Wait Which Model?
← Back to Voice models
xAINotable

Grok Voice 3.0

Released Oct 14, 2026

xAI's speech-to-speech model for voice agents, served through its Speech to Speech (Voice Agent) API. xAI also runs it on Starlink support lines and in Tesla cars.

Price per minute
$0.080/min
Inputs
audio, text
Tool calling
Yes
Languages
20
Voices
—
Big Bench Audio
98.5%
Access
Generally available
Licence
Proprietary
Strengths & weaknesses

Strengths

  • Big Bench Audio 98.5% and #2 on AA's S2S Index (86.9)
  • Flat $0.08 per minute, unchanged from Grok Voice 2.0
  • Built-in web search, X search, collections, MCP and custom function tools
  • A drop-in upgrade: agents on the grok-voice-latest alias moved over automatically

Weaknesses

  • 0.81s time to first audio on AA, up from 0.70s on Grok Voice 2.0
  • Anyone on grok-voice-latest was switched on launch day unless they had pinned a version
  • Single region (us-east-1); sessions capped at 120 minutes

Notable: frontier-lab clause (also top 10 on both AA boards: #2 S2S Index, #6 Big Bench Audio). AA lists the developer as 'SpaceXAI'. Moved from models.json; facts re-verified 2026-11-06. Launch figures (xAI citing AA): S2S Index 86.9 (2.0: 81.3), Big Bench Audio 98.5%, Speech Agent Arena Elo 1070, Task Success 94.8%, Full Duplex Bench 97.0%, τ-Voice 71.6% (GPT-Live-1 74.5%), ServiceNow EVA-Bench 74%. xAI's docs mark grok-voice-2.0 deprecated with no retirement date. Languages: the speech-to-speech guide lists 20 and says others work with varying accuracy. Voices: the guide points to the TTS voice list, and the TTS docs list 5 built-in voices (ara, eve, leo, rex, sal) while an xAI cookbook mentions 28, so the count is left null. The launch post doesn't name tool calling, but the docs list it.

Links

API model strings

  • xaigrok-voice-3.0
News