SpaceXAI plans to retire Grok Voice Transcribe 1.0 weeks after shipping a drop-in successor

"Twice as accurate" is SpaceXAI's own summary across four internal test sets ^[15]. 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% ^[14]. SpaceXAI calls that roughly a 67% reduction in errors ^[14]. Divide one rate by the other and errors on that set fell by a factor of three, more than the headline summary claims ^[1]. The other three sets were 8 kHz customer-support calls, English conversations with Grok, and spoken account codes and contact details ^[12]. 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 ^[13][11].

The public evidence is the Artificial Analysis streaming leaderboard, where SpaceXAI says it ranks first for accuracy among 32 systems ^[7]. The Artificial Analysis page currently describes its benchmark as covering 27 of 33 models, five fewer than the company's count ^[8][4]. 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 ^[9]. Network delay is included, and final and partial transcripts are scored separately after speech ends ^[10]. 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 ^[4]. 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 ^[5]. 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 ^[17], the API transcribes up to eight channels independently, and it accepts as many as 100 domain terms per request ^[18]. 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 ^[16], so streaming carries twice the rate ^[2] and 1,000 hours of recorded video costs $100 ^[3]. Loom took on that bill when it started transcribing every video ^[3]. xAI was founded in 2023 with a stated mission to understand "the true nature of the universe" ^[20], and announced a $20 billion Series E in January ^[21]. The shipping product formats dates, currencies and phone numbers ^[19].