With Grok Voice 3.0, xAI wants to turn the phone into a playground for its agents

Grok Voice 3.0 tops the Artificial Analysis voice leaderboard. xAI announces more reliable, fast, and cost-effective conversations for enterprises.

Answering a customer, understanding an accent, handling an interruption, and resolving an issue without transferring the call to an agent. These are the capabilities where Grok Voice 3.0 aims to stand out. xAI’s new voice model takes first place on Artificial Analysis’s Speech-to-Speech Index, ahead of compared solutions from OpenAI and Google. Already used at scale at Tesla and Starlink, it specifically targets companies looking to automate their phone conversations. A score of 86.9 against OpenAI and Google

Grok Voice 3.0 scores 86.9 points on Artificial Analysis’s Speech-to-Speech Index, compared to 83.2 for OpenAI’s GPT-Live-1 in its Astra Medium configuration, and 82.6 for Google’s Gemini 3.8 Live with extended reasoning.

This independent evaluation combines several criteria: spoken comprehension and reasoning, task completion capability, human preferences in blind comparative conversations, and agent performance in complex exchanges.

On Big Bench Audio, dedicated to voice reasoning, the model reaches 98.5%. It also scores 97% on Full Duplex Bench, which notably measures the handling of pauses, interruptions, and turn-taking changes.

The results show progress compared to Grok Voice 2.0, previously credited with 81.3 points on the overall index. Fast conversations, even with interruptions

Responsiveness remains an important criterion for phone agents. A response that is too slow disrupts the flow of the exchange and quickly gives the impression of talking to an automated system.

Grok Voice 3.0 shows a delay of about 0.81 seconds before starting its voice response, compared to 1.34 seconds for GPT-Live-1 in the compared configuration. However, it remains slightly slower on this criterion than Grok Voice 2.0, which measured at 0.70 seconds.

On τ-Voice, a test reproducing support conversations with different accents, noise, and interruptions, the model reaches a 71.6% success rate. GPT-Live-1 maintains a lead here at 74.5%, but responds more slowly.

The new generation thus improves on the overall quality of exchanges without dominating every individual indicator. Starlink automates 70% of its calls

Beyond evaluations, xAI highlights the use of its technology across the group's services.

At Starlink, call centers reportedly handle approximately 20,000 conversations per day with Grok Voice. According to the company, nearly 70% of requests are resolved without human intervention. The agent can verify a customer's identity, diagnose a hardware issue, grant a commercial credit, or arrange for replacement equipment to be shipped.

xAI also claims that one out of four sales calls leads to a purchase before the conversation ends.

In Tesla vehicles, Grok Voice 3.0 is said to participate in several million daily exchanges. Drivers can check their calendar, draft an email response, or delegate certain tasks to Grok Bot—including ordering coffee or making a restaurant reservation—hands-free.

These volumes and resolution rates are reported by xAI. They do not correspond to independent audits of Tesla or Starlink operations. An API at 8 cents per minute

Grok Voice 3.0 is available via the Voice Agent API, priced at $0.08 per minute, or $4.80 per hour of conversation, with no price increase from the previous version.

Developers already using the `grok-voice-latest` identifier automatically switch to the new generation without changing their integration.

On Big Bench Audio, Artificial Analysis estimates the average cost of a task at $0.025 with Grok Voice 3.0, compared to $0.030 for GPT-Live-1 and $0.055 for GPT-Realtime-2.1 in the tested configurations. Voice becomes an execution interface

With Grok Voice 3.0, xAI positions conversation as a way to accomplish actions, not just to get answers. The value proposition for businesses relies on combining three elements: reliable understanding, responsiveness close to human dialogue, and the ability to complete a request without systematic operator intervention.

This automation nevertheless raises questions about the control of authorized actions, conversation privacy, and the management of complex or sensitive situations. Results obtained in test scenarios are not enough to guarantee consistent quality with every customer.

However, deployment at Tesla and Starlink gives xAI a significant testing ground. Competition among voice assistants is now fought as much on their ability to conduct a natural conversation as on their capacity to effectively resolve the requests entrusted to them.