Meet Grok Voice 3.0: SpaceXAI Voice Agent Model Ranks #1 on Artificial Analysis

SpaceXAI has released Grok Voice 3.0, its latest realtime speech-to-speech model. The company says it listens, reasons, calls tools, and speaks back, and on Artificial Analysis it now sits at the top of the Speech-to-Speech Index with a score of 86.9.

That lead matters because the model is still aimed squarely at developers, not hobbyists. It runs through SpaceXAI’s hosted Voice Agent API, exposed as grok-voice-3.0 on the /v1/realtime WebSocket endpoint. The grok-voice-latest alias already points there, so existing agents were switched over automatically. And if someone was hoping for model internals, SpaceXAI isn’t sharing those. No parameter count. No architecture details. No weights to download.

The benchmark picture is strong, but not spotless. Grok Voice 3.0 does well across the board, though it doesn’t win every slice of the test. It scores 98.5% on Big Bench Audio, 97.0% on Full Duplex Bench, 94.8% on task success rate, and 1070 Elo in Speech Agent Arena. Its weakest published result is τ-Voice at 71.6%, where GPT-Live-1 (Astra, medium) does better at 74.5%. Still, Grok Voice 3.0’s overall index score beats GPT-Live-1’s 83.2 and Gemini 3.8 Live Extended Thinking’s 82.6.

Speed and pricing also look competitive. Artificial Analysis measures time to first audio at 0.81 seconds, which lines up with SpaceXAI’s claim of 40% lower latency than GPT-Live-1’s 1.34 seconds. The catch is that Grok Voice 2.0 was faster at 0.70 seconds. Pricing stays unchanged from the prior version at $0.08 per minute, while the listed cost per Big Bench Audio task is $0.025. For teams already on the API, the practical story is simple: same price, better benchmark results, and a higher ceiling on the leaderboard.

Deployment is meant to be straightforward. The API is compatible with OpenAI Realtime, supports WebSocket, WebRTC demos, and SIP, and offers 28 built-in voices plus custom voice cloning from a reference clip of up to 120 seconds. SpaceXAI also says the service covers 20 locales, has server-side voice activity detection, and supports session resumption for up to 30 minutes. Grok Voice 2.0 is now deprecated, so the latest alias is doing the heavy lifting whether users notice or not.