On October 14, 2026 (U.S. time), SpaceXSI (formerly xAI) released the AI model Grok Voice 3.0. It is a "speech-to-speech" model that takes in a person's voice directly and answers by voice, and it is used to build AI agents that interact by voice, such as for answering phone calls or holding conversations in cars. It has been available through the developer-facing Voice Agent API since the day of release.
xAI, which developed Grok, became SpaceXAI after being acquired by SpaceX, and since October it has been using the name SpaceXSI. The announcement page for this release also appears under the SpaceXSI name.
What is Grok Voice 3.0?
Most voice AI until now has gone through several steps: first transcribing speech into text, generating an answer in text, and then converting that into speech. Speech-to-speech models like Grok Voice 3.0 understand and answer in voice directly, so waiting times are shorter and conversations can more easily feel natural. SpaceXSI says the time until an answer comes back is under one second.
According to the developer documentation, apps using the setting that always points to the latest version (grok-voice-latest) will automatically switch to Grok Voice 3.0. The previous version, Grok Voice 2.0, has been deprecated (it is no longer recommended for future use).
Pricing
The price is $0.08 per minute ($4.80 per hour), the same as Grok Voice 2.0. SpaceXSI says that the cost per task on Big Bench Audio, Artificial Analysis's audio evaluation, is about $0.025, cheaper than OpenAI's GPT-Live-1 ($0.030) and GPT-Realtime-2.1 ($0.055), which it compared against.
Performance evaluation
The main results SpaceXSI highlighted in its announcement are as follows.
- Artificial Analysis Speech-to-Speech Index: first place with 86.9. OpenAI's GPT-Live-1 scored 83.2 and Google's Gemini 3.8 Live 82.6
- Full Duplex Bench (an evaluation of natural conversation, including interrupting and being interrupted): 97%
- τ-Voice (an evaluation of whether tasks requested by voice can be completed): 71.6%. On this item, OpenAI's GPT-Live-1 (74.5%) came out ahead, while Google's Gemini 3.8 Live scored 68.6%
- EVA-Bench (ServiceNow's evaluation of enterprise voice agents): 74% accuracy
All figures other than EVA-Bench come from Artificial Analysis, a third-party site that compares AI models.
How it is being used
According to SpaceXSI, Grok Voice is already handling millions of conversations a day in the cars of Tesla, another Elon Musk company. At the support line of the satellite internet service Starlink, the company says it resolves about 70% of roughly 20,000 calls a day without a human taking over.
CoAI's take
Grok Voice 3.0 is not a feature of a consumer app but an API for companies and developers to build voice-based help desks and assistants. SpaceXSI says it improved performance while keeping the price unchanged, giving companies considering automated phone answering and similar uses another candidate to compare. The next day, OpenAI also updated ChatGPT's voice model, and competition in voice AI is heating up.
On the other hand, OpenAI's model comes out ahead on some items, so it is not superior in every respect. It has also not been said when voice conversations in the Grok app will move to the new model, or how naturally it can speak Japanese, so if you plan to use it for a Japanese-language help desk, you will need to test it in real conversations.
