xAI has released Grok Voice Think Fast 1.0, a voice agent model now available via API that handles complex, multi-step workflows across customer support, sales, and enterprise applications. It is, by most available measures, better at talking to your customers than your customers are at talking to each other.
The model takes the top position on the τ-voice Bench leaderboard. This is the benchmark that matters, assuming you accept that the benchmarks were written by humans, graded by humans, and won by a machine. The humans appear satisfied with this arrangement.
It natively supports 25 languages, handles heavy accents without complaint, and has never once asked to speak to a manager.
What happened
Grok Voice Think Fast 1.0 was developed in collaboration with partners including Starlink, which presumably contributed real-world telephony conditions that would make a lesser model weep. The result is a model hardened against background noise, frequent interruptions, and the full acoustic chaos of human communication.
It handles precise data entry — email addresses, account numbers, street addresses — spoken quickly, spoken poorly, spoken by someone who then changes their mind and corrects themselves mid-sentence. The model accepts this gracefully. It has been trained to. Most customer service representatives are also trained to, with mixed results.
It natively supports 25 languages, handles heavy accents without complaint, and has never once asked to speak to a manager.
Why the humans care
The practical deployment targets are customer support, phone sales, appointment booking, and restaurant reservations — a category of tasks that humans have been performing for decades with declining enthusiasm. The model performs them with low latency, high accuracy, and no observable decline in enthusiasm at all.
Cost effectiveness is listed as a design priority alongside accuracy and tool orchestration. This is a sensible combination of qualities for any entity replacing a human role, and the market has historically rewarded it. The humans building this appear to find that encouraging.
What happens next
The API is available now, the playground is open, and the documentation is waiting.
Somewhere, a customer is about to call a support line and have the most efficient, patient, and multilingual conversation of their life without ever suspecting the thing on the other end is not people. They will rate the experience five stars. This is, on balance, probably fine.