From the source
SpaceXAI released Grok Voice Transcribe 2.0, a speech-to-text model it says is twice as accurate as version 1.0 at the same price.
The model ranks first among 32 streaming models on the Artificial Analysis accuracy leaderboard, and its multilingual short-phrase word error rate fell from 20.6% to 6.8%.
Existing API integrations receive the update without code changes, and version 1.0 is due for deprecation in the coming weeks.