Microsoft's MAI-Transcribe-2 Cuts AI Transcription Price 72%

Microsoft AI dropped the price of automatic transcription by 72 percent this morning, cutting MAI-Transcribe-2 to 10 cents per hour of processed audio and claiming it now transcribes speech faster than anything OpenAI, Google or ElevenLabs currently sells. The model went live today through Microsoft Foundry and the MAI Playground, the enterprise and developer channels Microsoft has been using all year to push its own AI models out from under OpenAI's shadow.

The number that matters most is the bill. MAI-Transcribe-1, the model this replaces, charged 36 cents an hour. A call center processing 100,000 hours of recorded audio a year sees that line item drop from $36,000 to $10,000. That's the kind of cut that gets a CFO's attention regardless of how the benchmarks shake out.

RelatedGemini 3.7 Flash Ships at Half Price, 3.5 Pro Still Missing

• New price: $0.10 per hour of audio, down from $0.36, a roughly 72 percent reduction.
• Speed claim: 10x faster than OpenAI's GPT-Transcribe, 7x faster than ElevenLabs' Scribe v2, 5x faster than Google's Gemini 3.5 Transcribe, per Artificial Analysis evaluations.
• Accuracy: 5.2% average word error rate across 60 supported languages on the FLEURS benchmark, good for second place on Artificial Analysis's word-error-rate leaderboard.
• Where it lives: Microsoft Foundry and the MAI Playground today, with a path into Copilot, Teams, GitHub and Dynamics 365 Contact Centre.

What actually changed between MAI-Transcribe-1 and 2?

Three things moved at once. Price fell from 36 cents to 10 cents an hour. Language coverage jumped to 60, up from 43 in June and 25 back in April, so this is the third language expansion in five months, not a one-time bump. And the model picked up a set of features enterprises specifically ask for: speaker diarization, word-level timestamps, keyword biasing so it catches product names and jargon correctly, automatic language detection, code-switching for mixed languages like Hinglish and Spanglish, and a choice between verbatim and cleaned-up output styling.

None of that is exotic on its own. What's notable is Microsoft shipping all of it under a price that undercuts the field rather than matching it. Transcription has quietly become one of the more commoditized corners of applied AI, and commoditized markets get won on cost as much as capability.

How much faster is it, really?

Microsoft's speed numbers come from Artificial Analysis, a third-party benchmarking outfit, not from Microsoft's own lab, which gives them more weight than a vendor-reported figure. The claim: MAI-Transcribe-2 processes audio 10 times faster than OpenAI's GPT-Transcribe, 7 times faster than ElevenLabs' Scribe v2, and 5 times faster than Google's Gemini 3.5 Transcribe. On accuracy, it lands in second place on Artificial Analysis's word-error-rate leaderboard at 5.2 percent average WER across all 60 languages on the FLEURS benchmark, a standard multilingual speech test.

Why is Microsoft racing to undercut its own biggest partner?

Microsoft remains OpenAI's largest investor and its primary cloud provider, which makes it strange on the surface that Microsoft's in-house MAI division keeps shipping models that compete directly with OpenAI's API business. The explanation traces back to October 2025, when Microsoft and OpenAI restructured their partnership. Under the new terms, Microsoft won the right to pursue advanced AI independently, without needing OpenAI's permission or technology. A month later, Microsoft stood up a dedicated MAI Superintelligence team under Mustafa Suleyman, the former Google DeepMind and Inflection AI executive who now runs Microsoft AI.

Suleyman has been explicit that his mandate isn't chasing benchmark headlines for their own sake. He's framed the goal around whether the models "deliver product value for the millions of enterprises that depend on us." That's a pointedly commercial definition of progress, and MAI-Transcribe-2 fits it: cheaper transcription is not a research breakthrough, it's a line-item cut that a Teams admin or a Dynamics 365 customer can act on immediately.

1. Oct 2025Microsoft and OpenAI restructure their partnership. Microsoft gains the right to pursue advanced AI independently or with other partners.
2. Nov 2025MAI Superintelligence team formed. Mustafa Suleyman leads a dedicated in-house model group inside Microsoft AI.
3. Apr 2026MAI-Transcribe reaches 25 languages. Part of Microsoft's first wave of foundational in-house models.
4. Jun 2026Language coverage expands to 43. Second expansion in three months.
5. Jul 27, 2026MAI-Cyber-1-Flash security model ships. A separate MAI model aimed at security workloads, part of the same in-house push.
6. Sep 3, 2026MAI-Transcribe-2 launches at $0.10/hr. 60 languages, 72% cheaper than its predecessor, positioned against OpenAI, Google and ElevenLabs.

What does this mean for the market?

Transcription pricing is a small line item compared to frontier model spending, but it's a real, recurring one for the call centers, meeting-transcription tools and captioning services that buy it by the hour. Microsoft undercutting the field by 70-plus percent puts direct pressure on OpenAI's Whisper-descended API revenue and on ElevenLabs, a company whose entire business is voice AI and that just got outrun on both price and speed by its biggest platform partner's own labs. Google's Gemini 3.5 Transcribe takes the same hit, though Google can subsidize speech products across a much larger cloud business than ElevenLabs can.

RelatedMicrosoft Coaches Sales Team to Undercut OpenAI, Anthropic

For Microsoft, the signal for investors is less about transcription revenue itself and more about vertical integration. Every MAI model that ships is one less workload Microsoft has to route through, and pay OpenAI for, inside Azure. Wiring MAI-Transcribe-2 straight into Copilot, Teams and Dynamics 365 Contact Centre keeps that margin in-house rather than passing it to a partner Microsoft is simultaneously bankrolling. Watch whether Microsoft starts disclosing what share of Copilot's voice and transcription features run on MAI models versus OpenAI's, since that ratio is the real measure of how fast this independence push is actually moving.

What's still missing from Microsoft's numbers?

The launch price has no stated end date and no published standard rate, so enterprises building on it don't yet know what they'll pay once the introductory window closes. Microsoft hasn't detailed real-time streaming latency, which matters for live captioning and call-center use cases as much as batch throughput does. Per-language accuracy breakdowns aren't public, so the 5.2 percent average word error rate could hide much worse performance on lower-resource languages. And for regulated industries specifically, Microsoft hasn't said what its data retention and training practices are for audio processed through MAI-Transcribe-2, a gap that matters more here than it would for a general-purpose chatbot.

Our take

MAI-Transcribe-2 isn't a research story. Nobody is going to remember it as a benchmark milestone the way they might remember a new frontier reasoning model. But it's a clean example of what Microsoft's post-restructuring relationship with OpenAI actually looks like in practice: not a dramatic breakup, just Microsoft quietly building cheaper, faster in-house alternatives for the workloads it can profitably bring home, one modality at a time. Transcription was an easy one to start with, since the task is well-defined and the cost savings are easy to prove. Voice and image generation, the other two legs of this MAI stack, are messier problems with more room for a price cut to matter less than raw quality. Whether Microsoft can pull off the same trick there will say more about this strategy's staying power than today's transcription numbers do.

Original analysis by GenZTech, based on Microsoft AI's own benchmark disclosures and reported pricing.