A faster, more accurate, and better-calibrated Grok voice for agents

xAI's Grok Voice Think Fast 2.0 reduces the voice response delay to 0.70 seconds. A gain in speed and accuracy for real-time agents.

Designed for real-time voice interactions, Grok Voice Think Fast 2.0 improves speech comprehension, conversational flow, and the reliability of actions triggered during a conversation. xAI indicates that the model can reason while speaking, allowing it to prepare or execute certain tool calls before even finishing its first sentence.

On Artificial Analysis's Speech-to-Speech Quality index, this version scores 82.9%, compared to 75.7% for the previous one, 79.1% for GPT-Realtime-2.1, and 69.5% for Gemini 3.1 Flash. The delay before the first sound also drops from 1.25 to 0.70 seconds.

Transcription improves across 24 languages, with a particular focus on noisy environments and compressed phone communications. According to xAI's tests, the model is reportedly 1.5 to 2 times more accurate than Deepgram Nova 3 and ElevenLabs Scribe v2 on the evaluated sentences. The gap could become much wider when background noise severely degrades the signal.

Voice responses have also been shortened to closer mimic natural conversation: one question at a time, fewer detours, and greater continuity in complex workflows. The median consumption of reasoning tokens is reportedly reduced by 60% compared to Grok Voice Think Fast 1.0.

The `grok-voice-latest` reference will automatically switch to this version. Pricing remains set at $0.08 per audio minute.