SpaceXAI unveils grok voice 3.0 as top performing speech-to-speech model

SpaceXAI has launched Grok Voice 3.0, a speech-to-speech model that outperforms its predecessors and competitors on the Artificial Analysis Speech-to-Speech Index, achieving a score of 86.9. The model retains the same endpoint and pricing structure as Grok Voice 2.0, with a rate of $0.08 per minute. It is now the default version, accessible via the grok-voice-3.0 endpoint and the grok-voice-latest alias. The model excels in several benchmarks, including a top accuracy score of 74% on ServiceNow’s EVA-Bench and a latency of 0.81 seconds, which is 40% lower than GPT-Live-1. Grok Voice 3.0 also has a cost advantage over GPT-Live-1 and Gemini 3.8 Live Extended Thinking, with a task cost of $0.025 versus $0.030 and $0.055 respectively. The API is compatible with the OpenAI Realtime API, simplifying migration for developers. SpaceXAI has not disclosed parameter counts or architecture details, but the model supports reasoning at a high effort level and offers faster turn times with the 'none' setting. It integrates tools such as collections search, web search, and custom functions within the session. Grok Voice 2.0 is now deprecated, and developers are advised to use the latest version for production.