xAI Unveils Grok Voice 3.0 Voice Model and Takes the Lead in Voice Agents

On October 14, 2026, xAI introduced Grok Voice 3.0, its speech-to-speech model built for voice agents. The model is available through the Voice Agent API at $0.08 per minute. The company describes Grok Voice 3.0 as the world’s smartest voice model, saying it set a new record for resolving customer support calls and that its latency stays under one second. The price was kept the same as the previous version, Grok Voice 2.0.

First place in independent tests

In its announcement, xAI cites Artificial Analysis’s tests dated October 13, 2026. According to these results, Grok Voice 3.0 ranks first on the Speech-to-Speech Index with 86.9 points. In the comparison table, GPT-Live-1 (Astra, Medium) scores 83.2; Gemini 3.8 Live (Extended Thinking, High) scores 82.6; and Grok Voice 2.0 remains at 81.3 points.

For the index’s subcomponents, xAI shared the following results: 98.5% in Big Bench Audio speech reasoning, an Elo preference score of 1070 in Speech Agent Arena, a task success rate of 94.8%, and 97.0% on Full Duplex Bench, which measures pauses, turn-taking, and interruptions.

On the tau-Voice test, which consists of simulated customer support calls, Grok Voice 3.0 scores 71.6%; xAI describes this result as comparable to GPT-Live-1 (Astra) but with 40% lower latency. The table shows GPT-Live-1 (Astra, Medium) at 74.5%, Gemini 3.8 Live at 68.6%, and Grok Voice 2.0 at 56.5%. In Artificial Analysis’s Big Bench Audio run, the average cost per task is $0.025 for Grok Voice 3.0; $0.030 for GPT-Live-1 (Astra, Medium), and $0.055 for GPT-Realtime-2.1 (High).

How the index is calculated

Artificial Analysis defines the Speech-to-Speech Index as an equally weighted average of four components: speech reasoning, agent performance, arena preference, and task success rate.

Big Bench Audio consists of 1,000 audio questions adapted from Big Bench Hard; models receive the question as audio and respond with audio. In tau-Voice, an audio user simulator calls the model being tested; the model plays the role of a customer support representative with access to domain-specific tools and a policy document. The arena preference component is based on pairwise Elo scores from humans who try the same scenario with two anonymous models. Full Duplex Bench measures pauses, turn-taking, and user interruptions.

Highest accuracy on ServiceNow EVA-Bench

According to xAI, Grok Voice 3.0 achieved the highest accuracy score on ServiceNow’s EVA-Bench, designed for enterprise voice agents, with 74%, and became the best-performing model when accuracy and customer experience are scored together.

Nitin Aggarwal, senior director of product management for Agentic Evals and RAI at ServiceNow, said in a statement in the announcement that Grok Voice 3.0 set a new Pareto frontier on the open-source EVA-Bench, which scores voice agents on accuracy and experience across more than 200 realistic enterprise scenarios.

ServiceNow AI Research describes EVA-Bench as an open-source framework that measures speech-based, multi-turn conversations using a bot-to-bot architecture. The test spans customer service, IT service management, and human resources; agents are scored on two dimensions: task accuracy (EVA-A) and conversational experience (EVA-X).

Daily use at Tesla and Starlink

According to xAI, Grok Voice 3.0 handles millions of conversations per day in Tesla vehicles. Drivers can use Connectors to do things like check their calendar or reply to emails hands-free, and use Grok Bot for larger tasks such as ordering coffee or booking dinner.

In Starlink call centers, Grok Voice answers 20,000 customer calls per day in multiple languages. According to xAI, about 70% of calls are resolved without human intervention, and one quarter of sales calls result in a purchase before the caller hangs up. A single agent built on Grok Voice 3.0 can verify the customer, troubleshoot the hardware issue, issue a credit, and send replacement equipment.

Pricing and access

Grok Voice 3.0 is offered on the Voice Agent API at $0.08 per minute, or $4.80 per hour; this is the same price as Grok Voice 2.0. According to xAI, agents using grok-voice-latest were automatically switched to the new version, and developers do not need to make any additional changes.

Source: Unite.AI