Microsoft's 10-cent transcription hour undercuts its own prior pricing, requiring rivals at the old rate to find 3.6 times the volume

An early-bird rate is a term rather than a price, and the whole question is which of the two 10 cents an hour turns out to be ^[1][2]. The cost side of the answer sits in throughput: Microsoft says the model clears audio 10 times faster than OpenAI's GPT-Transcribe, seven times faster than ElevenLabs' Scribe v2 and five times faster than Google's Gemini 3.5 Transcribe ^[14], and in batch work that ratio is the unit cost, because a model running at 300 times real time burns a fraction of the GPU-hours of one running at 30 ^[15]. Mustafa Suleyman told The Verge in April that the first model in the line came out of a 10-person team and ran at half the GPU cost of other state-of-the-art models ^[18]. If those multiples hold, a dime an hour can sit above marginal cost and stay there. If they do not, this is acquisition spend with a renewal date attached.

For anyone selling transcription by the hour, the number that matters is 36 divided by 10: a vendor that matched Microsoft's April rate now needs 3.6 times the volume to hold revenue flat ^[1]. Scale that down and it gets starker, since at 10 cents an hour a full year of continuously recorded audio, all 8,760 hours of it, bills at $876 ^[3], and a $1m revenue line requires 10 million hours of customer audio ^[4]. That is the arithmetic that makes standalone speech-to-text hard to fund, or rather, the more useful version is that it makes the features the funding, and Microsoft has moved speaker diarization, word-level timestamps, keyword biasing for domain jargon and automatic language identification into the base rate ^[6], which is the same menu the specialists have historically billed as extras.

The buyer, meanwhile, barely notices. On the 100,000 hours of call-center audio Microsoft uses as its own illustration, the bill goes from $36,000 to $10,000 ^[3], a saving of $26,000 a year ^[2], which is less than a single hire and nowhere near the cost of revalidating a compliance workflow. The volume moves on defaults, not on price, and Microsoft owns Teams and its meeting audio, Nuance and its clinical documentation, and the Azure speech services enterprises already call ^[17]. Bloomberg reported in July that MAI models had begun answering a portion of prompts in Word and Excel, products previously advertised as running on OpenAI and Anthropic ^[16], which is what substitution looks like when nobody signs anything.

What the evidence does not yet establish is whether the transcript is good enough for the specific buyer. The 5.2% average word error rate that Microsoft ranks first with comes from FLEURS ^[8], a benchmark built from native speakers reading about 2,000 sentences per language ^[9], and it is up from the 3.7% reported in June, roughly 41% higher in relative terms ^[10][5], with no per-language breakdown in the release ^[11]. Microsoft's other yardstick is Artificial Analysis, where it says the model now ranks second ^[12] on a leaderboard that had the previous version third at 2.4% ^[13]. A buyer whose traffic is Hinglish or Spanglish, the two cases Microsoft names for its code-switching mode ^[7], is buying an average.

The release names four rivals and omits Deepgram, AssemblyAI, Speechmatics and Rev ^[19], which reads to me as a judgment that those firms are not competing on API list price. The thesis fails if 10 cents expands the pool rather than dividing it: at $876 for a year of continuous audio, categories nobody bothered transcribing become worth transcribing, and 3.6 times the volume is a target the specialists can actually hit.