On September 3, 2026, Microsoft AI released MAI-Transcribe-2, a speech-recognition model the company says is faster, more accurate, and cheaper than competing offerings from OpenAI, Google, and ElevenLabs. The model is priced at $0.10 per hour of audio, a roughly 72% cut from the $0.36 per hour Microsoft charged for the original MAI-Transcribe-1 just five months ago.

For an enterprise processing 100,000 hours of call-center audio a year, the annual bill drops from $36,000 to $10,000. The model is available now on Microsoft Foundry, MAI Playground, and OpenRouter, and is also exposed through Azure AI Speech services in public preview.

What's new

  • Price: $0.10 per audio hour (early-bird pricing), down from $0.36 for MAI-Transcribe-1.
  • Languages: 60 languages supported, up from 43 in MAI-Transcribe-1.5 (June 2026) and 25 in the April 2026 original release.
  • Benchmarks: Microsoft claims first place on the FLEURS multilingual benchmark across 60 languages with an average word error rate (WER) of 5.2%, and second place on the Artificial Analysis WER leaderboard. The company also says MAI-Transcribe-2 defines the Artificial Analysis accuracy-latency Pareto frontier.
  • Speed claims (vendor-cited, per Artificial Analysis): 10x faster than OpenAI's GPT-Transcribe, 7x faster than ElevenLabs Scribe v2, 5x faster than Google Gemini 3.5 Transcribe.
  • New features: Speaker diarization, word-level timestamps, keyword biasing for domain jargon, automatic language identification, code-switching for mixed-language audio (e.g., Hinglish, Spanglish), and configurable transcription styles (verbatim vs. clean).
  • Availability: Microsoft Foundry, MAI Playground, OpenRouter, and Azure AI Speech (public preview, not recommended for production workloads per Microsoft's own documentation).

Context: three releases in five months

The cadence is notable. MAI-Transcribe-1 launched April 2, 2026 with 25 languages at $0.36/hour. MAI-Transcribe-1.5 arrived June 2, 2026 with 43 languages, keyword biasing, and a third-place Artificial Analysis ranking. MAI-Transcribe-2 now ships with 60 languages, diarization, timestamps, code-switching, a second-place ranking, and the $0.10 price point. Each release expanded language coverage by roughly 40% while adding features competitors typically gate behind premium tiers.

Microsoft AI chief Mustafa Suleyman credited the first model to a "small, focused 10-person team" liberated from bureaucracy, with a larger group handling vendor management and data acquisition. He also said the model ran at "half the GPU cost of the other state-of-the-art models." That architecture stability is now letting the team turn the crank on data and scale — the phase where speech models tend to improve quickly and predictably.

Why it matters

In our read, the pricing and feature bundle appears to put direct pressure on specialist transcription vendors such as Deepgram and AssemblyAI, who have historically charged premiums for diarization, word-level timestamps, and keyword biasing — Microsoft is including all of them in the base price. The move also reads as part of a broader strategy: after restructuring its OpenAI partnership in October 2025 and again in April 2026 — ending exclusive access and revenue-share payments — Microsoft appears to be steadily swapping its own MAI models into products that once ran on OpenAI technology. On this reading, transcription looks like the first modality where that substitution is moving at commercial speed, and Microsoft's ownership of distribution channels (Teams, Nuance, Azure Speech Services) gives it the option to make the switch at scale. These are analytical judgments, not outcomes the sources confirm.

Confirmed

  • Launch and price: MAI-Transcribe-2 was announced September 3, 2026 at $0.10 per audio hour (early-bird pricing), down from $0.36 per hour for MAI-Transcribe-1 five months earlier.
  • Features as announced: Speaker diarization, word-level timestamps, keyword biasing, automatic language identification, code-switching for mixed-language audio, and configurable transcription styles, across 60 supported languages.
  • Availability: Microsoft Foundry, MAI Playground, and OpenRouter now; Azure AI Speech in public preview, which Microsoft's own documentation says is not recommended for production workloads.
  • Release history: MAI-Transcribe-1 shipped April 2, 2026 with 25 languages at $0.36/hour; MAI-Transcribe-1.5 arrived June 2, 2026 with 43 languages and a third-place Artificial Analysis ranking.
  • Vendor-reported benchmarks: Microsoft reports a 5.2% average WER and first place on FLEURS across 60 languages, second place on the Artificial Analysis WER leaderboard, and speed multiples of 10x vs. OpenAI's GPT-Transcribe, 7x vs. ElevenLabs Scribe v2, and 5x vs. Google Gemini 3.5 Transcribe — figures cited to Artificial Analysis in Microsoft's announcement.

Unknown

  • Independent verification: The accuracy rankings and speed multiples trace to Microsoft's announcement and vendor-cited Artificial Analysis figures; this coverage includes no third-party benchmark results.
  • Low-resource language performance: The 5.2% average WER across 60 languages sits above 1.5's 3.7% average across 43, implying weaker per-language results somewhere — but a per-language breakdown isn't provided in this coverage.
  • Real-world savings: Whether the claimed speed multiples and Suleyman's "half the GPU cost" statement translate into equivalent customer savings depends on individual hardware and Azure GPU quota constraints, which the sources don't address.
  • Early-bird terms: How long the $0.10 early-bird price lasts and what standard pricing follows isn't detailed in this coverage.

Our take

Microsoft is commoditizing the transcription primitive before rivals can lock in enterprise contracts. The $0.10 price signals that internal GPU economics — half the cost of peer models, per Suleyman — are real enough to sustain margin while undercutting the market. What remains unproven is whether the FLEURS average (5.2% WER across 60 languages) holds up on the low-resource languages that dragged the average up from 1.5's 3.7% on 43 languages, and whether the Artificial Analysis speed multiples translate to equivalent savings on Azure GPU quotas for customers not on Microsoft's preferred hardware.

Sources