Microsoft's New AI Model Is Built to Replace OpenAI, Anthropic

Open Copilot on your phone today and ask it to transcribe a voice note, and there's a good chance the words on your screen never touched a model from OpenAI or Anthropic. Microsoft built that transcription engine itself, and it did so specifically to stop paying its AI rivals for the privilege.

A speech model built to prove a point

The model is called MAI-Transcribe-2, and Microsoft quietly pushed it into public preview in September, calling it the fastest transcription system it has shipped. Independent benchmarking from Artificial Analysis put its word error rate — essentially how often it mishears or mistranscribes a word — at 2.0% across 60 languages, among the best scores on that board. More strikingly, Microsoft says it processes audio at roughly 410 times real-time speed, turning an hour-long recording into clean text in about ten seconds, priced at just $0.10 per hour of audio.

Microsoft claims that's roughly ten times faster than OpenAI's comparable transcription model, about seven times faster than ElevenLabs' Scribe v2, and nearly five times faster than Google's Gemini transcription tools. It's the third speech model Microsoft has shipped in five months, part of a run of seven in-house "MAI" models — spanning coding, reasoning, image generation and voice — that the company unveiled at its Build conference earlier this year.

The real reason: Microsoft is tired of the bill

None of this is happening in a vacuum. For most of the last two years, Copilot — Microsoft's flagship AI assistant across Windows, Office and Bing — has leaned heavily on models licensed from OpenAI and, more recently, Anthropic. That arrangement works, but it isn't cheap, and Microsoft has started routing everyday requests in apps like Excel and Outlook to its own MAI models instead, even when they aren't quite as capable as the premium alternative.

Mustafa Suleyman, the former DeepMind co-founder who now runs Microsoft's AI division, said the quiet part out loud earlier this year.

"We pay a lot of money to Anthropic – so our goal is to reduce and ultimately eliminate that cost," Suleyman said, laying out Microsoft's thinking in blunt terms.

It helps that Microsoft badly needs a win. By its own numbers, Copilot had around 23 million monthly active users between April and June this year — respectable, but dwarfed by ChatGPT's roughly 400 million weekly users and Gemini's reported 650 million monthly users over the same stretch. Building cheaper in-house models won't close that adoption gap on its own, but it buys Microsoft room to keep bolting Copilot onto every product it sells without the per-query cost spiraling.

Why Indian BPOs and startups should care

This isn't just a boardroom story out of Redmond. India runs the world's largest business process outsourcing industry, and a huge chunk of that work — call centre quality audits, customer support logs, meeting transcriptions for enterprise clients — depends on converting speech to text at massive scale, often in English layered with Hindi or regional accents. For an Indian BPO processing millions of call-minutes a month, a transcription model that's several times faster and a fraction of the cost of the going rate isn't a nice-to-have; it shows up directly on the margin line.

The same logic applies to Indian startups building voice products, from vernacular-language support bots to accessibility tools, who've so far had to budget around OpenAI or Google pricing. If Microsoft's bet pays off and pushes rivals to cut prices in response, the way OpenAI already did earlier this year with its own API pricing, that's a direct, measurable saving for anyone building on these platforms out of Bengaluru, Gurugram or Hyderabad.

A cheaper model isn't always the better one

There's a catch, and Microsoft doesn't hide it: its in-house models generally trail the best that OpenAI and Anthropic can offer on raw capability. Independent testing has placed some MAI models closer to open-source alternatives than to frontier models like Anthropic's latest Claude releases — the same lab that recently opened an India office and courted local partnerships. Microsoft is making a conscious trade: good enough, at a fraction of the price, for the huge volume of routine requests that don't need a frontier model's full power.

• MAI-Transcribe-2: $0.10 per audio hour, 2.0% word error rate, 60 languages supported
• Roughly 410x real-time processing speed, per Microsoft's own testing
• Third Microsoft speech model shipped in five months
• Part of seven in-house MAI models spanning coding, reasoning, voice and images

Whether that trade-off holds as Microsoft pushes MAI models deeper into Copilot's coding and reasoning features — the parts of the product people actually judge it on — is the real test still to come. For now, the message to OpenAI and Anthropic is clear enough: Microsoft would rather ship something slightly less capable and fully its own than keep paying someone else's invoice.