Grok Voice 3.0 arrives in the real-time AI API
xAI launches Grok Voice 3.0, a voice model with responses in less than one second and a price of $0.08 per minute. Artificial Analysis tests place it ahead of several competitors, while it already handles calls for Tesla and Starlink.
xAI has launched Grok Voice 3.0, a voice model that understands and responds to users in real time with latency of less than one second. It is now available through the voice agent API for $0.08 per minute, the same price as the previous version.
The company says the model sets new benchmarks for customer service. In independent tests by Artificial Analysis, Grok Voice 3.0 scored 86.9 on the Speech-to-Speech Index, ahead of Grok Voice 2.0, GPT-Live-1 and Gemini 3.8 Live.
The index combines several capabilities: audio reasoning, conversation understanding and task completion success. In a blind human preference test, the model reached an Elo score of 1070. It also achieved 94.8% task success and 97% on the Full Duplex Bench, a test that measures whether an AI can respect turn-taking, manage pauses and respond to interruptions without sounding artificial.
Less waiting on calls
Grok Voice 3.0 scored 71.6% on τ-Voice, a test focused on support calls with diverse accents, background noise and interruptions. The result is similar to GPT-Live-1, but with 40% lower latency, according to Artificial Analysis.
Latency matters because a phone conversation does not work like a written chat. If the AI takes too long to answer, interrupts the user or waits until they finish long sentences, the call becomes frustrating. A faster response lets the system confirm details, ask questions and correct mistakes at a more natural pace.
On EVA-Bench, ServiceNow's evaluation of enterprise voice agents, the model achieved 74% accuracy and delivered the best combined result when measuring accuracy and customer experience.
Already in use at Tesla and Starlink
According to xAI, Grok Voice already handles millions of conversations every day across Tesla and SpaceX products. In Tesla vehicles, drivers can use connectors to check their calendar or reply to an email without touching the screen. Grok Bot can also handle more complete tasks, such as ordering a coffee or booking dinner.
At Starlink call centers, the system handles 20,000 calls a day in several languages. Approximately 70% are resolved without human intervention, while one in four sales calls ends in a purchase before the customer hangs up.
A single agent can verify the customer's identity, diagnose hardware problems, issue credits and send replacement equipment. That is the practical shift: not just holding a conversation, but completing processes from start to finish.
Users of grok-voice-latest have already been migrated automatically to the new version and do not need to change their integrations. For businesses, the combination of fast responses, lower costs than other comparable models and stable pricing makes it easier to test voice agents in support, sales and operations. The key question now is whether those independent results hold up at scale and in real calls, where noise, accents and unexpected situations tend to test any system.