xAI announced Grok Voice 3.0 on October 14, 2026, as a real-time speech-to-speech model for voice agents and said it was available in the Voice Agent API. The company also said agents using grok-voice-latest switched to the new model automatically. (xAI announcement)
For developers, the practical details are in xAI’s API guide: it shows a WebSocket connection to /v1/realtime, with session settings for voice, instructions, server-side voice activity detection (VAD) and tools. The guide recommends ephemeral tokens for client-side connections so an API key need not be exposed in the app. (xAI Voice Overview)
What developers can configure
The documented session controls let an agent use a selected voice, follow supplied instructions, detect when a speaker has finished a turn, and call tools. xAI’s example enables web_search. These are the settings shown in the API guide, rather than a complete account of every possible integration.
The guide lists 28 built-in voice IDs. It also demonstrates creating a custom voice from a reference clip of up to 120 seconds; the resulting voice ID can be used with the real-time speech-to-speech API. (xAI Voice Overview)
What the benchmark comparison measures
The launch announcement cites Artificial Analysis results dated October 13, 2026, one day before xAI announced the model. In the Speech-to-Speech Index comparison, Grok Voice 3.0 scored 86.9. The listed entries and settings are: (xAI announcement; Artificial Analysis methodology and leaderboard)
| Model and variant | Speech-to-Speech Index |
| Grok Voice 3.0 | 86.9 |
| Grok Voice 2.0 | 81.3 |
| GPT-Live-1 (Astra, Medium) | 83.2 |
| Gemini 3.8 Live (Extended Thinking, High) | 82.6 |
Artificial Analysis calculates the index as an equal-weighted combination of four measures: Big Bench Audio speech reasoning, τ-Voice agentic performance, Arena Preference and task success. The score therefore combines distinct benchmark results into one index; it is not a score for every aspect of a live customer conversation.
xAI also reported a 74% accuracy result for Grok Voice 3.0 on ServiceNow’s EVA-Bench. ServiceNow describes the benchmark as evaluating complete, multi-turn spoken conversations, with separate measures for task accuracy and conversational experience. (xAI announcement; ServiceNow EVA-Bench)
The announced rate and other listed charges
xAI announced a speech-to-speech rate of $0.08 per minute in USD, the same per-minute rate it gave for Grok Voice 2.0. Its pricing page also lists $0.004 per text input; the page says it was last updated October 10, 2026, before the Voice 3.0 announcement. (xAI announcement; xAI API pricing)
That page lists speech-to-text and text-to-speech as separate API services, with their own rates: $0.10 per hour for REST transcription, $0.20 per hour for streaming transcription and $15 per million characters for text-to-speech. Those are prices for the separate services, not additional rates to assume for every speech-to-speech integration. (xAI API pricing)
xAI also reported that Grok Voice 3.0 handles millions of conversations each day in Tesla vehicles. For Starlink call centers, the company said it answers 20,000 calls per day, with roughly 70% resolved without a human. (xAI announcement)
