Loading page…
Developers using the latest-model setting have already switched. xAI’s performance case pairs faster responses with customer-support scores that do not lead every comparison.
Listen to this story
Existing developers face a small migration decision rather than an API redesign: sessions using grok-voice-latest moved automatically, while integrations pinned to version 2.0 must update the model setting; the endpoint and events are unchanged. The evidence positions 3.0 as a latency advantage, not an across-the-board accuracy leader: xAI reports 71.6% on τ-Voice versus 74.5% for GPT-Live-1, while claiming 40% lower latency. xAI also says the model handles 20,000 Starlink calls daily, with about 70% resolved without human help—a company-reported deployment result.
The API remains $0.08 per minute; Artificial Analysis’s benchmark costs are per task and should not be treated as API rates.
On Big Bench Audio, xAI reports $0.025 per task for Grok Voice 3.0, versus $0.030 for GPT-Live-1 with Astra Medium and $0.055 for GPT-Realtime-2.1 High.
For browser and mobile clients, xAI recommends short-lived tokens; direct API-key authentication is intended for server-side use.
Developers building phone agents and voice assistants on xAI’s API now have a new model—and some have already switched automatically. xAI released Grok Voice 3.0 on October 14, 2026, keeping its predecessor’s $0.08-per-minute price while promising faster responses and stronger customer-support performance. The release leaves the existing API unchanged.
The Speech to Speech API sends audio and text in both directions over a WebSocket connection, letting an application exchange speech and messages with Grok in real time. Client events supply conversation details and user audio; server events return audio and text responses. Developers configure the voice and instructions through a session update rather than opening a separate connection for each setting.
Turn-taking is configurable. With server-side voice activity detection enabled, the service detects when a user has finished speaking. Developers can also set an idle timer that prompts the agent to speak again after silence. That timer resets after each response and keeps firing until the user speaks; it is off by default.
For existing integrations, the October 14 release notes distinguish between applications that follow the latest model and those selecting a specific version:
Authentication also separates browser and mobile apps from server software. xAI recommends short-lived tokens for client-side applications, keeping the API key off the device. Direct API-key authentication through the connection’s authorization header is reserved for server-side use.
xAI’s launch post cites October 13 testing by Artificial Analysis, placing Grok Voice 3.0 first on its Speech-to-Speech Index with 86.9 points. That index combines speech reasoning, support-call testing, human preferences and task success. The company also reports a 97% Full Duplex Bench score, measuring how the model handles pauses, turn-taking and interruptions.
xAI reproduces Artificial Analysis results showing higher success for version 3.0 on τ-Voice, which tests support calls with varied accents, background noise and interruptions.
The support-call comparison is not a clean sweep. In xAI’s published chart, GPT-Live-1 with the Astra Medium configuration scores 74.5%, above Grok Voice 3.0’s 71.6%. xAI’s argument is responsiveness: it says Grok delivers comparable support-call performance at 40% lower latency. Its headline claim is a response delay of under one second, not the highest success rate on every test.
The cost comparison is task-specific, too. xAI cites an average cost of $0.025 per task for Grok Voice 3.0 in Artificial Analysis’s Big Bench Audio run. The figures are $0.030 for GPT-Live-1 with Astra Medium and $0.055 for GPT-Realtime-2.1 High. Those are measured costs for that benchmark run, distinct from the API’s per-minute rate.
For enterprise calls, xAI also reports 74% accuracy on ServiceNow’s EVA-Bench. ServiceNow’s Nitin Aggarwal says the model combines top accuracy with responsive conversation across more than 200 realistic enterprise scenarios. That is a separate evaluation from the support-call comparison above.
xAI says Grok Voice 3.0 already handles millions of daily conversations in Tesla vehicles and powers Grok Bot’s voice mode. For Starlink, the company describes an operational customer-service role: Grok Voice answers 20,000 calls a day across multiple languages, with roughly 70% resolved without a human.
The company says a single agent built on version 3.0 can verify customers, troubleshoot hardware, issue credits and ship replacement equipment. It also says one in four Starlink sales calls ends in a purchase before the caller hangs up. These are company-reported deployment outcomes, separate from the benchmark scores, and describe agents taking actions as well as answering questions.
Loading discussion...
Join the conversation
Explain when easier maintenance should give way to tighter control.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.