While GPT-6 got the headlines, Microsoft quietly made transcription almost free
MAI-Transcribe-2 costs $0.10 an audio hour — 72% less than its previous model. The unglamorous half of the AI market is where prices are actually falling.
Stated by the maker, publisher, or platform holder.
The same week OpenAI launched a $50-per-million-token flagship, Microsoft went the opposite direction with a much less glamorous product — and it might tell you more about where AI is actually heading for most people.
Microsoft AI released MAI-Transcribe-2, a speech-to-text model priced at $0.10 per audio hour — about 72% cheaper than the $0.36 it charged for the previous version earlier this year. It's the kind of launch that gets no attention next to a new frontier model, and it's arguably more consequential for the average business using AI.
What's confirmed
Diarization and word-level timestamps aren't minor. They're the difference between a wall of text and a usable transcript that says who said what and when — the features that make transcription actually useful for meetings, interviews, and captioning rather than just a novelty.
Why the price matters more than the model
Speech-to-text is one of the few AI markets where the product is close to a commodity — several providers do it well, so the competition happens on price and speed rather than on being the smartest. That's exactly the kind of market where prices fall fastest, and MAI-Transcribe-2 dropping 72% from one version to the next shows how fast.
For a business, $0.10 an audio hour changes what's worth doing. Transcribing every support call, every meeting, every interview was a cost decision at $0.36; at a tenth of that it becomes something you just turn on. That's how falling prices create new uses — not by enabling something impossible, but by making something previously-too-expensive routine.
The split this completes
Put this next to GPT-6 Astra, launched the same week at 2.5x the previous flagship's price, and the shape of the AI market in late 2026 is clear.
Frontier models — the smartest, aimed at the hardest problems — are getting more expensive, because buyers will pay for the best. Commodity models — transcription, cheap chat, routine tasks — are getting dramatically cheaper, because when everyone's product is good enough, price is all that's left to compete on.
There isn't one "AI price trend" anymore. There are two, moving in opposite directions, and which one you feel depends entirely on whether you're buying the cutting edge or the commodity. Most people and most businesses live at the commodity end — which is why a boring transcription price cut may touch more lives than a headline-grabbing frontier launch.