SpaceXSI launches Grok Voice 3.0: speech-to-speech model tops Artificial Analysis S2S Index at 86.9, price held at $0.08/min

Background

Voice agents must handle accents, noise and barge-ins while answering in under a second; developers have long traded cascaded ASR→LLM→TTS intelligence against end-to-end speed. SpaceXSI (SpaceX Super Intelligence, x.ai) released Grok Voice 3.0 on 2026-10-14 17:04 UTC, claiming the frontier on both axes.

What shipped

An end-to-end speech-to-speech model served through the Voice Agent API WebSocket endpoint (wss://api.x.ai/v1/realtime) with server VAD turn detection, tool use (e.g. web_search) and optional reasoning.effort. SpaceXSI says it is the same model behind Grok voice mode, Tesla in-car assistance and Starlink support. Parameter count and training details were not disclosed.

Benchmarks (Artificial Analysis, Oct 13, as cited by the vendor)

• Speech-to-Speech Index: 86.9 vs 81.3 (Grok Voice 2.0), 83.2 (GPT-Live-1 Astra), 82.6 (Gemini 3.8 Live)
• Big Bench Audio 98.5%, Speech Agent Arena Elo 1070, task success 94.8%, Full Duplex Bench 97.0%
• τ-Voice 71.6%, comparable with GPT-Live-1 (Astra) at ~40% lower latency; top accuracy (74%) on ServiceNow's EVA-Bench

These are vendor-cited third-party results; broader independent replication is pending.

Production claims

Starlink call centers handle ~20,000 calls/day on Grok Voice with ~70% resolved without a human, per SpaceXSI.

For developers

• Apps on grok-voice-latest were upgraded automatically; same endpoint, session parameters and events, still $0.08/min ($4.80/hr).
• Pin grok-voice-3.0 in production; grok-voice-2.0 is deprecated.
• Use ephemeral tokens for client-side apps. See the Speech to Speech docs and release notes.