Microsoft calls its new transcription model "the fastest, most accurate, and cheapest in the world" -- an independent leaderboard ranks it 2nd on both

Microsoft AI released MAI-Transcribe-2 on September 3, calling it "the fastest, most accurate, and cheapest speech recognition model in the world" in its own headline. The model ranks first on FLEURS, a standard multilingual benchmark, across 60 languages, with an average word error rate of 5.2% by Microsoft's own measurement. It launches at $0.10 per hour of audio -- a 72% cut from the $0.36-per-hour price of MAI-Transcribe-1 -- as a stated limited-time offer through the end of 2026.

That three-part claim -- fastest, most accurate, cheapest -- is the kind of sentence a vendor writes about its own model on launch day. Artificial Analysis, an independent benchmark firm Microsoft's own post cites for context, runs a continuously updated leaderboard that tests transcription models against one shared set of audio rather than each vendor's own selected clips. Checked against that board, MAI-Transcribe-2 ranks second, not first, on both of the two claims it can actually measure.

On word error rate, Artificial Analysis measures MAI-Transcribe-2 at 2.0% -- a different, broader test set than Microsoft's own FLEURS run, so the two numbers aren't directly comparable, but still behind Fun-Realtime-ASR-preview's 1.7%, a smaller model most readers won't have heard of, and just ahead of ElevenLabs' Scribe v2 at 2.2% and Microsoft's own prior model, MAI-Transcribe-1.5, at 2.4%. On raw speed, MAI-Transcribe-2 processes audio at 410.7 times real-time -- more than double MAI-Transcribe-1.5's 192.6x -- ranking second behind Nova-3's 572.8x. Microsoft's own head-to-head comparisons -- 10x faster than OpenAI's GPT-Transcribe, 7x faster than Scribe v2, 5x faster than Google's Gemini 3.5 Transcribe -- are all plausible against the leaderboard's own numbers for those specific rivals. They just leave out the two models that beat MAI-Transcribe-2 outright, because neither is a company whose name moves a headline.

This isn't a new pattern for the leaderboard Microsoft cites. When Google's Gemini 3.5 Transcribe launched in late August claiming a 2.6% word-error rate across 85+ languages, the same board placed it fifth -- behind Microsoft's own prior model among others -- a rank Google's own launch post didn't mention either. Two vendor launches within two weeks of each other, checked against the same independent board, both overstated where they actually rank once someone opened the leaderboard instead of the press release. MAI-Transcribe-1.5, the model Google's Gemini 3.5 Transcribe was ranked behind in August, is the direct predecessor to the model launching in this story -- which makes Microsoft's own current claim the third data point in the same pattern, not a new one.

The one claim Artificial Analysis's leaderboard can't fully settle is price -- it does not publish a single ranked cost comparison the way it does for accuracy and speed. On the figures this newsroom could confirm, $0.10 per hour genuinely undercuts the list prices for GPT-Transcribe, Gemini 3.5 Transcribe, and Scribe v2, so "cheapest" holds up better than "fastest" or "most accurate" does. (It's also the one number Microsoft flagged as temporary in its own post -- the launch price runs through the end of 2026, with no standard rate named for after.) New capabilities came along with the price cut and the speed gain: speaker diarization, word-level timestamps, keyword biasing for domain-specific terms, automatic language identification, and code-switching support for conversations that mix languages mid-sentence -- Hinglish and Spanglish are the two Microsoft names specifically. None of those are independently benchmarked by Artificial Analysis's leaderboard, which measures word error rate and speed only; they're real product features, just not ones this newsroom can check against a shared, third-party measurement the way the headline claims can be.

Second on an independent leaderboard against every serious rival in the category is still a strong launch.

None of this makes MAI-Transcribe-2 a bad model. The FLEURS result is Microsoft's own honest benchmark, not a fabricated one, and beating three named, real competitors by wide speed margins is a genuine result regardless of what sits above it on a broader board. What it shows is a familiar gap between a company's own superlative and what a shared, continuously updated measurement finds when it's checked the same way for everyone on it -- the exact gap an independent leaderboard exists to close, if a reader bothers to open it. Fun-Realtime-ASR-preview and Nova-3, the two models actually ahead of MAI-Transcribe-2 on accuracy and speed, are not named anywhere in Microsoft's launch materials -- not because Microsoft hid them, but because the standard move in this industry is to pick the comparisons that flatter the launch and let the leaderboard be the fine print, if anyone checks it at all.