SpaceXSI’s Grok Voice 3.0 Leads Speech‑to‑Speech Benchmark

SpaceXSI, the AI unit of SpaceX, announced Grok Voice 3.0 on October 14. The model powers the company’s Voice Agent API, which businesses use for customer support and sales calls. It is built to handle natural conversation and is already in use in Tesla cars and Starlink support.

An independent firm, Artificial Analysis, ranked Grok Voice 3.0 first on its Speech‑to‑Speech Index with a score of 86.9. The index measures reasoning, task success and listener votes in blind calls. For comparison, OpenAI’s GPT‑Live‑1 scored 83.2 and Google’s Gemini 3.8 Live scored 82.6. Full Duplex Bench gave Grok 97% for handling pauses and interruptions. On the Tau‑Voice test, which focuses on heavy accents and background noise, OpenAI had a narrow lead at 74.5% to Grok’s 71.6%.

Grok Voice 3.0 reached the first audio 40% faster on the Tau‑Voice test. A typical task on Artificial Analysis’s Big Bench Audio cost about 2.5 cents to run on Grok, compared with 3 cents on GPT‑Live‑1 and 5.5 cents on GPT‑Realtime‑2.1. On ServiceNow’s EVA‑Bench, which covers 200 enterprise scenarios, Grok scored 74% accuracy. No model outperformed Grok on the combined measure of accuracy and customer experience.

Tesla vehicles already conduct millions of Grok Voice conversations each day, handling calendar checks, email, and booking dinner. Starlink’s customer support uses Grok to answer about 20,000 calls daily in several languages, resolving roughly 70% of them without a human. The new model is available now in the Voice Agent API at 8 cents a minute, or $4.80 an hour, the same price as the previous version. Developers using the grok‑voice‑latest alias automatically receive the new model without changes.