Grok Voice 3.0 from SpaceXSI outscores OpenAI and Google on speech-to-speech index

The new Grok Voice 3.0 model's performance suggests a notable improvement in the competitive landscape for voice AI. Its top ranking on Artificial Analysis's Speech-to-Speech Index, which incorporates reasoning, task success, and listener votes, indicates a strong overall capability. The model also demonstrated superior performance in handling natural conversational elements, scoring 97% on Full Duplex Bench for coping with pauses and interruptions.

While Grok Voice 3.0 led in general accuracy and natural conversation, OpenAI's GPT-Live-1 maintained a narrow lead in specialized tests like Tau-Voice, which is built around support calls with heavy accents and background noise. However, SpaceXSI reported that its model achieved first audio about 40% faster in that specific test, which could offer a practical advantage where rapid initial response times are important.

Beyond performance metrics, the cost efficiency and existing deployments of Grok Voice 3.0 are notable. The model is priced at eight cents a minute, matching its predecessor, and demonstrated lower task completion costs in Artificial Analysis's Big Bench Audio run compared to OpenAI's models. Its current use in millions of daily Tesla vehicle conversations and handling approximately 20,000 daily Starlink customer calls, with a reported 70% resolution rate without human intervention, suggests a degree of proven operational reliability and scalability.