xAI ships Grok Voice Think Fast 2.0 — time-to-first-audio hits 0.70s
xAI’s Grok Voice Think Fast 2.0 ships speech-to-speech at 0.70s TTFA and $0.08/min, becoming the latest grok-voice alias for low-latency voice agents.
On July 29, 2026, xAI published Grok Voice Think Fast 2.0: improved transcription, reasoning, and time-to-first-audio of 0.70 seconds, priced at $0.08 per minute. On August 5, grok-voice-latest routes to 2.0 unless you pin 1.0. Starlink A/B tests, xAI says, lifted conversion and support containment. Release notes live at docs.x.ai.
This is a speech-to-speech SKU, not Grok 4.5 and not Grok Bot.
Latency and accuracy claims
Artificial Analysis numbers in the post: overall speech-to-speech quality index 82.9% versus 75.7% on 1.0; conversational dynamics 95.1% versus 77.8%; agentic τ-voice 56.5%; time-to-first-audio 0.70s versus 1.25s on 1.0 and 2.98s on Gemini 3.1 Flash (High). Those are named benches. Use them as xAI-cited AA figures.
Transcription: xAI claims 1.5–2.0× better word-error than Deepgram Nova 3 and ElevenLabs Scribe v2 across thousands of short phrases in 24 languages, 1.4× better than Voice Think Fast 1.0, and about 10× better than dedicated STT in noisy / telephony conditions. That last gap is the product claim. Voice models that die on hold music are toys.
Reasoning runs in parallel with speech. 2.0 uses about 0.4× the reasoning tokens of 1.0 at P50, which xAI says makes tool calls fire before the first sentence ends.
Pricing and latest-alias migration
$0.08 per minute of audio. On August 5, 2026, the grok-voice-latest alias moves from 1.0 to 2.0. No prompt rewrite required, xAI says. To stay on 1.0, pin grok-voice-think-fast-1.0 before that date. That is a foot-gun for anyone who treated latest as a constant.
Starlink A/B results
xAI says A/B testing on Starlink support (+1 888 GO STARLINK) showed a significant increase in sales conversion and support containment. Those are vendor A/B claims on a captive voice line. They are more interesting than a demo reel and less interesting than a public methodology. Containment up means fewer human handoffs. Conversion up means the voice agent closed more.
When to pin 1.0 vs upgrade
Pin 1.0 if you have a tuned IVR that 2.0’s shorter, less-fluffy style will break. Upgrade if latency and noise robustness are the ticket. Read the docs release notes before August 5 if latest is in production.
Grok Voice Think Fast 2.0 is a 0.70s voice model with a price and an alias clock. The 2.0 string is the one to put in the ticket. The alias is the one that will move without you.
