Microsoft Cut Transcription Pricing 72%. If You Bill by the Audio Hour, Redo Your Math.
Microsoft AI shipped MAI-Transcribe-2 on September 3 at an introductory price of $0.10 per hour of audio, through Azure Speech's public preview. The previous model, MAI-Transcribe-1/1.5, listed at $0.36 an hour. That's a 72% cut, and it landed with a straight claim from Microsoft: fastest, most accurate, and cheapest speech recognition model available right now. If you're running a podcast tool, a meeting-notes app, a voice-memo transcription side hustle, or anything where transcription cost sits in your cost-of-goods-sold line, your unit economics moved this week whether you touched your stack or not.
What shipped
MAI-Transcribe-2 ranks first on the FLEURS benchmark across 60 languages, with an average word error rate of 5.2%, and adds speaker diarization, configurable transcription styles, and word-level timestamps on top of the raw transcript. Microsoft is positioning it directly against Gemini's transcription model, OpenAI's, Whisper V3-Large, and ElevenLabs's Scribe, claiming it beats all of them on price and speed simultaneously. It's in public preview through Azure Speech, not a stable GA product yet, which matters more than it sounds like it should.
The $0.10/hour rate is explicitly introductory, through December 31, 2026. Microsoft has not said what it costs after that date. That's the detail that should slow down anyone tempted to rebuild a pricing page around this number today.
Do the math before you touch anything
I keep a rough COGS spreadsheet for every API-dependent feature I ship, and transcription is exactly the line item this kind of announcement should send you back to. Say you run a transcription add-on that processes 500 hours of audio a month. At the old $0.36/hour rate, that's $180 in COGS. At $0.10/hour, it's $50, a $130/month swing on volume that isn't even that large. Scale that to a few thousand hours a month, which isn't unusual for a real transcription product with a modest user base, and the swing is real money, the kind that changes whether a freemium tier is sustainable or whether you've been quietly overpricing a paid tier because you built your margin around the old rate.
But here's the part worth sitting with: this pricing has an expiration date, no published successor rate, and lives in a preview product, not a GA SKU with an SLA you'd want to hard-commit a paying customer's product to. If you rearchitect your entire transcription pipeline around a Microsoft preview endpoint's promotional pricing, you're making a bet on Microsoft's post-promo pricing decision in Q1 2027, a decision Microsoft hasn't made yet either, or at least hasn't disclosed.
What I'd do
Don't migrate your production pipeline today. Do redo your pricing spreadsheet today. If your product's pricing page or internal margin model assumes $0.36/hour for transcription COGS, that assumption is now stale, and you should know your actual current cheapest-available-option cost even if you don't switch providers immediately. Separately, if you're prototyping something new that needs transcription, this is a legitimately good moment to build against MAI-Transcribe-2 in preview, get the accuracy and speed benefits now, and just budget for the possibility that the price normalizes upward at the end of the year rather than assuming $0.10/hour is the permanent floor.
The pattern that matters more than this specific number: transcription, like most foundation-model-adjacent commodity APIs, is in a price war between three or four large providers right now. That means the "true cost" of running a transcription feature keeps dropping every few months, and if you built pricing assumptions into a business plan even six months ago, they're probably already wrong in your favor, not against it. Check, don't assume.
The honest take
I think the actual story here isn't "Microsoft has the best model," it's "commodity transcription pricing is in freefall and nobody's updating their spreadsheets to match." The risk I'd flag against my own advice: chasing the cheapest API of the month is how you end up with three transcription integrations in your codebase and none of them well-tested, because you kept switching providers every time someone undercut the last one by a few cents an hour. My actual recommendation is boring: track the market so you know when your assumptions are stale, but don't rebuild anything around a promotional rate with a hard expiration date and no successor price published.
Author
Lukas
@lukcombinator