xAI Launches Voice-to-Voice Model Grok Voice 3.0, Claims It Is the World's Smartest Voice-to-Voice Model
xAI announced the launch of Grok Voice 3.0, claiming it to be the world's smartest voice-to-voice model, now available for use in the Voice Agent API and Voice Agent Builder.
The model is priced at $0.08 per minute ($4.80 per hour), maintaining the same price as its predecessor Grok Voice 2.0, with a response latency of less than 1 second.
In independent testing by Artificial Analysis, Grok Voice 3.0 scored 86.9 on the Speech-to-Speech Index, ranking first, surpassing Grok Voice 2.0's 81.3, GPT-Live-1 (Astra, Medium) at 83.2, and Gemini 3.8 Live (Extended Thinking, High) at 82.6.
The model scored 97% on the Full Duplex Bench, which measures its natural handling of pauses, turn-taking, and interruptions.
Grok Voice supports millions of conversations daily in Tesla and SpaceX scenarios. The Starlink call center uses the model to handle 20,000 customer calls each day, resolving about 70% without human intervention, with a quarter of sales calls completing purchases before the call ends.
A single agent built on Grok Voice 3.0 can complete customer verification, hardware troubleshooting, credit issuance, and replacement device shipping.
Who is buying: Developers access voice agent capabilities via API, and businesses use it for customer service and sales scenarios. Who is selling: xAI opens the model at the same price as its predecessor. Event-driven: Voice agent performance and cost competition. Funding flows towards low-latency, high task completion rate voice infrastructure. Beneficiaries are businesses deploying voice customer service and sales agents; those under pressure are competitors with higher latency or costs.
Source: Public Information
ABAB AI Insight
xAI previously launched the Grok Voice series and deployed it in Tesla vehicles and its own applications, with Grok Voice 2.0 being the version prior to this upgrade. The company has long embedded voice capabilities into the actual operations of SpaceX and Tesla, including Starlink customer support, rather than just remaining at general conversation.
On the capital path, xAI opens the voice model to developers at a fixed per-minute price while also using it extensively within its own ecosystem to validate effectiveness. Resources are focused on reducing latency and improving task completion rates, enabling a single agent to cover the entire process from identity verification to device replacement. The motivation is to accumulate real call data through large-scale internal deployment and then open it to external users at the same price, forming a closed loop of data and product.
Similar cases include OpenAI's Realtime API and Google's Gemini Live voice capabilities. The current industry is in an expansion phase, with voice agents transitioning from demonstration tools to production systems capable of independently handling customer service and sales, with capital concentrating on providers with low latency and high self-resolution rates.
Essentially, this represents a technological substitution and industrial chain restructuring. Traditional customer service relies on human agents to handle repetitive inquiries; with voice-to-voice models achieving latency below 1 second and improved task success rates, they can directly replace some human roles. The mechanism lies in the model's simultaneous handling of voice understanding, reasoning, and action, reducing the proportion of human intervention, shifting call centers from labor-intensive to model-driven, and altering the cost structure of customer service and sales processes.
ABAB News · Cognitive Laws
Latency below one second, human seats begin to give way
Voice models are not selling conversations, but self-resolution rates
Internal deployment accumulates data, external API sells at the same price