xAI released grok-voice-transcribe-2.0, a new speech-to-text model, on September 17, 2026, and made it the default for its transcription API. On October 2, 2026, the previous model, grok-voice-transcribe-1.0, reached end of life, and requests to it are now routed to version 2.0 at the same price. Any team calling xAI's speech-to-text endpoint is therefore already on the new model, whether or not it changed any code.
Two entries in the xAI release notes define this release:
grok-voice-transcribe-2.0 became available. Both 1.0 and 2.0 could be selected, with 2.0 as the default.grok-voice-transcribe-1.0 reached end of life. The release note says all requests to that model name are routed to 2.0 at the same price, and describes 2.0 as having higher accuracy.xAI's release notes do not publish an accuracy figure for the claim, so treat "higher accuracy" as xAI's description rather than a measured result. The practical implication of the forced migration is that transcripts produced after October 2 may differ from earlier ones even for identical audio. Pipelines that compare transcripts over time, or that tuned post-processing to version 1.0 output, should re-check.
The speech-to-text documentation describes two ways to use the model:
https://api.x.ai/v1/stt, uploading a file or passing a URL for the server to fetch.wss://api.x.ai/v1/stt, sending raw audio frames and receiving JSON transcript events as the audio is processed.Features listed in the documentation include:
Existing xAI speech-to-text users are already on it; the action item is to validate transcripts and any downstream parsing.
Teams building voice agents may find the streaming endpoint's turn-detection controls the most relevant feature, since deciding when a speaker has finished is a common source of awkward pauses and interruptions in voice interfaces.
Teams processing recorded calls and meetings can use multichannel mode, diarization and key terms together.
For multimodal models that understand audio as part of a broader task rather than transcribing it, see Qwen3.8-Omni-Flash. For xAI's main language model release this cycle, see Grok 4.7.
When was grok-voice-transcribe-2.0 released?
September 17, 2026. Version 1.0 reached end of life on October 2, 2026.
Do I need to change my code?
Not to keep working: requests to grok-voice-transcribe-1.0 are routed to 2.0 at the same price. You should still validate output, since the model changed.
Does it support real-time transcription?
Yes, through the WebSocket endpoint, with configurable endpointing and Smart Turn detection.
Discover more content: