SpaceXAI Introduces Grok Voice Transcribe 2.0

In an official blog post on September 18, 2026, xAI released Grok Voice Transcribe 2.0, positioning it as its "latest speech-to-text model" and claiming "unmatched accuracy and cost efficiency." The core quantitative claim from the company is that in its real-world evaluations, 2.0 is twice as accurate as Grok Voice Transcribe 1.0 while the price stays the same.

This is not an isolated product update but a switch of the default path: the company states explicitly that 2.0 will soon become the default version of the speech-to-text API, that 1.0 will be deprecated within the coming weeks, and that developers who want to keep using the old version during the transition must explicitly pin grok-voice-transcribe-1.0. For teams already integrated with the API, this means the accuracy improvement can be obtained without code changes, but it also means the old version's exit has a clear time window.

The public evidence cited by the company is the Artificial Analysis leaderboard: Grok Voice Transcribe 2.0 ranks first in accuracy among 32 streaming models. Note that this is xAI's citation of a third-party leaderboard result; this article has not independently verified the leaderboard's evaluation methodology, timing, or model set.

On internal evaluations, xAI says 2.0 improves across the board over 1.0 on four datasets drawn from production traffic: customer service phone audio (8 kHz, English), conversations with Grok (English), read-aloud credentials such as account codes and email addresses (English), and short spoken commands in 19 languages. The company specifically notes that on phone audio, 2.0 leads all models it has tested.

The number that best reflects the "real-world audio" positioning is the short-command set: word error rate fell from 20.6% in 1.0 to 6.8%. The company explains that short phrases like in-car commands give the model very little language-identification context, so the room for improvement is greatest there; it also says multilingual is the biggest improvement of 2.0 over 1.0, with the ability to transcribe dozens of languages, automatically detect language, and follow language switches mid-recording within a single pass.

On features, 2.0 supports both batch and streaming modes, can transcribe recorded files and URLs, and can transcribe audio streams in real time; it provides word-level timestamps (with start and end times and confidence), speaker diarization (at no extra charge), independent transcription of up to 8 channels, keyword biasing of up to 100 domain terms per request, written formatting of numbers/dates/currency/phone numbers/email addresses, filler-word removal, and smart turn detection for voice agents.

Pricing is the most easily verifiable constraint of this release: the company says 2.0 is priced exactly the same as 1.0, at $0.10 per hour of audio for batch and $0.20 per hour of audio for streaming, with speaker diarization, timestamps, and keywords all included. The company also says 2.0 is built on the audio foundation model that powers Grok Voice, trained on real, noisy, multilingual audio recorded across a variety of environments and then fine-tuned with post-training.

The customer case provided by the company is Atlassian Loom: Atlassian found that Grok Voice Transcribe 2.0 was more accurate than its existing solution when transcribing Loom videos. The company also describes a workflow—recording an action plan in Loom, feeding the transcript into Cursor, and having the latter directly complete code updates; a quote from Sanchan Saxena, Atlassian's Senior Vice President of Team Collaboration Products, says this "closes the loop from context to code."

Facts and claims need to be distinguished: the accuracy comparison above, the leaderboard ranking, the internal evaluation numbers, and the "twice the accuracy" claim all come from xAI's official blog and are source claims; this article has not independently verified them. Loom's "more accurate" is likewise a customer statement relayed by the company, with no specific metrics or comparison baseline given.

In addition, the company says Grok Voice already powers tens of thousands of customer service calls per day, transcribes millions of hours of video voiceover, and runs voice agents in physical products (including the Grok assistant in Tesla vehicles). These scale figures also come from the company and lack third-party verification; and for the specific date of 1.0's deprecation and the exact timing of the 2.0 default switch, the company says only "soon" and "the coming weeks," leaving timing uncertainty.

Pulling the above together, 2.0's competitive narrative is not about "clean single-speaker audio" but about the hardest real-world scenarios: unstable phone lines, multiple people talking over each other, regional accents, and read-aloud phone numbers or email addresses. The company says most transcription models perform well on clean audio, while 2.0 was built for the hardest audio—this is the core of its positioning and the reason its internal evaluations chose datasets such as customer service calls and read-aloud credentials.

For developers, there are two actionable conclusions: first, the accuracy improvement takes effect for existing speech-to-text API integrations without code changes; second, if you depend on 1.0's behavior, you need to pin the old model version before deprecation. As for whether claims like "twice the accuracy" and "first on the leaderboard" can be reproduced on your own data, that still depends on each team's actual audio distribution; the company's data can only serve as a reference, not a guarantee.