xAI released grok-voice-think-fast-1.0 today, claiming the top spot on τ-voice Bench at 67.3% — well ahead of Gemini 3.1 Flash Live (43.8%), their own previous Grok Voice Fast 1.0 (38.3%), and GPT Realtime 1.5 (35.3%). The vertical breakdowns are even more lopsided: Telecom shows the model at 73.7% versus Gemini's 21.9% and GPT Realtime's 21.1%. Retail at 62.3%, Airline at 66%. Big gaps across the board, and xAI is making a serious claim on the full-duplex voice agent category.
Step back from the leaderboard for a second. τ-voice Bench is xAI's benchmark, modeled after Sierra's τ-bench framework but extended to noisy audio, accents, and interruption handling. Self-graded benchmarks are not automatically wrong, but the comparison set is also worth reading carefully: Gemini 3.1 Flash Live is Google's cheaper, lower-latency voice tier, not the top-end model, and GPT Realtime 1.5 is OpenAI's older voice product, not whatever they have brewing now. xAI did not benchmark against Gemini 3.1 Pro Live or against any of the production-deployed voice stacks Sierra and PolyAI run. The lead is real, but the comparison is curated.
The more useful data point is buried lower in the announcement: grok-voice-think-fast-1.0 is already running Starlink's live phone operations. The numbers xAI publishes from that deployment are 20% sales conversion on phone inquiries, 70% autonomous resolution on support, 28 distinct tools wired into hundreds of workflows, and 25+ language support. Those are production metrics from a customer base that does not stay on the line for a bad agent. The "background reasoning with zero added latency" framing — running reasoning passes in parallel with speech generation rather than serially — is the right architectural answer to the problem the older voice agents have, where you hear the model think before it answers.
For developers building voice products, the honest takeaway is that the API is worth a real evaluation, especially if you have noisy phone audio or need tool calls mid-conversation. Don't take the τ-voice Bench numbers as gospel — run your own conversation flows against Gemini 3 Pro Live and OpenAI's gpt-realtime, on your own audio, with your own tools, before you commit. The Starlink deployment is the strongest evidence that the model is actually production-grade; the leaderboard is the weakest. xAI has not published pricing or latency targets yet, which are the next questions anyone evaluating this for a real call center will need answered.