On September 3, Microsoft released MAI-Transcribe-2, a speech-to-text model that undercuts OpenAI, Google, and ElevenLabs on both price and accuracy at the same time. That combination almost never happens in AI pricing. Usually you get one or the other. Here’s what it actually changes if you’re sitting on hours of video or audio you’ve never done anything with.
What actually shipped#
MAI-Transcribe-2 launched at $0.10 per audio hour, a steep drop from the $0.36 per hour MAI-Transcribe-1 launched at. Independent benchmarking from Artificial Analysis puts it at a 2.0% word error rate, processing audio at roughly 410 times real-time speed, which works out to about $1.67 per 1,000 minutes transcribed. Microsoft’s own reported figure across all 60 supported languages is a 5.2% average word error rate, with 3.4% across its top 25 languages. It adds speaker diarization, word-level timestamps, and domain keyword biasing, so it can tell your speakers apart and pick up on industry-specific terms it would otherwise mishear.
Two details matter more than the headline speed claim. First, it handles code-switching, conversations that shift between languages mid-sentence, the kind of thing that trips up most transcription models completely. Microsoft’s own examples include Hinglish and Spanglish specifically, so if your customer base or your team routinely mixes languages on a call, this is one of the first models actually built to follow that instead of garbling it. Second, output isn’t one-size-fits-all: you can pull a “verbatim” transcript that captures every filler word and false start, useful for legal or compliance records where the exact words matter, or a “clean” version with the ums and false starts stripped out, useful for anything you’re about to publish or hand to someone else to read.
On raw speed, Microsoft’s own comparisons put it at 10 times faster than GPT-Transcribe, 7 times faster than ElevenLabs’ Scribe v2, and 5 times faster than Gemini 3.5 Transcribe, while leading Artificial Analysis’s own accuracy-versus-speed frontier. Microsoft is naming specific use cases it’s optimized for: clinical note-taking, legal documentation, accessibility services, and closed captioning, all fields where getting the words exactly right matters as much as getting them fast. It’s available now through Microsoft Foundry, the MAI Playground, and Open Router.
The catch worth knowing before you build anything on it#
Here’s the part that matters if you’re the kind of founder who reads past the headline: that $0.10 per hour price is an explicitly limited-time promotional rate through the end of 2026. Microsoft has not published what the permanent price will be. The accuracy and speed numbers are real and independently verified. The price is a launch offer, not a guarantee.
That doesn’t make this any less worth using right now. It means you shouldn’t build a business model that assumes $0.10 an hour is permanent. Use it. Just don’t bet your unit economics on a promotional rate holding forever.
Practically, that means running your backlog now, while the promotional price is live, rather than treating this as a someday project. It also means if you’re wiring transcription into an ongoing product feature rather than a one-time cleanup, price it out at something closer to the old $0.36 rate too, so a price change at the end of the year doesn’t blow up a cost model you built assuming the launch price was permanent.
Why this actually matters#
Transcription has quietly been one of the most expensive bottlenecks in content repurposing. Turning one podcast episode, sales call, or YouTube video into blog posts, social clips, and searchable text used to mean either paying real money per hour of audio or accepting mediocre accuracy to keep costs down. At 2% word error rate and a fraction of a cent per minute, that tradeoff mostly disappears.
Put an actual number on it. A hundred hours of recorded sales calls, at the old MAI-Transcribe-1 rate, would have run you $36. At the current promotional rate, the same hundred hours costs $10. That’s not a rounding difference, it’s the kind of gap that changes whether transcribing your entire call archive is a project you greenlight or one you keep deferring. If you’ve got a backlog measured in hundreds of hours across sales calls, webinars, and old video, this is the month that backlog stops being expensive to unlock.
If you’ve been sitting on a backlog of recorded calls, videos, or podcast episodes you never repurposed because transcription felt like a tax on doing it, that excuse is gone. The content you’ve already created is now cheap to turn into searchable text, which means it’s cheap to turn into more content, more SEO surface area, and more places a potential customer can find you.
That’s the real unlock here, not the novelty of a new model. Cheap, accurate transcription turns everything you’ve already recorded into raw material instead of a dead file sitting in storage. And the code-switching and speaker diarization features specifically matter if you’re an international team or you sell across markets where calls don’t happen in a single clean language — a lot of transcription tooling quietly assumes single-language, single-accent audio, and breaks down exactly at the calls that would tell you the most.
What to actually do with this#
Pull your highest-value recorded content first: sales calls that reveal real customer language, workshop or webinar recordings, anything long-form you made once and never touched again. Transcribe it, then look for the moments where you explained something clearly and specifically. Those moments are usually better written content than anything you’d draft from scratch, because they came from a real conversation, not a blank page.
Decide your output style before you run the batch, not after. If you’re building a searchable internal archive or need a compliance-grade record, pull verbatim transcripts and keep every false start. If you’re pulling quotes for a blog post or a social clip, pull clean transcripts so you’re not editing filler words out of every single line by hand. Doing this once at the start saves a second pass through everything you transcribe.
Use the word-level timestamps for more than search. They let you jump straight to the exact five-second clip where a customer said the thing you want to quote, instead of scrubbing through a forty-minute recording by ear. That’s the actual time savings here, not just the transcription cost, it’s the hours you stop spending hunting for the moment you remember happened but can’t pinpoint.
If content repurposing is part of how you’re trying to turn attention into customers, the same operator discipline applies whether the channel is a podcast, a sales call archive, or YouTube. I’ve written the full system for turning YouTube specifically into a customer-acquisition channel in Built the Product. Now Get Customers., if that’s the channel you’re building toward.
What’s sitting in your recordings right now that you never got around to repurposing?