as specific to that service, not as official xAI setup instructions.

What the benchmark claims do and do not show

xAI cites Artificial Analysis results dated October 13, 2026. Its announcement reports Grok Voice 3.0 at 86.9 on the overall Speech-to-Speech Index, compared with 83.2 for GPT-Live-1 and 82.6 for Gemini 3.8 Live in the listed comparison. It also reports a 97.0% Full Duplex Bench score.

These are benchmark results, not a guarantee that the model will outperform alternatives on your calls. A support agent may need to identify a customer, follow a policy, retrieve account information, or transfer a conversation correctly. Test those exact workflows with your own prompts, tools, and failure cases.

The announcement also cites a 74% score on ServiceNow’s EVA-Bench. That figure is attributed to xAI’s account of the benchmark; verify the benchmark materials directly if that result matters to a purchasing or deployment decision.

FAQ

When was Grok Voice 3.0 released?

xAI’s announcement is dated October 14, 2026, and says the model is available in the Voice Agent API. The launch information in the research does not establish a separate date for availability in consumer apps.

Is Grok Voice 3.0 free?

The announcement lists API pricing of $0.08 per minute ($4.80 per hour). It does not confirm a free tier, so check the current xAI console for any trial credits or account-specific terms.

Does Grok Voice 3.0 have an API?

Yes. xAI says it is available in the Voice Agent API. Consult the official documentation for current setup instructions, model identifiers, and supported controls.

What languages does it support?

The cited announcement does not give a complete supported-language list. It says Grok Voice is used for Starlink calls across multiple languages, but that company claim is not a full API language specification. Check the current docs or test the languages your agent needs.

How should I test it before using it for customer calls?

Test representative calls end to end: common questions, noisy audio, interruptions, tool use, policy-sensitive requests, and transfers to a human. Compare results against your current system using task completion, response timing, and failure handling, rather than relying only on the headline benchmark score.