Grok Voice 3.0
modelYour notes
SpaceXAI's third-generation speech-to-speech model, and currently the top-ranked model on Artificial Analysis's Speech to Speech Index. AA scores Grok Voice 3.0 at 86.9, first of all models measured, ahead of GPT-Live-1 (Astra, medium) at 83.2, Gemini 3.8 Live Extended Thinking at 82.6 and the 2.0 generation at 81.3 — a 5.6-point generational gain on an equal-weighted composite of Speech Reasoning (Big Bench Audio), agentic performance (τ-Voice), arena preference and task success rate. AA's component figures: τ-Voice 71.6%, task success rate 94.8%, Full-Duplex-Bench 97.0%, arena preference Elo 1064, and time to first audio 0.81 s, at a measured $4.80 per hour of input audio on the Big Bench Audio subset. AA's Big Bench Audio number is inconsistent on its own page — 99% in the summary table against 98.5% in the FAQ — so it is left out of the structured scores here.
The release itself is a changelog line, not a launch post. x.ai published no announcement that is reachable (x.ai/news returns 403), and the only primary record is the API release notes: "grok-voice-3.0 is now available with Speech to Speech, at the same price as Grok Voice 2.0," with "The API is unchanged," grok-voice-latest now pointing at 3.0 "so sessions that use the alias are upgraded automatically," and grok-voice-2.0 deprecated. The notes are dated only "October"; the October 14 date here comes from secondary coverage and is not confirmed by a primary source. Price parity with 2.0 implies $0.08 per audio minute, the figure x.ai published for the previous generation, but 3.0's own rate card is not stated in the notes.
Nothing technical is disclosed. No parameter count, architecture, training data, language coverage or vendor benchmark table — where the 2.0 release at least described RL for conversational behaviour and tool use with reasoning in parallel with speech, 3.0 arrives as a drop-in model-string swap whose only public evidence is third-party. Note that AA now lists the predecessor as "Grok Voice 2.0" at 81.3, below the 82.9% recorded on its Speech-to-Speech Quality Index at the time that model was filed; the index has been renamed and recomposed since, so the two generations should only be compared within AA's current table.
Model Details
Benchmark Scores
| Benchmark | Score | Mode |
|---|---|---|
| AA Speech to Speech Index | 86.9 (#1) | — |
| tau-Voice (AA agentic) | 71.6% | — |
| Task success rate (AA) | 94.8% | — |
| Full-Duplex-Bench (AA conversational dynamics) | 97.0% | — |
| Time to first audio (AA) | 0.81 s | — |