SpaceXAI has launched Grok Voice 3.0, its latest real-time voice model accessible via the v1/realtime WebSocket endpoint. The model is now the default for users of the grok-voice-latest alias, while the previous version, Grok Voice 2.0, has been officially deprecated.
In the Artificial Analysis Speech-to-Speech Index, Grok Voice 3.0 achieved a score of 86.9, outperforming GPT-Live-1 at 83.2 and Gemini 3.8 Live at 82.6. The model features a time-to-first-audio latency of 0.81 seconds, which SpaceXAI claims is 40% lower than GPT-Live-1.
The model supports advanced features including server-side voice activity detection and integrated tools like web and X search. Reasoning is enabled by default with a high effort setting, though it can be disabled to improve response speed.
Regarding costs, Grok Voice 3.0 is priced at $0.025 per Big Bench Audio task, compared to $0.030 for GPT-Live-1 and $0.055 for GPT-Realtime-2.1. However, Gemini 3.8 Live remains more cost-effective for hourly input audio at $3.50 versus $4.80.
SpaceXAI reports a 74% accuracy score on the ServiceNow EVA-Bench. The API is compatible with the OpenAI Realtime API, allowing developers to migrate by updating their URL and authentication keys.
