SpaceXAI released Grok Voice Transcribe 2.0 on September 18, a speech-to-text model that the company says is twice as accurate as its predecessor while holding the same price: $0.10 per hour of audio for batch transcription and $0.20 per hour for streaming. On the company's hardest internal test, short voice-assistant commands spoken across 19 languages, the word error rate fell from 20.6% to 6.8%, a cut of roughly two-thirds. The model, the first major speech release under the SpaceXAI name since xAI was folded into SpaceX and rebranded in July, debuted at the top of the Artificial Analysis streaming accuracy leaderboard, ranking first among 32 models.
The company announced the model with characteristic bravado on its X account: “Introducing Grok Voice Transcribe 2.0. It's the world's most accurate speech transcription model.” The independent benchmarking firm Artificial Analysis offered a narrower but still striking endorsement the same day, writing that the model was “taking the #1 spot for Final Transcript accuracy and First Partial Transcript accuracy on AA-WER Streaming with 2.7% WER at 0.49s after end of speech.” On the firm's non-streaming benchmark, the model scored a 2.3% word error rate, down from 4.0% for version 1.0.
Transcribe 2.0 is built on the same audio foundation model that runs Grok Voice, which SpaceXAI says already handles tens of thousands of customer-support calls a day, transcribes millions of hours of video narration, and powers the Grok assistant inside Tesla vehicles. The company says the model was trained on live, noisy, multilingual audio recorded across a wide range of environments rather than clean studio clips, then refined with post-training. That focus shows up in the four internal evaluation sets SpaceXAI published alongside the public leaderboard result: 8 kHz telephony audio from English customer-support calls, conversations with Grok, spoken credentials such as account codes and email addresses, and the multilingual short-phrase set. The company says version 2.0 improves on 1.0 across all four, and on telephony it beat every model tested, a list that according to its published charts included Gemini 3.5 Transcribe, MAI-Transcribe-2, ElevenLabs Scribe v2, Deepgram Nova-3, and Whisper Large v3. On the telephony set, the first final transcript came in at 2.7% WER, down from 3.9%.
SpaceXAI is bundling features that rivals often sell as add-ons. Word-level timestamps with confidence scores, speaker diarization, multichannel transcription of up to eight channels, key-term biasing of up to 100 domain terms per request, written-form formatting of numbers and email addresses, filler-word removal, and smart turn detection for voice agents are all included at the base rate. The model is API-only, reachable through the REST endpoint for batch jobs and a WebSocket endpoint at api.x.ai/v1/stt for streaming, with interim results emitted roughly every 500 milliseconds. There are no open weights. Existing integrations get the accuracy gains without code changes, and SpaceXAI says 2.0 will become the API default shortly, with version 1.0 deprecated in the coming weeks. Teams that need the old behavior can pin grok-voice-transcribe-1.0 during the transition.
SpaceXAI's marquee customer for the launch is Atlassian, which evaluated the model against its incumbent vendor and now uses it to transcribe every video on Loom, its screen-recording product. “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,” said Sanchan Saxena, senior vice president of Teamwork Collection at Atlassian. “It's a glimpse of where AI-assisted development is headed.”
Why It Matters
The pricing is the real story. Speech-to-text has been a commodity market for years, but the floor keeps dropping, and SpaceXAI is now well below the incumbents on list price. Deepgram's Nova-3 runs about $0.0043 per minute for pre-recorded audio, or roughly $0.26 an hour, and $0.0077 per minute for streaming, roughly $0.46 an hour. AssemblyAI's batch tier starts at $0.15 an hour on its published pricing, with its Universal-3.5 Pro real-time offering listed at $0.45 an hour. ElevenLabs' Scribe v2 is listed at about $0.22 an hour for batch and $0.39 for real-time, and OpenAI's gpt-4o-mini-transcribe works out to about $0.18 an hour. Against that field, Grok Voice Transcribe 2.0 at $0.10 batch and $0.20 streaming, with diarization and timestamps thrown in, is somewhere between 30% and 60% cheaper depending on the comparison, and SpaceXAI is claiming better accuracy on top of that. For a customer-support operation transcribing a million hours of calls a year, the difference between $0.10 and $0.26 an hour is $160,000.
The multilingual result deserves attention on its own. A 20.6% word error rate on short commands is the kind of number that makes voice interfaces feel broken; 6.8% is usable. Short in-car utterances give a model almost no context to guess the language, and the improvement is a reminder that the audio foundation model powering Tesla's in-vehicle Grok assistant has a very direct commercial reason to get this right.
The independent leaderboard result is the kind of validation SpaceXAI has struggled to earn for its language models, where Grok has often trailed OpenAI, Anthropic, and Google on third-party evaluations. In speech, the company can now point to a public number one. Whether that holds depends on how quickly Deepgram, AssemblyAI, and ElevenLabs respond.
What to Watch
The deprecation of Transcribe 1.0 in the coming weeks will be the first real test of whether the accuracy gains hold up under full production load and across the long tail of languages beyond the 19 in the short-phrase set. Watch also for whether competitors cut prices to match the $0.10 batch rate, and for any independent verification of the telephony numbers, which so far rest entirely on SpaceXAI's internal test sets. The one thing SpaceXAI has not offered is open weights, which leaves on-premises and privacy-sensitive buyers, a meaningful slice of the call-center and healthcare transcription market, still shopping elsewhere.
"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, Senior Vice President of Teamwork Collection, Atlassian