10 Cents an Hour, and Still Not the Cheapest on the Board

TL;DR

On September 3, Microsoft AI shipped MAI-Transcribe-2 at $0.10 per hour of audio, down from the $0.36 its April predecessor charged. Microsoft claims first place on FLEURS across 60 languages at a 5.2% average word error rate, and second on the Artificial Analysis WER leaderboard. The speed is the part nobody is arguing about: a speed factor of 410.7, which works out to an hour of audio transcribed in under nine seconds. The announcement's title claims three superlatives. The board Microsoft cites backs two of them.

What actually shipped

MAI-Transcribe-2 is a multilingual speech-to-text model, available in preview through Microsoft Foundry, the MAI Playground, and OpenRouter. Microsoft's post lists 60 languages with automatic language identification, code-switching for mixed-language speech (Hinglish and Spanglish get named), speaker diarization, word-level timestamps, keyword biasing for domain jargon, and two output styles: verbatim, which keeps every "um", and clean, which does not.

The competitors Microsoft names directly are GPT-Transcribe, Scribe v2 from ElevenLabs, Gemini 3.5 Transcribe, and Whisper large-v3. Against them it claims 10x, 7x and 5x speed advantages respectively.

Those multiples are self-reported. The speed factor is not.

Why the speed number matters more than the price

A speed factor of 410.7 means the model transcribes 410 seconds of audio per second of wall clock. If you run a podcast pipeline, a meeting-notes product, or anything that ingests long recordings in batches, this is the number that changes your architecture. At 53.8, a two-hour recording takes about two and a half minutes and you build a queue with a status endpoint. At 410.7 it takes 17 seconds and you can just make the user wait. OpenRouter currently reports a P50 latency of 0.18 seconds on the endpoint.

The "cheapest in the world" claim does not survive its own leaderboard

Microsoft titled the announcement "the fastest, most accurate and cheapest speech recognition model in the world." The Artificial Analysis speech-to-text board, which Microsoft cites in the same post for its own ranking, lists MAI-Transcribe-2 at $1.67 per 1,000 minutes. That is exactly the ten cents an hour, and it is not the lowest number in the column. StepFun's Step-Audio line appears as StepAudio 2.5 ASR at $0.37 per 1,000 minutes, roughly 4.5x cheaper.

The defensible version of the claim is "cheapest at this accuracy tier", and that one holds comfortably. Every model on the board with a WER under 2.5% costs at least twice as much per minute. The accuracy superlative is scoped too: Microsoft says first on FLEURS, which is true, while the same post concedes second place on Artificial Analysis, where a preview model called Fun-Realtime-ASR sits ahead at 1.7%. Only the speed claim is unqualified and unbeaten.

Five months, three models, one honest arithmetic trap

The lineage moves fast. MAI-Transcribe-1 landed April 2 with 25 languages at $0.36 per hour. MAI-Transcribe-1.5 followed on June 2 with 43 languages, an hour of audio in under 15 seconds, and keyword biasing that Microsoft said cut WER by up to 30% on FLEURS. MAI-Transcribe-2 is the third release in five months.

Now the trap. The FLEURS average went the wrong way across generations, and it means nothing. VentureBeat reports 3.7% for the 43-language version against 5.2% for this one, which reads like a regression until you notice the denominator changed. Adding 17 harder, lower-resource languages drags the mean up even when every individual language improves. It is the same arithmetic that tanks a school's average score the year it stops testing only the honors class.

The Artificial Analysis composite is the apples-to-apples view, because it holds the test set fixed: 2.6%, then 2.4%, now 2.0%. That board runs a non-streaming blend of AA-AgentTalk (50%), VoxPopuli-Cleaned-AA (25%) and Earnings22-Cleaned-AA (25%), so it is a different mix of audio than FLEURS and worth reading as a second opinion rather than a verdict.

What it costs you, and what expires

VentureBeat runs the obvious enterprise math: 100,000 hours of audio a year drops from $36,000 to $10,000. For a small product, the number that matters more is that a 90-minute podcast episode now costs 15 cents to transcribe with diarization and word timestamps included, which is cheaper than the storage you will put the audio in.

Three caveats before you rip out your pipeline.

• The price is a promotion. Microsoft's wording is exact: $0.10 per hour "as a limited-time offer until the end of the year." No successor rate is published. This is a pricing strategy with a countdown timer attached, and January 1 is the reveal.
• It is a preview. Unite.AI notes the listing ships as a public preview without a service-level agreement and is not recommended for production workloads. Microsoft's own post frames Foundry, the MAI Playground and OpenRouter as places to try it.
• The comparison multiples are Microsoft's. The 10x/7x/5x figures come from the vendor. The 410.7 speed factor and 2.0% WER come from a third party running its own harness, which is why those are the numbers worth quoting.

Where this leaves Whisper

Self-hosted Whisper large-v3 still wins on the two things a price cut cannot buy: your audio never leaves your machine, and nobody can reprice it in January. But the gap in raw economics is now wide enough that "we run our own ASR to save money" has stopped being true for most workloads. If you are still paying for a GPU to sit around transcribing, the reason had better be privacy, not the bill.

Key Takeaways

• MAI-Transcribe-2 launched September 3 at $0.10 per hour of audio, a 72% cut from the $0.36 Microsoft charged for MAI-Transcribe-1 in April.
• Artificial Analysis measures a 410.7 speed factor, roughly an hour of audio in under nine seconds, and a 2.0% word error rate for second place behind Fun-Realtime-ASR at 1.7%.
• The "cheapest in the world" claim fails on Microsoft's own cited board: StepAudio 2.5 ASR lists at $0.37 per 1,000 minutes versus $1.67, at 4.7% WER. Cheapest under 2.5% WER is the accurate claim.
• The 5.2% FLEURS average is higher than the previous generation's because the language set grew from 43 to 60, not because accuracy fell. Compare on a fixed test set or do not compare.
• The price expires December 31 with no published successor rate, and the model ships as a preview with no SLA. Price your pipeline for the unknown January number.
• Diarization, word-level timestamps, keyword biasing and verbatim/clean styles are included rather than metered as add-ons, which is a real difference from providers that charge per feature.

Sources: Microsoft AI: MAI-Transcribe-2, Artificial Analysis Speech to Text leaderboard, VentureBeat, Unite.AI, Microsoft AI: MAI-Transcribe-1.5, Microsoft AI: MAI-Transcribe-1, OpenRouter model page