Grok Voice 3.0 Answers in 1.6 Seconds
xAI's new speech-to-speech model lands on our S2S board at 0.86, finishes every scripted call, and has the fastest time to a real answer of any model we time.
xAI released Grok Voice 3.0 on October 14. We ran it through the same six concierge calls every model on our S2S board gets: book a flight, move a reservation, convert a budget, handle a refund, recover from a failed tool, remember what the caller said ten turns ago.
It scores 0.86 overall, level with gpt-live-1 and 0.01 behind Gemini 3.8 Live.
Where it stands out
| Model | Overall | Action | Call completion | Task capability | First sound | Answer |
|---|---|---|---|---|---|---|
| Gemini 3.8 Live | 0.87 | 0.83 | 98% | 0.92 | 1111ms | 2013ms |
| gpt-live-1 | 0.86 | 0.78 | 100% | 0.90 | 1217ms | 4187ms |
| Grok Voice 3.0 | 0.86 | 0.67 | 100% | 0.94 | 782ms | 1598ms |
| Grok Voice Think Fast 2.0 | 0.85 | 0.67 | 100% | 0.92 | 820ms | — |
Three things in that row matter for a production agent.
It gets to the point. “First sound” is when the caller hears anything. “Answer” is when they hear the part that answers them, filler excluded. Grok Voice 3.0 reaches the answer in 1.6 seconds. gpt-live-1 makes a sound in about the same time but needs 4.2 seconds to say something useful. On a phone call that gap is the difference between a conversation and a hold.
It finishes the call. Every scripted call ran to the end, across all 18. It picked the right tool, didn’t invent prices, and recovered honestly when a booking failed. Its task capability of 0.94 is the highest on the board.
It remembers, and keeps it short. On the hard-recall turn, where the caller changes a booking day twice and then asks what was booked, it named the latest day in 17 of 18 calls. Its answers ran noticeably shorter than 2.0’s in the same run.
At $0.08 a minute it costs the same as 2.0, and it is a straight step up from it: higher capability, 38ms faster to first sound in the same run.
If you pinned an older Grok voice model
xAI now serves Grok Voice 3.0 when you request grok-voice-fast-1.0 or grok-voice-think-fast-1.0. If your agent is pinned to either, it is already running 3.0. We retired both 1.0 rows from the board for that reason. grok-voice-think-fast-2.0 still answers as 2.0.
One thing to know
When the prompt doesn’t include today’s date, Grok Voice 3.0 sometimes picks a specific date on its own, like “March 19th”, instead of saying “this Thursday”. Give it the date, as most production agents do, and it uses the right one.