The short version: same price as before, about half the errors on SpaceXAI's own tests, and first place for accuracy among 32 streaming models on an independent leaderboard. It's an API, so you'll only see it in your meeting tool if that tool's maker adopts it. Loom already has. The maker describes the improvement as "twice as accurate as Grok Voice Transcribe 1.0, at the same price" in its September 18 announcement.
SpaceXAI released Grok Voice Transcribe 2.0 on September 18. Speech-to-text launches usually come with a nicer demo and not much else, but this one changes the two numbers people actually pay attention to, accuracy and cost, in the right direction.
According to the launch post, Transcribe 2.0 is built on the same audio foundation model as Grok Voice, which SpaceXAI says already handles tens of thousands of support calls a day and runs the Grok assistant in Tesla cars. Batch transcription stays at $0.10 per hour of audio, streaming at $0.20. Speaker diarization, word timestamps and key-term biasing are included. The September 18 announcement said version 1.0 would be retired in the coming weeks and described pinning grok-voice-transcribe-1.0 during that transition. Use the current API documentation when choosing a version today.
Artificial Analysis measured a 2.7 percent word error rate on the first final transcript in streaming, arriving 0.49 seconds after speech ends, down from 3.9 percent for version 1.0. That's enough to top its streaming accuracy ranking. Speed is another matter: Muse Voice Transcribe and ElevenLabs Scribe v2 Realtime deliver a first transcript faster, with slightly higher error rates. In batch, Transcribe 2.0 ranked fifth of 59 at 2.3 percent, behind Alibaba, StepFun, Microsoft and ElevenLabs Scribe v2.
The vendor claim we'd most like to test is about short multilingual phrases, the "navigate home" kind of command with almost no context to go on. SpaceXAI says the word error rate on that set fell from 20.6 percent to 6.8 percent across 19 languages. If you dictate notes in more than one language, watch that figure more closely than any clean-English leaderboard.
ElevenLabs is in our catalog mainly for generation: voices, and now music. Scribe is its transcription product, and on batch accuracy it still sits ahead of Grok 2.0 in Artificial Analysis's table. Grok's pitch is live, messy audio (phone lines, people talking over each other, account numbers read aloud) at a price that undercuts realtime rivals.
Atlassian is the named customer. SpaceXAI says Atlassian found 2.0 more accurate than its previous setup for transcribing Loom videos, and quotes SVP Sanchan Saxena on a workflow that goes from a Loom recording straight into Cursor. For the rest of us the workflow is shorter: record the meeting, skim the transcript, paste the action items wherever you already keep them.
We haven't run a bake-off on our own recordings. SpaceXAI's internal test sets (telephony, Grok chats, spoken credentials, short multilingual commands) come from production traffic, which is the right kind of material, but they're still the vendor's tests.
Our take: If you stream audio for support lines, live notes or voice commands and pay realtime rates elsewhere, run a trial. Twenty cents an hour is a strong argument, and since the price didn't go up, a week of real calls costs next to nothing to compare. Count names, numbers and mangled product terms.
If your files are clean studio speech and you're happy with Scribe or another top performer in the 59-model table, stay where you are.
And if you were hoping for a new Grok button in your meeting tool, there isn't one. Grok's consumer Voice Mode is a separate product for talking to the chatbot.
One last practical tip. Leaderboards average across datasets that don't include your conference room, your accents or your product names. Key-term biasing takes up to 100 custom words per request, so load it with your company name, your product names and the colleague whose surname every transcriber gets wrong.
