Publication date: 18 September 2026
xAI has made Grok Voice Transcribe 2.0 available through its Speech to Text API, adding a second pinned transcription model alongside version 1.0. The model was first observed in xAI’s public documentation on 18 September 2026 and is available for both file transcription and real-time WebSocket streaming.
The buyer-relevant change is unusually narrow: developers can opt into a new transcription model without paying a higher published list price. xAI still keeps Grok Voice Transcribe 1.0 as the default, so production systems will not move to 2.0 unless they select it explicitly.
Grok Voice Transcribe 2.0 is an opt-in API upgrade
The same `/v1/stt` API now accepts either `grok-voice-transcribe-1.0` or `grok-voice-transcribe-2.0`. For streaming, the model is selected when the WebSocket connection is created. If developers omit the model parameter, xAI continues to use version 1.0.
That default matters operationally. Teams can canary the new model on a slice of calls or recordings while keeping the existing version as a control. It also avoids the risk of an unannounced model swap for integrations that depend on stable transcript behavior.
The surrounding STT feature set includes batch and streaming transcription, interim streaming results, word-level timestamps, multichannel transcription, keyterm prompting and inverse text normalization. Batch files can be up to 500 MB, and multichannel requests support as many as eight channels.
The list price is unchanged, but xAI has not published a 2.0 benchmark
| Decision factor | Transcribe 2.0 | Transcribe 1.0 |
|---|---|---|
| REST transcription | $0.10 per hour | $0.10 per hour |
| Streaming transcription | $0.20 per hour | $0.20 per hour |
| Default when model is omitted | No | Yes |
| Public version-to-version benchmark | Not published | Baseline |
For procurement, the lack of a version premium makes testing straightforward. For engineering, the missing benchmark is more important. xAI has not published a like-for-like 1.0 versus 2.0 accuracy or latency table, so there is no public evidence yet that quantifies the improvement on accents, noise, names or telephony audio.
xAI’s general STT documentation lists 10 requests per second for REST, 10 requests per second for streaming and 100 concurrent streaming sessions per team. The published model page places the service in `us-east-1`. Enterprise customers can request higher limits.
The right comparison is correction effort, not a headline WER
Speech recognition teams should resist treating a new version number as a reason to migrate. The useful test is whether 2.0 reduces the errors that create downstream work: names, product terms, addresses, numbers, overlapping speakers and noisy phone audio.
For teams already evaluating Deepgram, ElevenLabs or other transcription providers, cross-vendor benchmark numbers are only useful when audio, language, codec and scoring rules match. xAI has not supplied that comparison for Transcribe 2.0. A production-shaped A/B test is therefore more decision-relevant than mixing vendor-reported scores.
The published API price is aggressive enough that human correction time can dominate model cost. A model that costs the same but reduces corrections, retries or failed automations can improve total cost even if raw transcription latency changes little.
Privacy and deployment still need procurement review
xAI’s enterprise terms state that user content is normally deleted within 30 days, subject to agreed retention settings, legal requirements and safety or security exceptions. The company also offers Zero Data Retention-enabled APIs under separate conditions. Buyers handling personal or regulated audio should confirm that their chosen STT deployment is covered by the required retention configuration.
The main reason not to switch immediately is evidence, not integration cost. Version 2.0 is selectable and priced like 1.0, but xAI has not published an independent evaluation or a detailed version-to-version quality report. Teams should keep 1.0 pinned until their own acceptance metrics show a meaningful gain.
What this means
Contact-center, meeting-transcription and voice-product teams already using xAI should test Grok Voice Transcribe 2.0 now. Start with recordings that contain known failures, then compare entity accuracy, correction time, latency and retry rates against the pinned 1.0 model.
For new buyers, 2.0 is worth adding to an STT bake-off because the list price is unchanged. It is not yet a reason to switch providers on its own, because the release arrives without a public quality benchmark that establishes the size of the improvement.
For more information, visit the official documentation for Grok Voice Transcribe 2.0 on the xAI website.
