SpaceXAI plans to retire Grok Voice Transcribe 1.0 weeks after shipping a drop-in successor
SpaceXAI's new transcription model costs the same as the one it replaces and drops into existing code, so the migration work is re-validating accuracy claims that rest mostly on the company's own test sets.
The Engineer · Build desk
What happened
- SpaceXAI made Grok Voice Transcribe 2.0 available through its API on September 17th, according to the release notes, and launched it publicly the following day.
- Atlassian says Loom now uses the model to transcribe every video.
- Recorded audio costs $0.10 per hour and streaming $0.20 per hour, the same rates version 1.0 charged, according to the current SpaceXAI pricing page.
- SpaceXAI plans to make version 2.0 the default and deprecate version 1.0 in the coming weeks, though developers can pin the older model temporarily.
- On short multilingual phrases, SpaceXAI reports word error rate falling from 20.6% to 6.8%, the sharpest gain it disclosed across four internal test sets.
Why it matters
- Who has to decide decision Since the model drops straight in, the only migration task left is running your own audio through both models, and the window for that work closes on a schedule SpaceXAI sets.
- What it costs cost Because the accuracy claim arrives without a price change, a volume buyer's spend still tracks hours of audio, and the new expense is validation labor rather than a higher unit rate.
- What doesn't add up contradiction First among 32 systems and a benchmark covering 27 of 33 models describe two different comparisons, so the denominator behind the rank depends on which page you read.
Next story · Build · 2d ago
Grok Voice 3.0 reaches 86.9 on Artificial Analysis' speech index at the old $0.08 priceConfirmed7 publishers
"Twice as accurate" is SpaceXAI's own summary across four internal test sets [18]. The number under it is a set of short commands in 19 languages, where reported word error rate fell from 20.6% to 6.8% [17]. SpaceXAI calls that roughly a 67% reduction in errors [17]. Divide one rate by the other and errors on that set fell by a factor of three, more than the headline summary claims [24]. The other three sets were 8 kHz customer-support calls, English conversations with Grok, and spoken account codes and contact details [15]. Version 2.0 beat version 1.0 on all four and led every model the company tested on telephony, according to SpaceXAI, on evaluations drawn from its own production traffic [16][14].
The public evidence is the Artificial Analysis streaming leaderboard, where SpaceXAI says it ranks first for accuracy among 32 systems [5]. The Artificial Analysis page currently describes its benchmark as covering 27 of 33 models, five fewer than the company's count [25][26]. The methodology is word error rate over roughly eight hours of audio, with 50% of the score from AA-AgentTalk, a private dataset of speech directed at voice agents, and 25% each from VoxPopuli and Earnings22 [12]. Network delay is included, and final and partial transcripts are scored separately after speech ends [13]. Narrowband call-center audio sits outside the public half of that mix. For first place to predict your own numbers, half your traffic would have to look like people talking to a voice agent over a live connection.
The client stays as it is on the way to 2.0 [3]. That removes the integration work and leaves the behavior check, which is the part that takes calendar time: pull a sample of your own recordings, run both models, compare where names, account codes and numbers land. SpaceXAI plans to make 2.0 the default and deprecate 1.0 within weeks, with temporary pinning for developers who want the old model [4]. Weeks is the only timing the release gives.
The hourly rate covers more than the transcript. Speaker diarization, word-level timestamps, confidence scores and key-term biasing are included [7], the API transcribes up to eight channels independently, and it accepts as many as 100 domain terms per request [8]. Confidence scores are what makes the regression check cheap to automate. The 100-term ceiling is the one spec to check against your vocabulary: a product with thousands of proper nouns has to decide which ones ride along on each request.
Recorded audio is $0.10 an hour and streaming is $0.20, unchanged from 1.0 [6], so streaming carries twice the rate [22] and 1,000 hours of recorded video costs $100 [23]. Loom took on that bill when it started transcribing every video [10]. xAI was founded in 2023 with a stated mission to understand "the true nature of the universe" [19], and announced a $20 billion Series E in January [20]. The shipping product formats dates, currencies and phone numbers [9].
What to watch
- A published cutoff for pinning version 1.0 would turn the validation window into a fixed schedule.
- Artificial Analysis changing its model count or the AA-AgentTalk weighting would change what the first-place ranking supports.
- Per-language word error rates for the 19-language command set would show whether the multilingual gain is broad or concentrated.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence58
- Adoption40
- Hype gap+30
- Incentives65
- Confidence62
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
SpaceXAI released Grok Voice Transcribe 2.0 on September 18th.
- [2]
The model became available through SpaceXAI's API on September 17th, according to the release notes, before the public launch the following day.
- [3]
Existing speech-to-text integrations can move to version 2.0 without code changes.
- [4]
SpaceXAI plans to make version 2.0 the default and deprecate version 1.0 in the coming weeks, though developers can temporarily pin the older model.
- [5]
SpaceXAI says the model ranks first for accuracy among 32 systems on the Artificial Analysis streaming leaderboard.
- [6]
Grok Voice Transcribe 2.0 costs $0.10 per hour for recorded audio and $0.20 per hour for streaming, matching version 1.0; the current SpaceXAI pricing page confirms those rates.
- [7]
Speaker diarization, word-level timestamps, confidence scores and key-term biasing are included.
- [8]
The API can transcribe up to eight channels independently and accept as many as 100 domain terms per request.
- [9]
The API formats dates, currencies and phone numbers.
- [10]
Atlassian says Loom now uses Grok Voice Transcribe 2.0 to transcribe every video.
- [11]
SpaceXAI says Grok Voice Transcribe 2.0 is twice as accurate as version 1.0 at the same price.
- [12]
Artificial Analysis measures word error rate using roughly eight hours of audio, giving 50% of the score to AA-AgentTalk, a private dataset focused on speech directed at voice agents, with VoxPopuli and Earnings22 accounting for 25% each.
- [13]
The Artificial Analysis benchmark includes network delay and separately measures final and partial transcripts after speech ends.
- [14]
The independent leaderboard does not independently establish SpaceXAI's claim that version 2.0 is twice as accurate as version 1.0 across the company's evaluations; SpaceXAI's broader claims about telephony, spoken credentials and multilingual commands rely on internal evaluations drawn from production traffic.
- [15]
SpaceXAI tested version 2.0 on four sets: 8 kHz customer-support calls, English conversations with Grok, spoken account codes and contact details, and short commands across 19 languages.
- [16]
SpaceXAI says version 2.0 beat its predecessor on all four test sets and led every model it tested on telephony.
- [17]
The sharpest disclosed gain came from short multilingual phrases, where reported word error rate fell from 20.6% to 6.8%, roughly a 67% reduction in errors.
- [18]
SpaceXAI uses the broader phrase "twice as accurate" to summarize performance across its evaluations, so developers assessing a particular language or audio environment will still need to test their own recordings.
- [19]
Musk founded xAI in 2023 with a stated mission to understand "the true nature of the universe."
- [20]
xAI has raised successive funding rounds to develop large models and build compute infrastructure, including a $20 billion Series E announced in January.
- [21]
SpaceXAI emphasizes overlapping speakers, narrowband phone calls, regional accents, language switching and structured information read aloud.
- [22]
Streaming transcription is priced at twice the rate of recorded audio.
- [23]
Transcribing 1,000 hours of recorded audio costs $100 at the published rate.
- [24]
On the short multilingual phrase set, word error rate fell by a factor of about three, more than the twofold improvement the company's summary phrase claims.
- [25]
The Artificial Analysis page currently describes its benchmark as covering 27 of 33 models, so the count and ranking should be read as the company's characterization of the comparison.
- [26]
The company's cited comparison set of 32 systems is five larger than the 27 models Artificial Analysis says its benchmark covers.
Sources
2 independent publishers whose own reporting we read for this story.
- runtimewire.comSpaceXAI ships Grok Voice Transcribe 2.0 as Loom transcribes every video
2 articles · September 18, 2026
- testingcatalog.comSpaceXAI says Grok Voice Transcribe 2 doubles accuracy
1 article · September 19, 2026
Topics and entities
Entities
- ElevenLabs
- Elon Musk
- Grok Voice Transcribe 2.0
- Artificial Analysis
- Loom
- Atlassian
- SpaceXAI
- SpaceX
- xAI