Grok Voice 3.0

Grok Voice 3.0 is a real-time speech-to-speech model for developers building voice assistants, phone agents and spoken customer-support workflows.

A conversation that can reach your tools

Grok Voice 3.0 was announced on October 14, 2026. It takes spoken input and returns spoken responses, with text and tool events available to the application. The intended job is an interactive conversation, such as helping a caller explain a problem, retrieving account information and continuing the discussion after a tool returns a result.

This is different from generating a finished voiceover from a script or uploading a recording for transcription. Text-to-speech and speech-to-text have their own endpoints and prices. A developer choosing this model is choosing the conversational part of an application, rather than a complete contact-center system.

What a voice-agent integration involves

• Audio and text move in both directions over the realtime WebSocket endpoint. Session configuration controls instructions, the selected voice and turn detection.
• Server voice-activity detection supports conversational turns and interruptions. The application's audio capture, playback and connection handling still affect the experience.
• Function calls can connect the conversation to business systems. Your application runs client-side functions and returns their results; it remains responsible for access checks and the actions performed.
• The documentation lists server-side search, collections and remote MCP tools alongside custom functions. Configure the tools needed for the particular task rather than assuming the voice model already has access to customer records.

The Speech to Speech guide also documents more than 20 languages, automatic language detection and language hints. A supported language list does not establish equal accuracy for every accent, noisy phone line or specialist vocabulary. Those are useful cases for an application-specific evaluation.

Pricing: the minute rate and the extra line

The official API pricing page lists Grok Voice 3.0 at US$0.08 per minute, equivalent to US$4.80 per hour. It also lists US$0.004 per text input for Speech to Speech. These figures were checked on October 16, 2026. Do not treat the minute rate alone as the entire bill for an application that also uses telephone infrastructure, external tools or other hosted services.

Voiceover and transcription prices should not be substituted for this rate: they describe separate API modes. Likewise, a free-trial button does not establish a permanent free realtime allowance. The current pricing page is the relevant source before estimating a production workload.

Version pinning changes what upgrades automatically

The voice guide says grok-voice-latest now resolves to grok-voice-3.0. Applications using that alias have therefore moved to the new model. Grok Voice 2.0 is deprecated; applications pinned to grok-voice-2.0 are told to select 3.0 or the latest alias. The release documentation says the endpoint, session parameters and event interface remain unchanged, at the same price.

Use the explicit version when evaluating a particular release, and the alias when automatic model updates fit your deployment. An unchanged API contract does not guarantee identical conversational behavior, so replay representative tasks before relying on an upgrade in a customer-facing workflow.

Published results, with their scope attached

The launch announcement cites Artificial Analysis's October 13 evaluation: a Speech-to-Speech Index score of 86.9 and 97% on Full Duplex Bench. Those are named evaluation results reported in the announcement, not performance measurements conducted by Seailife or a guarantee for every live call.

The same announcement describes deployment in Tesla and SpaceX services, including Starlink support. It reports roughly 70% of 20,000 daily Starlink calls resolved without human intervention. That is the vendor's customer-deployment account, not a forecast of the resolution rate another business will achieve.

Start with a bounded caller task

The official console provides a voice playground, and the documentation links tester applications for web, mobile and telephony integrations. For browser or mobile clients, it recommends short-lived ephemeral tokens so a long-lived API key is not exposed in the client. API-key authentication is intended for server-side use.

A practical first evaluation is one caller task with a known successful outcome, such as finding a support article or checking an order through a read-only function. Check interruptions, language switching, tool failures and the handoff to a person before broadening the workflow. Seailife has not placed a call, generated audio or independently tested latency or task completion. The product image is the official announcement artwork, not a tested application screen.