News · SI
SpaceXSI has released Grok Voice 3.0, a speech-to-speech model for real-time voice agents that ranks first on Artificial Analysis's index with under one second of latency, at the same $0.08 a minute. Here is what changed and what it means for teams running phone agents.
SpaceXSI, the company formerly known as xAI, has released Grok Voice 3.0, a speech-to-speech model for real-time voice agents[^1][^3]. It ranks first on Artificial Analysis's Speech-to-Speech Index, answers in under a second, and costs the same US$0.08 a minute as the model it replaces[^1].
01 / 04
In Artificial Analysis's testing, Grok Voice 3.0 scored 86.9 on the Speech-to-Speech Index, which combines spoken reasoning, a support-call test called τ-Voice, and preference and task success from blind head-to-head calls[^1]. OpenAI's GPT-Live-1 scored 83.2 and Google's Gemini 3.8 Live 82.6[^1][^3]. It also scored 97% on Full Duplex Bench, which measures how naturally a model handles pauses, turn-taking and interruptions[^1].
It is not first on everything. On τ-Voice, which tests support calls with accents, background noise and interruptions, it scored 71.6% against GPT-Live-1's 74.5%; SpaceXSI's argument is that it gets there with about 40% lower latency[^1]. Artificial Analysis also puts its cost per task on Big Bench Audio at about US$0.025, against US$0.030 for GPT-Live-1[^1].
02 / 04
SpaceXSI says the model handles millions of conversations a day in Tesla vehicles and runs Starlink's call centres, which take 20,000 calls a day[^1]. It says about 70% of those calls finish with no human involved, and one in four sales calls ends in a purchase before the caller hangs up[^1]. Those are the company's own figures.
03 / 04
Agents using the grok-voice-latest alias were moved to the new model automatically, and the API is unchanged[^2]. Grok Voice 2.0 is now deprecated, so anyone who pinned it should switch the model name to grok-voice-3.0[^2].
04 / 04
Voice agents have mostly been held back by two things: the awkward pause before an answer, and the cost of a model smart enough to finish the task. A sub-second model that resolves most routine calls at under five dollars an hour makes an after-hours line or a first-line support agent a realistic project for a small business. Start with the calls you already have scripts for, keep a clean hand-off to a person, and measure resolution on your own calls rather than on benchmark scores.
FAQ
Grok Voice 3.0 is a speech-to-speech model from SpaceXSI, the company formerly called xAI, released on 14 October 2026 (US time). It listens and replies in audio directly, for real-time voice agents such as customer support and sales lines, and it is the model behind the Voice Agent API.
It costs US$0.08 a minute, or US$4.80 an hour, in the Voice Agent API, the same price as Grok Voice 2.0.
Not if your agent uses the grok-voice-latest alias, which SpaceXSI has already pointed at grok-voice-3.0. The API is unchanged. If you pinned grok-voice-2.0, which is now deprecated, set the model to grok-voice-3.0 or grok-voice-latest.
In Artificial Analysis's testing published on 13 October 2026, Grok Voice 3.0 scored 86.9 on the Speech-to-Speech Index, ahead of OpenAI's GPT-Live-1 at 83.2 and Google's Gemini 3.8 Live at 82.6. On the tau-Voice support-call test it scored 71.6%, slightly behind GPT-Live-1 at 74.5%, but SpaceXSI says it responds with about 40% lower latency.
SpaceXSI says it handles millions of conversations a day in Tesla vehicles and powers Starlink call centres that take 20,000 calls a day across several languages. It says about 70% of those Starlink calls are resolved with no human involved. These are the company's own figures.
That’s the story
Spectrum Web Co ships SI and Web3 products, places talent and builds social narratives as monthly services.
