Microsoft AI on Thursday launched MAI-Transcribe-2, a speech-recognition mannequin the corporate says is quicker, extra correct, and cheaper than something OpenAI, Google, or ElevenLabs at the moment sells. Then it priced the factor at 10 cents per hour of audio.
That determine deserves a pause. When Microsoft AI shipped the primary mannequin on this line simply 5 months in the past, it charged $0.36 an hour. Thursday’s early-bird worth cuts that by roughly 72%. For an enterprise processing 100,000 hours of call-center audio a yr — a modest quantity for a big financial institution or telecom — the invoice drops from $36,000 to $10,000. At that stage, transcription stops being a line merchandise anybody argues about.
The discharge arrives as Microsoft executes a technique that may have appeared implausible two years in the past: constructing its personal frontier-class fashions one modality at a time, then steadily swapping them into merchandise that after ran on OpenAI’s expertise. Transcription is the modality the place that plan has moved quickest, and MAI-Transcribe-2 is its clearest proof level but. It additionally presents a preview of how the world’s Most worthy software program firm intends to compete in AI with out relying on the companion it spent $13 billion to domesticate.
What MAI-Transcribe-2 does and why the function record issues to enterprise patrons
The mannequin transcribes audio in 60 languages, up from 43 in June’s MAI-Transcribe-1.5 and 25 in April’s authentic launch. It runs on Microsoft Foundry, the corporate’s mannequin market for builders, and in MAI Playground, its testing surroundings. Microsoft says it constructed the mannequin for the messy audio that actual companies generate — background noise, low-quality recordings, overlapping speech — somewhat than clear studio circumstances.
Extra vital than the language rely is what Microsoft has bundled into the bottom product. Speaker diarization kinds out who stated what in a multi-person recording, which is the distinction between a wall of textual content and a usable assembly transcript. Phrase-level timestamps connect a exact time marker to each phrase, enabling search, modifying, and alignment with video. Key phrase biasing lets builders feed the mannequin a listing of drug names, product codes, or worker names so it stops mangling area jargon. Automated language identification means customers not should declare the language upfront.
Two options stand out for his or her specificity. A configurable output model presents a “verbatim” mode that preserves each “um,” false begin, and stutter for compliance and authorized groups, alongside a “clear” mode that strips fillers for readable captions and notes. And code switching handles conversations that drift between languages mid-sentence; Microsoft explicitly names Hinglish and Spanglish, a nod to the Indian and U.S. Hispanic markets the place a single customer-service name would possibly toggle languages a dozen occasions. Specialty distributors have traditionally charged premiums for every of those capabilities. Microsoft is together with all of them for a dime.
Methods to learn Microsoft’s FLEURS and Synthetic Evaluation benchmark claims
Microsoft makes three efficiency claims, every resting on a distinct measuring stick, and technical patrons ought to perceive what every one captures and what it misses.
The primary is that MAI-Transcribe-2 ranks primary on FLEURS throughout 60 languages with a mean phrase error fee of 5.2%. FLEURS is a benchmark Google researchers printed in 2022, constructed from native audio system studying roughly 2,000 sentences in every of 102 languages — about 12 hours of speech per language. It’s the usual yardstick for multilingual speech recognition as a result of it enables you to examine a mannequin’s Swahili in opposition to its Swedish on equivalent content material. Phrase error fee, its metric, merely counts substitutions, insertions, and deletions in opposition to a human reference; 5.2% means roughly one phrase in 20 is mistaken. However FLEURS is learn speech, not dialog, and Microsoft’s common has really risen from the three.7% it reported for MAI-Transcribe-1.5 in June. That nearly definitely displays broader protection somewhat than regression — averaging throughout 60 languages as an alternative of 43 means folding in low-resource languages the place each mannequin struggles — however patrons ought to request the per-language breakdown.
The second declare is that the mannequin ranks second on the Synthetic Evaluation word-error-rate leaderboard and defines that agency’s accuracy-latency Pareto frontier. Synthetic Evaluation is an impartial benchmarker that assessments fashions by their public APIs, measuring what a buyer really will get. Its index blends simulated agent conversations, European Parliament speeches, and company earnings calls, weighting closely towards English enterprise speech. In June, the agency ranked MAI-Transcribe-1.5 third at 2.4% WER, behind Alibaba’s Enjoyable-Realtime-ASR-preview and ElevenLabs’ Scribe v2, whereas calling it the quickest mannequin within the high 10. Climbing to second suggests Microsoft has cleared ElevenLabs. “Pareto frontier” is the phrase practitioners ought to be aware: it means no rival beats the mannequin on accuracy with out being slower, and none beats it on velocity with out being much less correct.
The third declare is uncooked velocity — 10 occasions sooner than OpenAI’s GPT-Transcribe, seven occasions sooner than ElevenLabs’ Scribe v2, 5 occasions sooner than Google’s Gemini 3.5 Transcribe, per Synthetic Evaluation evaluations. In batch transcription, velocity issues much less as a result of anybody is ready and extra as a result of throughput is value. A mannequin operating at 300 occasions real-time wants a fraction of the GPU-hours of 1 operating at 30 occasions. That effectivity is what lets Microsoft cost a dime and, presumably, nonetheless earn a living.
Three speech fashions in 5 months: inside Microsoft AI’s speedy launch cadence
The tempo is the story inside the story. On April 2, MAI-Transcribe-1 launched with 25 languages at $0.36 per hour. On June 2, MAI-Transcribe-1.5 arrived with 43 languages, key phrase biasing, and a third-place rating on Synthetic Evaluation. Right this moment, MAI-Transcribe-2 shipped with 60 languages, diarization, timestamps, code switching, a second-place rating, and a worth of $0.10.
Three releases in 5 months, every increasing language protection by roughly 40% whereas including options opponents gate behind premium tiers. That cadence is attribute of a workforce that has settled on a secure structure and is now turning the crank on knowledge and scale — the part the place speech fashions have a tendency to enhance rapidly and predictably. It’s also the cadence of an organization that intends to make transcription a commodity earlier than anybody else can.
The organizational wager behind that velocity is one Mustafa Suleyman, Microsoft AI’s chief government, described to The Verge in April. He credited the primary mannequin to “a small, centered 10-person workforce” that had been “liberated from any of the forms,” with a bigger surrounding group dealing with vendor administration and knowledge acquisition.
He additionally advised The Verge the mannequin ran at “half the GPU value of the opposite state-of-the-art fashions,” calling it “an enormous cost-saving” for Microsoft. Meta, Amazon, Google, and Anthropic have all experimented with comparable flattened buildings, The Verge famous. Microsoft’s transcription line is essentially the most seen check but of whether or not the method produces business outcomes somewhat than analysis papers.
Why Microsoft is constructing its personal AI fashions regardless of its $13 billion OpenAI wager
Microsoft has invested greater than $13 billion in OpenAI, and hosts OpenAI’s fashions throughout Azure, Workplace, and Copilot. For a lot of the previous 4 years, the apparent query about any Microsoft-built mannequin has been: why hassle? The reply has sharpened over the previous yr, and it begins with independence.
When Microsoft employed Suleyman from Inflection AI in March 2024, together with most of Inflection’s employees, Salesforce CEO Marc Benioff learn it as a declaration of intent. “Microsoft is constructing their very own AI and I do not assume Microsoft will use OpenAI sooner or later. They will have their very own frontier fashions,” Benioff advised CNBC in January 2025. “That is why they employed Mustafa Suleyman.” Benioff had his personal motives — Salesforce competes with Microsoft and invests in Anthropic — however occasions have largely borne him out.
In October 2025, Microsoft and OpenAI restructured their partnership in a deal that, per Microsoft’s personal announcement, allowed Microsoft to “independently pursue AGI alone or in partnership with third events” for the primary time. Suleyman advised The Verge that renegotiation “unlocked [Microsoft’s] capability to pursue superintelligence,” and Microsoft introduced its MAI Superintelligence workforce weeks later. In April 2026, the businesses amended the deal once more, ending Microsoft’s unique entry to OpenAI’s fashions and eliminating Microsoft’s revenue-share funds, in response to reviews on the time. Every modification loosened the tie. Each was adopted by extra MAI fashions.
How Microsoft’s in-house fashions are slicing prices throughout Groups, Phrase, and Excel
The second half of the reply is margin. Each immediate Microsoft routes to an OpenAI mannequin carries a value. Each immediate it routes to its personal mannequin by itself GPUs carries a smaller one. In July, Bloomberg reported that Microsoft had begun utilizing MAI fashions to reply a portion of person prompts in Phrase and Excel — merchandise it had beforehand marketed as powered by OpenAI and Anthropic. TechCrunch framed the shift as a part of a broader {industry} pullback on AI spending, with Amazon, Uber, Meta, and Accenture all reportedly trimming.
Transcription is the pure first goal for this substitution as a result of the issue is bounded and the metric is goal. Microsoft owns Groups, which generates an unlimited quantity of assembly audio. It owns Nuance, whose scientific documentation enterprise runs on speech recognition. It owns the Azure speech companies that 1000’s of enterprises already name. Each a type of workloads is a candidate to maneuver onto MAI-Transcribe-2, and each hour that strikes is an hour Microsoft not pays anybody else for.
Suleyman has been unusually candid that that is the purpose. Superintelligence, he advised The Verge in April, “is de facto about, ‘Are these fashions able to delivering product worth for the hundreds of thousands of enterprises that rely on us to ship world-class language fashions?'” No matter one thinks of making use of the phrase “superintelligence” to a transcription API, the business logic is apparent: construct the aptitude as soon as, deploy it throughout a dozen merchandise, and cease writing checks to a companion that’s more and more a competitor.
MAI-Transcribe-2 vs. OpenAI, Google, and ElevenLabs: the aggressive image
Microsoft’s launch names 4 rivals: OpenAI’s GPT-Transcribe, Google’s Gemini 3.5 Transcribe, OpenAI’s older Whisper V3-Massive, and ElevenLabs’ Scribe v2. It doesn’t point out Deepgram, AssemblyAI, Speechmatics, or Rev — the specialists which have offered transcription to enterprises for a decade. Microsoft is positioning in opposition to the frontier labs, not the incumbents.
That framing is partly advertising and marketing and partly true. The frontier labs have handled speech as a checkbox function of broader platforms, priced accordingly, and a devoted mannequin that beats them on velocity by 5 to 10 occasions whereas matching their accuracy is a real differentiator. However the specialists will really feel the worth stress most acutely. At $0.10 an hour, Microsoft is pricing at or beneath the place a lot of them promote high-volume enterprise contracts, and it’s bundling diarization, timestamps, and 60 languages into the bottom fee. The specialists’ remaining moat is area depth — medical vocabularies, authorized formatting, industry-specific integrations — and Microsoft’s key phrase biasing function is aimed squarely at it.
The one competitor Microsoft conspicuously doesn’t declare to beat on accuracy is Alibaba, whose fashions have posted main numbers on impartial leaderboards for a lot of 2026. TechCrunch reported in July that some U.S. corporations had begun evaluating Chinese language fashions as cheaper options regardless of safety issues. Microsoft’s pitch to these patrons is implicit however unmistakable: comparable accuracy, sooner inference, lower cost, and a vendor your compliance workforce already trusts.
The questions technical choice makers ought to ask earlier than switching transcription distributors
For all its specificity on benchmarks, the discharge leaves a number of sensible questions open. The primary is length: Microsoft calls $0.10 per hour a launch supply with out naming an finish date or an ordinary fee, and anybody constructing a value mannequin ought to get each in writing. The second is streaming. The discharge emphasizes batch throughput and long-form audio however says nothing about real-time transcription, which voice brokers and reside captioning require. Synthetic Evaluation maintains a separate streaming leaderboard, and Microsoft’s silence on it’s notable.
The third is per-language accuracy. A 5.2% common throughout 60 languages may imply 3% on main languages and 12% on low-resource ones, so patrons with particular wants ought to check these languages instantly. The fourth is diarization high quality. Phrase error fee doesn’t measure speaker attribution; a transcript can have near-perfect WER and nonetheless assign each different sentence to the mistaken particular person. The discharge presents no diarization error fee or comparable metric.
The fifth is knowledge dealing with. Enterprise transcription touches medical data, authorized privilege, and monetary disclosures, and the discharge says nothing about knowledge residency, retention, or whether or not audio submitted to Foundry feeds future coaching. Microsoft’s April bulletins described coaching knowledge as a mixture of human-curated recordings, contractor-recorded noisy audio, and “huge quantities of knowledge from the open internet,” per The Verge — an outline that ought to immediate pointed questions from regulated industries. None of those gaps is uncommon for a launch announcement, however they’re precisely the questions that separate a leaderboard win from a manufacturing deployment.
What Microsoft’s speech mannequin technique reveals about its broader AI ambitions
Step again from the speech-recognition particulars and a sample emerges that extends effectively past transcription. Microsoft’s AI unit now ships fashions for pictures, voice, transcription, code, reasoning, and cybersecurity. At Construct in June, it introduced seven new MAI fashions in a singlekeynote. Every follows the identical playbook: goal a well-defined modality, optimize aggressively for inference value, worth beneath the frontier labs, distribute by Foundry, and quietly swap the mannequin into Microsoft’s personal merchandise.
This isn’t an try to construct one mannequin that beats GPT or Gemini at every thing. It’s an try to construct a portfolio of specialised fashions that, in combination, let Microsoft serve most of its enterprise workloads with out paying anybody else — and to promote the excess capability to everybody else at costs the specialists can’t match. Transcription occurred to be the primary modality the place the method totally matured, however the launch notes for MAI-Transcribe-2 learn much less like a product announcement than a template.
Suleyman has spent two years speaking about “humanist superintelligence” and AI assistants which might be “accountable to them, on their aspect.” The vocabulary is lofty. The execution is a spreadsheet. 5 months in the past, Microsoft charged 36 cents to show an hour of speech into textual content. On Thursday it charged a dime, threw in six options its rivals promote individually, and claimed the highest spot on the {industry}’s commonplace multilingual benchmark. The corporate that spent $13 billion studying what frontier AI prices has determined it could somewhat personal the manufacturing facility than lease the output — and now it’s promoting the output for lower than the lease.
MAI-Transcribe-2 is on the market now by Microsoft Foundry and MAI Playground.
