Voice is becoming one of the fastest-moving parts of the X ecosystem, and this week it took another step forward. On October 14, 2026, xAI announced Grok Voice 3.0, a real-time speech-to-speech model built for voice agents, and made it available through its Voice Agent API.
The headline for developers is how little friction the upgrade involves. Applications that use the model alias grok-voice-latest were switched to the new version automatically, according to the release notes. Grok Voice 2.0 is now deprecated, and developers pinned to that older model are directed to move to grok-voice-3.0 or to the latest alias. The underlying API itself is unchanged, so existing connections should keep working without a rebuild.
Pricing also stayed put. Grok Voice 3.0 is listed at the same speech-to-speech rate as its predecessor, $0.08 per minute. For teams running support lines, phone agents, or in-app assistants, holding the price steady while raising performance changes the math on whether a voice agent is worth testing at scale.
On performance, xAI cited independent testing by Artificial Analysis dated October 13. In that comparison, Grok Voice 3.0 scored 86.9 on the Speech-to-Speech Index, up from 81.3 for Grok Voice 2.0 and ahead of the listed scores for rival models in the same table. The company also pointed to a 74 percent accuracy result on a benchmark built around complete, multi-turn spoken support conversations. As with any benchmark, those figures describe specific test conditions, not every live call, but they signal where xAI is aiming: customer support and other task-driven conversations.
The practical toolkit matters as much as the scores. The API guide describes a WebSocket connection for real-time audio, with session settings for the chosen voice, instructions, server-side voice activity detection to spot when a speaker has finished a turn, and tool calling, including web search in the documented example. Developers building for browsers or phones are advised to use short-lived ephemeral tokens, so a private API key does not have to live inside the app.
Choice of voice is part of the package, too. The guide lists 28 built-in voice IDs, and it shows how to create a custom voice from a reference clip of up to 120 seconds, then use that voice ID with the real-time API. That opens a practical path for brands and creators who want a consistent spoken identity without recording every line.
xAI says the technology is already working at scale inside its own orbit, reporting millions of conversations a day in Tesla vehicles and, for Starlink call centers, about 20,000 calls a day with roughly 70 percent resolved without a human. Those are company-reported figures, but they underline the intended audience: high-volume, real conversations, not a lab demo.
For everyday X users, Grok Voice 3.0 is not a new button in the timeline. Its impact is more likely to arrive indirectly, through faster and more capable voice assistants, support agents, and apps built on Grok. For builders, the message is simpler: check which model your agent is pinned to, because if you were on the latest alias, you may already be running the new voice.