SpaceXAI's Grok Voice 3.0 Tops Artificial Analysis Speech-to-Speech Index
SpaceXAI released Grok Voice 3.0, a hosted speech-to-speech agent model that scores 86.9 and ranks first on Artificial Analysis's Speech-to-Speech Index, though rivals still lead on some subtests.
The model is exposed as grok-voice-3.0 on the /v1/realtime WebSocket endpoint, and the grok-voice-latest alias already points to it, so existing agents switched over automatically. SpaceXAI has not published parameter counts or architecture details. Reasoning is enabled by default at reasoning.effort "high" and can be set to "none" for faster turns, while server-side voice activity detection handles barge-in. Tools execute inside the session and include collections search, web search, X search, remote MCP servers and custom functions. There are no open weights and no local or quantized variants.
The Artificial Analysis Speech-to-Speech Index is an equal-weighted average of four components: Big Bench Audio reasoning, τ-Voice agentic performance, Speech Agent Arena preference and task success rate. Behind Grok Voice 3.0's 86.9, GPT-Live-1 (Astra, medium) scores 83.2, Gemini 3.8 Live Extended Thinking scores 82.6 and Grok Voice 2.0 scores 81.3.
The component results are strong but not uniformly first. On Big Bench Audio, Grok Voice 3.0 scores 98.5%, ranked fourth, with StepAudio 3 Realtime leading at 99.7%. On Full Duplex Bench it scores 97.0%, also fourth, behind StepAudio 3 Realtime at 98.9%. Its task success rate is 94.8%, up from 94.6% for Grok Voice 2.0, and its Speech Agent Arena Elo of 1070 is the highest among the four models compared. On τ-Voice it scores 71.6%, up from 56.5%, but below GPT-Live-1 (Astra, medium) at 74.5%. SpaceXAI reports a top accuracy score of 74% on ServiceNow's EVA-Bench, which grades enterprise voice agents on accuracy and experience.
On speed, Artificial Analysis measures 0.81 seconds to first audio for Grok Voice 3.0 against 1.34 seconds for GPT-Live-1 (Astra, medium), which supports SpaceXAI's claim of 40% lower latency. Grok Voice 2.0 was faster at 0.70 seconds. On cost, SpaceXAI cites Artificial Analysis's cost per Big Bench Audio task: $0.025 for Grok Voice 3.0, against $0.030 for GPT-Live-1 (Astra, medium) and $0.055 for GPT-Realtime-2.1 (High). Measured per hour of input audio, Gemini 3.8 Live Extended Thinking is cheaper at $3.50 versus $4.80. The list prices are not directly comparable: GPT-Live-1 costs $0.05 per minute with backend model and tool usage billed separately, while Gemini 3.8 Live bills audio per minute plus text tokens.