September 7, 2026. Microsoft put a speech to text model on the market last week at 10 dollars per 100 hours of audio, and the number reprices every call recording, sales call review and AI receptionist transcript your business generates. On September 3 Microsoft AI launched MAI-Transcribe-2 at 0.10 dollars per hour of audio. Two things the launch post does not lead with decide whether you should move: the price is temporary, and the model is in public preview.
What the launch post says and leaves out
- The price is 0.10 dollars per hour, until the end of the year. Microsoft's own qualifier is a limited time offer until the end of the year. It sets no calendar date and publishes no successor rate anywhere.
- It is a preview, not a product. Microsoft Learn's page for the model, dated September 3, says the feature is in public preview, is provided without a service level agreement, and is not recommended for production workloads. The launch post does not mention any of that.
- The accuracy claim shrinks between the headline and the body. The page title calls it the most accurate model in the world. The body says it ranks first on FLEURS across 60 languages at a 5.2 percent average word error rate, and second on the Artificial Analysis word error rate leaderboard.
- No Azure billable meter exists yet. Azure's pricing table carries a row for it with the value masked, and Microsoft's retail price API returns no meter for it. The only published billable rate sits on OpenRouter, the distribution channel Microsoft's own post names.
What everyone else charges per hour of audio
We priced nine vendors from their own pages on September 7 and converted everything to dollars per hour of audio. Rev.ai's Reverb turbo also sits at 0.10, though it bills a 15 second minimum per job, which wrecks the rate on short clips. Speechmatics batch starts at 0.129, AssemblyAI at 0.15 for its older Universal-2 model and 0.21 for the current flagship, Azure batch at 0.18, ElevenLabs Scribe v2 at 0.22, Deepgram Nova-3 at 0.258, OpenAI's gpt-transcribe at 0.27, and AWS batch at 0.36.
The spread inside a single vendor is wider than the spread between vendors, and that is where the money is. Azure charges 1.00 dollars per hour for real time transcription and 0.18 for batch: the same service, 5.5 times the price, purely for delivery mode. Google charges 0.016 dollars per minute for speech to text with data logging and 0.024 without it, so opting out of having your audio used costs you 50 percent more. Deepgram's headline streaming rates are struck through promotional prices; the regular rate is 60 percent higher. Speechmatics advertises from 0.129, which is the cheapest of its five models, while standard batch is 0.24.
What it means for operators
If you transcribe sales calls, the honest move is to stop paying real time rates for work that is not real time. Reviewing yesterday's calls, scoring them, or feeding a CRM summary are all batch jobs, and on Azure alone that single reclassification is an 82 percent cut. Only live agent assist and live captioning need the streaming meter.
If you are quoting an AI receptionist or an outbound calling build, transcription is now a rounding error against the per minute cost of the voice agent itself, which we priced across eleven vendors in our AI receptionist cost breakdown. Do not let a vendor bundle transcription into a per minute rate without showing you the split. For home services operators comparing an AI calling setup against an answering service, the transcript is the cheapest part of the stack and the easiest to over pay for.
And do not migrate production workloads onto a preview model priced as a promotion. Microsoft has published neither an end date nor a replacement rate, and its nearest comparable published Azure rate is 0.36 dollars per hour. Run it in parallel, measure the word error rate on your own audio, and keep the incumbent contract alive until Microsoft prints a real price. If voice is core to your funnel, our AI voice agent builds treat the transcription layer as swappable for exactly this reason.
Frequently Asked Questions
Priced from vendor pages on September 7, 2026, the range runs from 0.10 dollars per hour to about 1.00. Microsoft's MAI-Transcribe-2 and Rev.ai Reverb turbo sit at 0.10, Speechmatics batch at 0.129, AssemblyAI at 0.15 to 0.21, Azure batch at 0.18, ElevenLabs Scribe v2 at 0.22, Deepgram Nova-3 at 0.258, OpenAI gpt-transcribe at 0.27 and AWS batch at 0.36. Azure real time transcription is 1.00.
Microsoft's own documentation says no. The Microsoft Learn page dated September 3, 2026 states the feature is in public preview, provided without a service level agreement, and not recommended for production workloads. The launch announcement does not repeat that warning.
Microsoft describes it as a limited time offer until the end of the year but publishes no specific expiry date and no successor rate. Azure has not published a billable meter for the model. The nearest published comparable on the same API surface is Fast Transcription at 0.36 dollars per hour, though Microsoft has not said the price will land there.
You are paying for latency, not accuracy. Azure charges 1.00 dollars per hour for real time and 0.18 for batch on the same service, a 5.5 times gap. Most business transcription work, including call review, CRM summaries and quality scoring, is not time sensitive and can run on the batch meter.