xAI has released Grok Voice Transcribe 2.0, a speech-to-text model the company says is twice as accurate as its predecessor at exactly the same price. The operative detail for developers is the clock: version 1.0 will be deprecated in the coming weeks, and 2.0 is set to become the default in the Speech-to-Text API whether production test suites are ready or not.
xAI Watch analysis
What happened
The announcement on x.ai went up on 18 September. Grok Voice Transcribe 2.0 is built on the audio foundation model behind Grok Voice, which xAI says already powers tens of thousands of customer-support calls a day, transcribes millions of hours of video narration and runs the Grok assistant in Tesla vehicles. The new model was trained on live, noisy, multilingual audio from a wide spread of environments, then refined with post-training. In other words, it was raised on the messy stuff rather than studio recordings.
On the benchmark front, xAI says the model ranks first for accuracy among 32 streaming models on the public Artificial Analysis leaderboard. Internally, the company reports improvements over 1.0 across four test sets drawn from production traffic: telephony calls, conversations with Grok, spoken credentials, and short voice-assistant phrases in 19 languages. The short-phrase results are the eye-catcher, with word error rate falling from 20.6% to 6.8%, and multilingual accuracy is described as the largest gain of all. The leaderboard position is the claim outsiders can test; the internal tables are xAI's own measurements on its own audio.
There is a customer story too. Per the announcement, Atlassian found 2.0 more accurate than its existing solution and now uses it to transcribe every video on Loom. Sanchan Saxena, SVP of Teamwork Collection at Atlassian, describes the ambition as closing the loop from context to code: record a plan in Loom, pipe the transcript into Cursor, and let the code updates follow.
Pricing is unchanged: $0.10 per hour of audio for batch, $0.20 for streaming, with diarization, timestamps and key terms included rather than itemised. The transition is the sharp edge. Version 2.0 will soon be the API default, 1.0 will be deprecated in the coming weeks, and pinning grok-voice-transcribe-1.0 is the only stay of execution.
One oddity for the anoraks: the page's masthead and copyright line read SpaceXAI LLC rather than xAI, a name the announcement itself does not explain.
Our top picks
- Batch and streaming
Transcribe recorded files and URLs, or run live audio through in real time. - Word-level timestamps
Every word carries precise start and end times plus a confidence score. - Speaker diarization at no extra cost
Speaker labels are folded into the base price rather than billed as an add-on. - Up to eight channels
Multichannel audio is transcribed independently, channel by channel. - Key term biasing
Pass up to 100 domain terms per request, from product names to medical vocabulary. - Written-form formatting
Numbers, dates, currencies, phone numbers and email addresses come back written out. - Smart turn detection
The model spots when a speaker has finished, which voice agents care about deeply.
Why it matters
Transcription is the unglamorous floor of the voice AI stack, and the floor just moved. If the accuracy claims hold, the cost of a reliable transcript drops for anyone running support lines, voice agents or in-car assistants, because the bill stays flat while the claimed quality doubles. xAI's own comparison charts name ElevenLabs Scribe v2 and Deepgram Nova-3 as the competition, so the pressure lands on rivals who must now answer a held price with something better than a held price.
The Loom anecdote is the strategic tell. Record a meeting, transcribe it, feed the transcript to a coding agent and let the work happen. That pipeline only functions if the transcript is trustworthy, which makes a speech-to-text upgrade quietly an agent-enabler. The deprecation, meanwhile, forces a re-test on every production system, because a model swap can change downstream behaviour even when the pricing does not move.
Our read
Treat 'twice as accurate' as a claim with a receipt attached, not a law of nature: the public leaderboard rank can be checked, the internal numbers cannot. The more impressive move is commercial. Same price, add-ons folded in, and a deprecation clock that pushes every existing customer to re-evaluate on xAI's schedule rather than their own. If you run speech-to-text in production, benchmark 2.0 against your worst audio this week, pin 1.0 only to buy time, and assume the default flips regardless.
What to watch
- Independent runs on the Artificial Analysis leaderboard as third parties put 2.0 through its paces.
- A calendar date for the 1.0 deprecation, replacing 'coming weeks'.
- Whether ElevenLabs and Deepgram answer with price moves or new versions of their own.
- Whether Grok Voice Transcribe 2.0 follows Grok Voice onto third-party hosting such as fal.ai.
Discussion spark: When an API provider deprecates the old model weeks after shipping its replacement, is a pin enough protection for production voice systems, or should providers guarantee old versions a full year of support?
Sources and evidence
- Introducing Grok Voice Transcribe 2.0 | SpaceXAI (18 September 2026, 00:00 UTC)
not affiliated with or endorsed by xAI, SpaceXAI, SpaceX, X Corp. or Cursor