xAI has introduced Grok Voice 3.0, a real-time speech-to-speech AI model designed to make voice assistants and customer service agents more responsive, capable, and natural to interact with. Already powering millions of daily conversations in Tesla vehicles and handling customer support calls for Starlink, the new model is now available to developers through xAI’s Voice Agent API.

Announced on October 14, Grok Voice 3.0 builds on its predecessor with improved benchmark performance, sub-second latency, and the ability to handle conversational interruptions and turn-taking. xAI is keeping the API price unchanged at $0.08 per minute, making the upgrade available without an increase in the model’s listed usage rate.

Grok Voice 3.0 aims to make AI conversations more natural

Unlike systems that separately convert speech into text, process a request, and then generate spoken output, a speech-to-speech model is designed to handle audio input and audio output directly. This approach can help voice agents respond more naturally, particularly when conversations involve interruptions, pauses, different accents, or background noise.

According to testing by AI benchmarking firm Artificial Analysis, Grok Voice 3.0 scored 86.9 on its Speech-to-Speech Index, placing it ahead of the previous Grok Voice 2.0 model and the competing systems included in xAI’s comparison.

The benchmark results published by xAI show the following scores:

• Grok Voice 3.0: 86.9
• GPT-Live-1 (Astra, Medium): 83.2
• Gemini 3.8 Live (Extended Thinking, High): 82.6
• Grok Voice 2.0: 81.3

Grok Voice 3.0 also achieved a 97% score on Full Duplex Bench, which evaluates conversational dynamics such as interruptions, pauses, and turn-taking. It scored 98.5% on Speech Reasoning Big Bench Audio and recorded a 94.8% task completion success rate in the reported tests.

On τ-Voice, a benchmark focused on customer support conversations involving varied accents, background noise, and interruptions, the model scored 71.6%. xAI says it achieved comparable performance to GPT-Live-1 (Astra) while delivering 40% lower latency on that test.

These results suggest that xAI is targeting more than speech generation alone. For voice agents, the ability to understand a request, respond quickly, and manage a conversation without awkward pauses can be just as important as the accuracy of the underlying AI.

The benchmark numbers come from the results cited in xAI’s announcement, rather than independent testing conducted for this article.

Tesla and Starlink already use Grok Voice at scale

Grok Voice 3.0 is not just a developer-facing model. It is already being used in two of the company’s most prominent real-world deployments: Tesla’s in-car voice experiences and Starlink’s customer support operations.

In Tesla vehicles, Grok Voice reportedly handles millions of conversations every day. Drivers can use voice interactions for everyday requests, including checking their calendars and answering emails through Connectors. More involved tasks, such as ordering coffee or booking dinner, can be handled through Grok Bot, according to xAI. The model also powers Grok Bot’s voice mode.

The integration highlights the potential value of faster speech models in cars, where drivers need to interact with an assistant without taking their hands off the wheel or their attention away from the road for extended periods. The actual availability of individual features can depend on the relevant Tesla software and service support.

Starlink is using Grok Voice for a different purpose: customer service.

According to xAI, Starlink’s call centers handle approximately 20,000 customer calls each day using Grok Voice, with support available in multiple languages. The company says roughly 70% of those calls are resolved without human intervention.

The AI agent can verify customers, troubleshoot hardware problems, issue credits, and arrange replacement equipment. xAI also claims that one in four sales calls ends in a purchase before the conversation concludes.

Those figures illustrate why businesses are investing in conversational AI. A voice agent that can do more than answer basic questions has the potential to handle complete support workflows, reducing the need to transfer customers between automated systems and human representatives. However, the reported resolution and sales figures are company-provided metrics, not independently verified results.

Grok Voice 3.0 pricing and availability

Developers can use Grok Voice 3.0 through the Voice Agent API for $0.08 per minute, or $4.80 per hour. The price is unchanged from Grok Voice 2.0, according to xAI.

Existing applications using the grok-voice-latest model alias have already been switched to Grok Voice 3.0 automatically. Developers using that alias do not need to change their integrations to receive the new model.

The previous grok-voice-2.0 model identifier has been deprecated, with xAI directing developers to use grok-voice-3.0 or grok-voice-latest instead.

The company has made the model available through its developer API, with documentation and an option to try it available through the official Grok Voice 3.0 announcement.

For developers building customer support systems, in-car assistants, and other real-time voice applications, the combination of improved benchmark results and unchanged per-minute pricing could make the upgrade appealing. The bigger test, however, will be how consistently the model performs across real conversations, particularly when users interrupt, change topics, or need help with tasks that go beyond simple questions.