SpaceXAI ships Grok Voice Transcribe 2.0, claiming twice the accuracy of 1.0 at the same price
Batch stays $0.10 an hour, streaming $0.20, with diarization included. The company cites first place among 32 streaming models on Artificial Analysis. The Next Web: “twice as accurate” is versus its own predecessor, and the credentials eval set is unexplained.
Deep dive3 minIn-depth
James Okonkwo3 min read
- SpaceXAI
- Grok Voice
- Speech-to-text
- Atlassian
- Tesla

Photo: SpaceXAI / xAISource page
SpaceXAI released Grok Voice Transcribe 2.0 on September 18, the speech-to-text model it says is twice as accurate as Grok Voice Transcribe 1.0 at the same price. Batch work stays $0.10 per hour of audio, streaming $0.20, with speaker diarization, word-level timestamps and key-term biasing included. The announcement is on x.ai under the SpaceXAI name the lab has used since the SpaceX acquisition. The Next Web, writing on Saturday, dated the ship to the 18th and noted that coverage still often says xAI. Existing Speech-to-Text API customers get the new weights with no code change. Version 1.0 will be deprecated in the coming weeks; anyone who needs it must pin `grok-voice-transcribe-1.0`.
Read the accuracy claim the way The Next Web did. “Twice as accurate” is versus Transcribe 1.0, not versus ElevenLabs or Deepgram. Against the field, SpaceXAI says the model ranks first for accuracy among 32 streaming models on the public Artificial Analysis leaderboard, and calls itself “one of the most accurate” rather than the most accurate. Internal charts, which the company ran, also put it ahead of ElevenLabs Scribe v2 and Deepgram Nova-3 on telephony. Those are vendor evaluations against vendor-chosen comparators. The public leaderboard is the narrower, named claim.
The audio already running through Grok Voice
Transcribe 2.0 is built on the audio foundation behind Grok Voice. SpaceXAI wrote that Grok Voice already powers tens of thousands of customer-support calls a day, transcribes millions of hours of video narration, and runs voice agents in physical products, including the Grok assistant in Tesla vehicles. The training set is described as live, noisy, multilingual audio recorded across a diverse set of environments, then refined with post-training. Clean studio speech is no longer the hard problem. The hard problem is flaky phone lines, overlapping talkers, local accents, and numbers or addresses read aloud. The company measures word error rate on four internal sets drawn from production traffic: 8 kHz customer-support telephony, conversations with Grok, spoken credentials (account codes, phone numbers, emails, addresses), and short multilingual voice-assistant phrases.
The largest published gain is on those short phrases. On a 19-language set of in-car-style commands, word error rate falls from 20.6 percent to 6.8 percent. The model detects language automatically and follows mid-recording switches in a single pass. SpaceXAI says multilingual accuracy is the biggest jump over 1.0. That is also the Tesla use case: a two-word climate command gives the model almost no context to guess the language. The Next Web put the commercial point next to that chart. At ten cents an hour, transcription is priced as a commodity. What is scarce is the production audio used to train and score it.
We’ve always believed the best way to move work forward is to capture context once and let it flow everywhere. With Grok powering Loom’s speech-to-text and Cursor turning that into code, we’re closing the loop from context to code: record what you mean, and the work gets done.
— Sanchan Saxena, SVP of Teamwork Collection, Atlassian, in SpaceXAI’s Transcribe 2.0 announcement, Sep 18, 2026
A named customer, and a corpus the blog does not explain
Atlassian is the proof point SpaceXAI named. The company said it found Transcribe 2.0 more accurate than Loom’s previous speech-to-text supplier and is switching. Saxena’s quote is about a workflow: record an action plan in Loom, pipe the transcript into Cursor, let the editor make the code change. That is a real buyer, which is worth more than a leaderboard row. The feature list around it is the usual production kit: batch files and URLs, real-time streams, up to eight independent channels, 100 domain terms per request, written-form numbers and dates, optional filler-word stripping, and end-of-turn detection for voice agents.
The Next Web’s caveat is the credentials set. SpaceXAI lists spoken account codes and email addresses as a standing evaluation corpus drawn from production traffic. The blog does not say how that audio is obtained, retained or consented to. It may sit inside enterprise terms. It does not say. A support call has two parties; only one of them signed with the vendor. Voice recordings are personal data in Europe. The Next Web did not accuse the company of a violation. It said the announcement leaves the obvious questions unanswered. That is the part of Saturday’s coverage that is not a price chart.


