SpaceXSI released Grok Voice 3.0 on Oct. 14, a speech-to-speech model for building voice agents that answer calls, take actions and talk back in under a second. It is live now in the Voice Agent API at $0.08 a minute ($4.80 an hour), the same price as Grok Voice 2.0.
Two of its largest use cases, Grok Bot Voice Mode and Grok in Tesla, already run on 3.0. Developers already pointing at grok-voice-latest were moved to 3.0 automatically, with no code changes.
What You Can Build With It
Grok Voice 3.0 runs over the Voice Agent API, a real-time WebSocket connection that handles turn-taking and tool calls. An agent can listen, look something up or trigger an action, and answer out loud in the same conversation. Client apps can connect with ephemeral tokens instead of exposing an API key, and SpaceXSI publishes demo apps for web, WebRTC, iOS and Twilio phone agents.
Every built-in voice speaks every supported language, with Eve as the default. You can also clone a custom voice from a reference clip of up to 120 seconds and use it in the speech-to-speech API. SpaceXSI says voice audio is processed in real time and never stored or used for training.
How It Ranks
Artificial Analysis says Grok Voice 3.0 debuts at No. 1 on its Speech to Speech Index with 86.9, ahead of the nearest rival model at 83.2. That is up 5.6 points on Grok Voice 2.0 and 13.2 points on Grok Voice 1.0. The index averages speech reasoning, a customer-support test called τ-Voice, human preference and task success.
SpaceXAI’s Grok Voice 3.0 is the new leading Speech to Speech model, debuting at #1 on the Artificial Analysis Speech to Speech Index at 86.9, while also being competitive on latency
Grok Voice 3.0 is @spacexai 's successor to Grok Voice 2.0, a Speech to Speech model available pic.x.com/eVtXVwEYGs
— Artificial Analysis (@ArtificialAnlys) October 14, 2026It does not top every test inside that index by itself, but that isn’t surprising. In Artificial Analysis’s Speech Agent Arena, where people pick between two unnamed agents, Grok Voice 3.0 ranks third on preference at Elo 1070, behind two rival models at 1100 and 1088.
It ranks first on task success at 94.8%, just ahead of Grok Voice 2.0. On τ-Voice it places second at 71.6%, behind a rival model at 74.5% and up from 56.5% for Grok Voice 2.0. It scores 98.5% on Big Bench Audio reasoning and 97.0% on Full Duplex Bench, which tests how it handles pauses, interruptions and turn-taking; both place it fourth on Artificial Analysis’s leaderboard.
Artificial Analysis measured an average time to first audio of 0.81 seconds, faster than the leading rival models but slower than Grok Voice 2.0 at 0.70 seconds. Its cost of $4.80 per hour of input audio is unchanged from Grok Voice 2.0 and sits mid-pack among the rival models Artificial Analysis compared. SpaceXSI also says Grok Voice 3.0 posted the highest accuracy, 74%, on ServiceNow’s EVA-Bench for enterprise voice agents.
Grok Voice 3.0 sets a new Pareto frontier on EVA-Bench, ServiceNow’s open-source benchmark scoring voice agents on accuracy and experience across 200+ realistic enterprise scenarios.
Nitin Aggarwal, Senior Director of Product Management, Agentic Evals and RAI, ServiceNow
Already Answering Starlink Calls
SpaceXSI says Starlink call centres use Grok Voice to answer about 20,000 customer calls a day in multiple languages. By SpaceXSI’s count, roughly 70% are resolved with no human involved, and one in four sales calls ends in a purchase before the caller hangs up. A single agent verifies the customer, troubleshoots hardware, issues credits and ships replacement equipment.
SpaceXSI says Grok Voice handles millions of conversations a day across Tesla and SpaceX. In Tesla cars, drivers can use Connectors for quick jobs such as checking a calendar or answering an email, and hand bigger ones, like ordering a coffee or booking dinner, to Grok Bot.
Where You’ll Hear It First
For most people, the change shows up in Grok Bot’s voice mode, which arrived alongside voice notes, and in Grok in Tesla vehicles. SpaceXSI hasn’t confirmed whether the standalone app and web Grok have moved over to 3.0 as well, but we expect that it has.
