We are evaluating spacexai/grok-stt on AI Gateway for production transcription. Vercel support sent me here because speech-to-text is in beta.
xAI released Grok Voice Transcribe 2.0 on 18 Sep (https://x.ai/news/grok-voice-transcribe-2) and made it the default on their API when no model is passed. The Gateway model id has no version, and the response and the generation record only show xai/grok-stt, so I can't tell which version we get.
On 21 Sep I re-transcribed recordings that I had transcribed with Grok on 17 Sep, and more than half of the transcripts changed. Two runs on the same day are nearly identical, so it looks like the upstream model changed. Passing providerOptions.xai.model does not seem to reach xAI either.
Questions for the AI Gateway team
- Which xAI model does
spacexai/grok-sttcall today, and did that change on 18 Sep? - Is there a way to pin the version, or could you add versioned ids (1.0 and 2.0)? We need to control when the model changes.
- When the upstream model changes, will the model page or the changelog say so?
I can share generation ids and details privately if that helps.