Microsoft AI released MAI-Transcribe-2 on 3 September, claiming first place on the FLEURS multilingual benchmark across 60 languages with a 5.2 percent average word error rate, a tenfold speed advantage over OpenAI's GPT-Transcribe on long recordings, and a price of 10 cents per audio hour through the end of the year.1 Microsoft AI 2026-09-03 First on FLEURS across 60 languages at 5.2 percent WER; second on the Artificial Analysis leaderboard; 10 times faster than GPT-Transcribe, 7 times faster than Scribe v2, 5 times faster than Gemini 3.5 Transcribe; 10 cents per hour through year end; Foundry, MAI Playground and OpenRouter. Open source The price is a 72 percent cut from the 36 cents Microsoft charged for MAI-Transcribe-1 in April.2 VentureBeat 2026-09-03 Price cut from 36 cents to 10 cents per hour in five months; 100,000 hours drops from 36,000 to 10,000 dollars; languages from 25 to 43 to 60; partnership restructuring in October 2025 and April 2026; MAI models replacing OpenAI models in Word and Excel per Bloomberg; 10 person team; specialists omitted from the comparison. Open source Speech to text is a commodity market, so the numbers matter less than the target: this is Microsoft's own lab shipping a cheaper substitute for an OpenAI product five months after the two companies ended exclusive access and revenue sharing.2 VentureBeat 2026-09-03 Price cut from 36 cents to 10 cents per hour in five months; 100,000 hours drops from 36,000 to 10,000 dollars; languages from 25 to 43 to 60; partnership restructuring in October 2025 and April 2026; MAI models replacing OpenAI models in Word and Excel per Bloomberg; 10 person team; specialists omitted from the comparison. Open source Our assessment, with moderate confidence, is that the 10 cent price is a market share play that will not survive 2026 intact, and that the durable news is the cadence: three speech models in five months from a small team, aimed squarely at the partner Microsoft no longer needs.

What was released

The model adds speaker diarization, word level timestamps, keyword biasing for domain terms, verbatim and clean output modes, automatic language identification and code switching for mixed speech.1 Microsoft AI 2026-09-03 First on FLEURS across 60 languages at 5.2 percent WER; second on the Artificial Analysis leaderboard; 10 times faster than GPT-Transcribe, 7 times faster than Scribe v2, 5 times faster than Gemini 3.5 Transcribe; 10 cents per hour through year end; Foundry, MAI Playground and OpenRouter. Open source Language coverage has grown from 25 in April to 43 in June to 60 now.2 VentureBeat 2026-09-03 Price cut from 36 cents to 10 cents per hour in five months; 100,000 hours drops from 36,000 to 10,000 dollars; languages from 25 to 43 to 60; partnership restructuring in October 2025 and April 2026; MAI models replacing OpenAI models in Word and Excel per Bloomberg; 10 person team; specialists omitted from the comparison. Open source On Artificial Analysis, the model ranks second on word error rate and sits on what Microsoft calls the accuracy and latency Pareto frontier.1 Microsoft AI 2026-09-03 First on FLEURS across 60 languages at 5.2 percent WER; second on the Artificial Analysis leaderboard; 10 times faster than GPT-Transcribe, 7 times faster than Scribe v2, 5 times faster than Gemini 3.5 Transcribe; 10 cents per hour through year end; Foundry, MAI Playground and OpenRouter. Open source Unite.AI reports a measured speed factor of 403.6, meaning roughly 10 seconds to process an hour of audio.3 Unite.AI 2026-09-03 Speed factor of 403.6, about 10 seconds per hour of audio; 2.0 percent error rate on Artificial Analysis; public preview with no SLA and not recommended for production; 300 megabyte file limit; Azure subscription required. Open source

Two caveats attach to the headline numbers. The 5.2 percent FLEURS figure is an average across 60 languages, and Microsoft's own June model posted 3.7 percent across a smaller set of 43, so the number rises as coverage widens and is not directly comparable to the earlier release.2 VentureBeat 2026-09-03 Price cut from 36 cents to 10 cents per hour in five months; 100,000 hours drops from 36,000 to 10,000 dollars; languages from 25 to 43 to 60; partnership restructuring in October 2025 and April 2026; MAI models replacing OpenAI models in Word and Excel per Bloomberg; 10 person team; specialists omitted from the comparison. Open source And the speed multiples are Microsoft's claims citing Artificial Analysis, not measurements this outlet has verified.1 Microsoft AI 2026-09-03 First on FLEURS across 60 languages at 5.2 percent WER; second on the Artificial Analysis leaderboard; 10 times faster than GPT-Transcribe, 7 times faster than Scribe v2, 5 times faster than Gemini 3.5 Transcribe; 10 cents per hour through year end; Foundry, MAI Playground and OpenRouter. Open source The comparison set is also selective: VentureBeat notes the announcement omits Deepgram, AssemblyAI, Speechmatics and Rev, the companies that have specialized in transcription for years.2 VentureBeat 2026-09-03 Price cut from 36 cents to 10 cents per hour in five months; 100,000 hours drops from 36,000 to 10,000 dollars; languages from 25 to 43 to 60; partnership restructuring in October 2025 and April 2026; MAI models replacing OpenAI models in Word and Excel per Bloomberg; 10 person team; specialists omitted from the comparison. Open source

The price is the product

At 10 cents an hour, an enterprise processing 100,000 hours a year pays 10,000 dollars instead of 36,000.2 VentureBeat 2026-09-03 Price cut from 36 cents to 10 cents per hour in five months; 100,000 hours drops from 36,000 to 10,000 dollars; languages from 25 to 43 to 60; partnership restructuring in October 2025 and April 2026; MAI models replacing OpenAI models in Word and Excel per Bloomberg; 10 person team; specialists omitted from the comparison. Open source That is small money for a large company and large money for a transcription vendor whose entire business is that line item. Microsoft has said the price is limited to the end of 2026 and has not published the rate that follows.1 Microsoft AI 2026-09-03 First on FLEURS across 60 languages at 5.2 percent WER; second on the Artificial Analysis leaderboard; 10 times faster than GPT-Transcribe, 7 times faster than Scribe v2, 5 times faster than Gemini 3.5 Transcribe; 10 cents per hour through year end; Foundry, MAI Playground and OpenRouter. Open source A promotional price with no disclosed successor is a standard way to pull workloads onto a platform and reprice once the integration is done.

The availability terms reinforce that reading. The Azure Speech integration is a public preview with no service level agreement and is described as not recommended for production; files are capped at 300 megabytes and use requires an Azure subscription and a Foundry resource.3 Unite.AI 2026-09-03 Speed factor of 403.6, about 10 seconds per hour of audio; 2.0 percent error rate on Artificial Analysis; public preview with no SLA and not recommended for production; 300 megabyte file limit; Azure subscription required. Open source The model is also on OpenRouter and the MAI Playground.1 Microsoft AI 2026-09-03 First on FLEURS across 60 languages at 5.2 percent WER; second on the Artificial Analysis leaderboard; 10 times faster than GPT-Transcribe, 7 times faster than Scribe v2, 5 times faster than Gemini 3.5 Transcribe; 10 cents per hour through year end; Foundry, MAI Playground and OpenRouter. Open source The production path runs through Azure, which is where Microsoft wants the workload to land.

Who gains and who loses

Microsoft gains a first party speech model that undercuts the OpenAI product it used to resell, and a proof point for the argument that a 10 person team can ship competitive models on a monthly cadence.2 VentureBeat 2026-09-03 Price cut from 36 cents to 10 cents per hour in five months; 100,000 hours drops from 36,000 to 10,000 dollars; languages from 25 to 43 to 60; partnership restructuring in October 2025 and April 2026; MAI models replacing OpenAI models in Word and Excel per Bloomberg; 10 person team; specialists omitted from the comparison. Open source Azure customers gain a cheaper transcription line for the rest of the year. Developers building on OpenRouter gain another interchangeable backend.

OpenAI loses a distribution channel it once counted on: VentureBeat reports Bloomberg found Microsoft already replacing OpenAI models with MAI models inside Word and Excel in July, and transcription now follows the same pattern.2 VentureBeat 2026-09-03 Price cut from 36 cents to 10 cents per hour in five months; 100,000 hours drops from 36,000 to 10,000 dollars; languages from 25 to 43 to 60; partnership restructuring in October 2025 and April 2026; MAI models replacing OpenAI models in Word and Excel per Bloomberg; 10 person team; specialists omitted from the comparison. Open source ElevenLabs and Google, named in the speed comparisons, face a price they will have to answer or explain.1 Microsoft AI 2026-09-03 First on FLEURS across 60 languages at 5.2 percent WER; second on the Artificial Analysis leaderboard; 10 times faster than GPT-Transcribe, 7 times faster than Scribe v2, 5 times faster than Gemini 3.5 Transcribe; 10 cents per hour through year end; Foundry, MAI Playground and OpenRouter. Open source The specialist transcription vendors lose most: their margins depend on per hour pricing that a hyperscaler can set below cost as a customer acquisition expense.

The counter case

The thesis that this is aimed at OpenAI could overstate the strategic intent. Speech to text is a natural first product for a new lab because the benchmarks are clean, the training data is abundant and the market is price sensitive; Microsoft may simply be shipping what its team can ship. The quality claims also rest on Microsoft's own framing of third party leaderboards, and the model ranks second, not first, on the Artificial Analysis error rate table.1 Microsoft AI 2026-09-03 First on FLEURS across 60 languages at 5.2 percent WER; second on the Artificial Analysis leaderboard; 10 times faster than GPT-Transcribe, 7 times faster than Scribe v2, 5 times faster than Gemini 3.5 Transcribe; 10 cents per hour through year end; Foundry, MAI Playground and OpenRouter. Open source If independent evaluations on noisy, domain specific audio show the specialists holding an accuracy edge, the 10 cent price buys volume without displacing the incumbents in clinical and legal work, where errors are expensive.3 Unite.AI 2026-09-03 Speed factor of 403.6, about 10 seconds per hour of audio; 2.0 percent error rate on Artificial Analysis; public preview with no SLA and not recommended for production; 300 megabyte file limit; Azure subscription required. Open source And a preview product with no SLA is not yet a production competitor, whatever its price.

What to watch

  • The post promotional price. If Microsoft publishes a 2027 rate at or below 20 cents an hour by December, the market share reading holds; a return toward 36 cents would mark the launch as a promotion, not a repricing.1 Microsoft AI 2026-09-03 First on FLEURS across 60 languages at 5.2 percent WER; second on the Artificial Analysis leaderboard; 10 times faster than GPT-Transcribe, 7 times faster than Scribe v2, 5 times faster than Gemini 3.5 Transcribe; 10 cents per hour through year end; Foundry, MAI Playground and OpenRouter. Open source
  • General availability with an SLA. Azure Speech moving MAI-Transcribe-2 from preview to general availability within two quarters would make it a production competitor; a longer preview would suggest the model is a showcase.3 Unite.AI 2026-09-03 Speed factor of 403.6, about 10 seconds per hour of audio; 2.0 percent error rate on Artificial Analysis; public preview with no SLA and not recommended for production; 300 megabyte file limit; Azure subscription required. Open source
  • Competitor price responses. A price cut from OpenAI, Google or ElevenLabs on transcription by the end of 2026 would confirm the pressure; no response would suggest they are ceding the low end.2 VentureBeat 2026-09-03 Price cut from 36 cents to 10 cents per hour in five months; 100,000 hours drops from 36,000 to 10,000 dollars; languages from 25 to 43 to 60; partnership restructuring in October 2025 and April 2026; MAI models replacing OpenAI models in Word and Excel per Bloomberg; 10 person team; specialists omitted from the comparison. Open source
  • The next MAI product. Microsoft has shipped three speech models in five months. A fourth model in a category OpenAI still supplies to Microsoft, by early 2027, would confirm the substitution pattern beyond transcription.2 VentureBeat 2026-09-03 Price cut from 36 cents to 10 cents per hour in five months; 100,000 hours drops from 36,000 to 10,000 dollars; languages from 25 to 43 to 60; partnership restructuring in October 2025 and April 2026; MAI models replacing OpenAI models in Word and Excel per Bloomberg; 10 person team; specialists omitted from the comparison. Open source

The 10 cent hour is a bid for workloads, not a price. The forward question is which OpenAI product Microsoft's lab replaces next.