Grok transcribes speech three times more accurately: 6.8% errors instead of 20.6%

It turns speech into text more accurately than the previous version, while xAI charges the same. Russian is among the supported languages.

Before and now. Version 1.0 confused 20.6% of words in short multilingual phrases; the new version confuses 6.8% (19 languages, xAI measurement, 18.09.2026). Streaming rivals trail behind: Meta Muse at 3.1%, ElevenLabs Scribe v2 at 3.6%, and v1.0 itself at 3.9% (Artificial Analysis, 21.09.2026). The final transcript arrives in 0.49 seconds.

What the price includes. $0.10 per hour of audio in batch mode and $0.20 per hour in streaming, the same as version 1.0. xAI includes speaker diarization, word-level timestamps, key-term prompts, and formatting at no extra charge.

You can try it without code in the console at console.x.ai. In code, it is a POST to https://api.x.ai/v1/stt with a key in the header and the fields model=grok-voice-transcribe-2.0 and file=@audio.mp3. Files can be up to 500 MB, across 12 formats, with up to 100 key terms per request.

Streaming runs over WebSocket at wss://api.x.ai/v1/stt: the client sends audio in binary frames, and with the interim_results parameter the model returns draft text about every 500 ms. Formatting works for 24 languages, including Russian, and the model catches language switching within a single recording in one pass.

At Atlassian, the model already transcribes recordings in Loom. xAI has not released the weights, so you cannot run it yourself: API only.

Version 1.0 has only its final weeks left: anyone who does not explicitly pin it will move to 2.0 automatically.