A cheap transcription API can become expensive when a floating model alias changes the words your downstream automation treats as facts.
Grok Voice Transcribe 2.0 is xAI’s new speech-to-text model for batch and streaming audio. xAI made it available on September 17, 2026 and published the launch article on September 18. The company lists $0.10 per audio hour for batch and $0.20 for streaming, with timestamps, confidence values, diarization, eight-channel support and vocabulary biasing.
The price difference is simple
| Mode | Published price | 100 audio hours | Best fit |
|---|---|---|---|
| Batch | $0.10 per audio hour | $10 | Recorded calls, podcasts and archives. |
| Streaming | $0.20 per audio hour | $20 | Live captions, agents and low-latency workflows. |
Those examples cover model usage only. Storage, upload, retries, silence, downstream summarization and human review can cost more than transcription. Measure billable audio duration and failed attempts rather than multiplying file count by average length.
The feature set targets calls and meetings
- Speaker diarization: separates speakers when the source is mixed.
- Eight audio channels: useful when participants are already recorded separately.
- Timestamps and confidence: support review, search and low-confidence escalation.
- Vocabulary biasing: accepts up to 100 terms for names and domain language.
- Language switching: handles multilingual changes within audio.
- Filler removal: can clean disfluencies for readable transcripts.
- Smart Turn: helps a realtime application decide when a speaker has finished.
Channels and diarization solve different problems
If a call platform records each participant on a separate channel, preserve those channels. The model can use explicit separation instead of guessing who spoke from one mixed waveform. Diarization remains useful for recordings where speakers share a channel, but it needs its own accuracy score because attribution errors can be more damaging than word errors.
Our Gemini transcription comparison explains why live and recorded speech are different products. The same test split applies here: evaluate latency and turn detection for streaming, then word accuracy, speaker assignment and timestamps for batch.
The default-model record changed across official pages
xAI’s release notes said Voice Transcribe 1.0 was the default at one point. The launch article said 2.0 would become the default soon. The current voice documentation now says 2.0 is the default. These statements can all be historically accurate, but a floating default makes an unpinned production evaluation hard to reproduce.
A similar naming gap appeared in our Grok Voice Think Fast 2.0 review. The reusable lesson is to treat product pages, release notes and model catalogs as separate records, then pin the exact model the client sends.
| Official record | Wording | Safe interpretation |
|---|---|---|
| Release notes | 1.0 documented as default | Historical state, not proof of today’s route. |
| 2.0 launch article | 2.0 would become default soon | Transition announcement. |
| Current voice docs | 2.0 documented as default | Current documentation, still pin for reproducibility. |
Pin the model and keep a transcript acceptance set
- Select the explicit 2.0 model ID rather than a default alias.
- Build a set with accents, noise, overlapping speech, code-switching and proper names.
- Score words and speaker attribution separately.
- Test all eight channels if the application can send them.
- Include domain terms inside and outside the 100-term bias list.
- Record latency to first partial, finalization delay and correction behavior.
- Keep a rollback path to the prior pinned model.
Treat filler removal as an editorial transformation
Removing filler can improve readability, but it also changes the literal record. Keep the raw audio and, where accuracy or compliance matters, preserve a verbatim transcript alongside the cleaned version. Do not use a polished transcript as the sole record of a legal, medical or employment conversation.
The performance claims are company-reported
xAI publishes word-error-rate comparisons for the launch. Those measurements are useful evidence of what the company tested, not independent proof for every accent, language or acoustic condition. Reproduce the evaluation on recordings that resemble the product’s users before making a quality claim.
The practical verdict
Vocabulary biasing and Smart Turn need adversarial tests rather than a happy-path demo. Put near-sounding product names inside and outside the 100-term list, then check whether biasing improves the intended term without forcing it into unrelated audio. For Smart Turn, include interruptions, long pauses, backchannels such as “uh-huh,” and speakers who restart a sentence. A premature turn decision can make a voice agent answer over the caller; an overly cautious one creates dead air. Track false end-of-turn and missed end-of-turn rates separately from transcription accuracy, because a transcript can be correct after the interaction has already felt broken.
Grok Voice Transcribe 2.0 has an aggressive public price and a feature set shaped for real call workflows. The model-version transition is the operational trap. Pin 2.0, retain raw evidence and test speaker attribution as carefully as the words themselves.
Primary sources
Checked September 20, 2026. Price and benchmark results are xAI-reported; MustHave.ai has not independently reproduced the word-error-rate claims.
