AI & Creative Tools

Grok's New Voice Model Cuts Response Time to 0.7 Seconds — and It's Rolling Out Today

Think Fast 2.0 listens, reasons, and speaks in one continuous pass rather than waiting for turns, with reasoning-token use down 60% and every 'grok-voice-latest' call switching over automatically starting August 5.

xAI announced Grok Voice Think Fast 2.0 on July 29 — less than three months after the original Think Fast debuted — and starting today, August 5, every application calling grok-voice-latest gets switched over to it automatically.

The architecture is the actual story

Most speech-to-speech systems still have an audible seam: you stop talking, there’s a beat while the model processes and formulates a reply, then it starts speaking. Think Fast 2.0’s core design listens, reasons, and speaks simultaneously — no waiting for a turn boundary, no scripted pause. The measurable result is time-to-first-audio down to roughly 0.70 seconds, compared to 1.25 seconds for the previous version, achieved partly by cutting reasoning-token use by about 60% — the model spends less computational effort “thinking” before it starts producing sound, without the reasoning quality visibly degrading in xAI’s own benchmarks.

The model also claims a real jump in transcription accuracy specifically in noisy environments, with xAI citing roughly a 10x accuracy gap versus existing speech-recognition systems under those conditions, and support for 25+ languages.

Why sub-second latency matters for creative work specifically

xAI’s own framing leans toward enterprise use cases — customer support, multi-step workflows, handling interruptions gracefully in a call-center context. But the same latency and interruption-handling properties are exactly what’s historically made AI voice tools frustrating for creative and performance applications: live improvisational voice work, interactive character voices in games or installations, and real-time voice-driven interfaces all break down the moment a model can’t handle a natural, messily-timed conversation. A model that can be talked over, interrupted mid-response, and still recover gracefully — rather than needing a clean, scripted turn-taking structure — is a meaningfully different tool for anyone building something that has to respond to an actual human in real time rather than a queued voice-note exchange.

Pricing and rollout

Think Fast 2.0 is priced at $0.08 per minute of audio. Beyond the automatic grok-voice-latest switch today, it’s available through Agent Builder and the standard xAI API for developers who want to target the specific model version directly rather than riding the “latest” alias.