共有:
Microsoft Copilot / MAI-Transcribe-2

Copilot's transcription just got dirt cheap — in-house.

Until last month, Copilot leaned on third-party general-purpose models for transcription, a mixed bag on cost and speed. The newly added in-house "MAI-Transcribe-2" brings a price cut that benefits heavy users of meeting notes and captions the most.

AI Navigate Editorial2026.09.046 min read

UNTIL LAST MONTH Relied on third-party models NOW MAI-Transcribe-2 (in-house) Cost: higher Cost: way down
01
Why Now

Copilot was
"renting" its transcription

Until last month, Microsoft Copilot had no dedicated in-house transcription model, so meeting recordings and live captions ran through general-purpose voice models from other vendors. Those models handle a wide range of tasks well, but without transcription-specific tuning, Copilot had to trade off cost against speed one way or the other.

Microsoft has now added a dedicated transcription model, "MAI-Transcribe-2," to Copilot. As the name suggests, it's a successor to an earlier "MAI-Transcribe-1," and it's a voice-specialized model Microsoft built itself. According to the official Copilot site, it runs cheaper and faster than the existing OpenAI- and Google-based transcription options.

Meeting minutes and live-stream captioning are a textbook case where processing cost scales with recording hours, so per-unit price differences compound fast. Ditching the "rented" general-purpose model for a purpose-built one pays off most for organizations running this kind of workload at scale. Full pricing details are published on the Microsoft Copilot Blog.

02
By the Numbers

How much cheaper
than the alternatives?

Cost per hour of audio, based on Artificial Analysis's transcription benchmark.

COST PER HOUR OF AUDIO OpenAI-based Google-based MAI-Transcribe-2 Up to 40% cheaper per hour
FIG. Hourly transcription cost based on Artificial Analysis's comparison, including figures from Microsoft's own announcement
up to 40%
Lower processing cost vs. competitors
MAI-Transcribe-1
The predecessor this replaces
3 routes
Available via Copilot app, Teams, and API
03
Head to Head

How it stacks up against the major transcription models

ModelApprox. cost per hour
MAI-Transcribe-2Among the cheapest available
OpenAI-based transcription modelsHigher than MAI
Google-based transcription modelsSomewhat higher than MAI

Stop renting the model, and the price drops.


04
Who It's For

Who benefits, and how

Engineers

If you're calling transcription through the Copilot API in your own app, switching the model reference alone should cut costs. Still worth a regression check on accuracy and latency before flipping the switch, especially for jargon-heavy industry use cases where misrecognition rates matter most.

Business / Back Office

Teams that churn through meeting notes and captions at volume benefit most. Anyone using it just a few times a month is unlikely to notice the cost difference at all. Teams with heavy volume may also find this price gap worth raising the next time an annual contract comes up for renewal.

01

Audit your current transcription spend

Tally up what you're currently paying for third-party models and see how much room there is to move to MAI-Transcribe-2.

02

Sample-test accuracy first

Run recordings with jargon and multiple speakers through MAI-Transcribe-2 and compare against your current model before fully migrating.


05
Risks & Limits

Not all upside

The "up to 40% cheaper" figure blends an Artificial Analysis benchmark with Microsoft's own announcement — it isn't a number AI Navigate independently remeasured. The actual cost gap may vary by language, recording quality, and number of speakers.

It also isn't yet clear whether MAI-Transcribe-2 can be called as a standalone API outside the Copilot ecosystem. If you're already calling a competitor's transcription API directly, switching over will require extra integration work through Copilot rather than a simple endpoint swap. Microsoft also hasn't given a clear timeline for how long the predecessor, MAI-Transcribe-1, will keep running alongside it before being retired.