Microsoft AI released a speech-recognition model this week that it says beats every major rival on speed, accuracy and price simultaneously — an unusually aggressive claim in a field where those three qualities are normally traded against each other.
MAI-Transcribe-2 ranks first on the FLEURS benchmark across 60 languages, with an average word-error rate of 5.2%, and second on the independent Artificial Analysis leaderboard, where Microsoft says it defines the frontier for balancing accuracy against latency. The company's own comparison puts it 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.
Ten cents an hour, for now
The pricing is the sharper edge. Microsoft is charging $0.10 per hour of transcribed audio through the end of the year — a 72% cut from the $0.36 rate its predecessor model charged five months ago, according to VentureBeat's reporting. The model bundles features that competitors often sell as add-ons, including speaker diarization, word-level timestamps, keyword biasing and automatic language identification, and it is available immediately through Microsoft Foundry, the MAI Playground and OpenRouter.
Notably, Microsoft did not claim to beat every competitor on raw accuracy — Alibaba's transcription models are absent from its superiority claims — but leaned instead on the combination of speed and cost as the differentiator for enterprise customers processing large volumes of call-center, meeting or media audio.
A pattern of narrow, cheap models
The release fits what analysts describe as Microsoft AI's emerging playbook under Mustafa Suleyman: target a specific modality, optimize aggressively for inference cost, undercut incumbents on price, and fold the result into Microsoft's own products rather than routing that traffic through OpenAI. Suleyman has attributed the unit's speed to small teams working outside normal corporate process — describing one internal effort to The Verge as coming from "a small, focused 10-person team" that had been "liberated from any of the bureaucracy."
That strategic drift has not gone unnoticed by Microsoft's partners and rivals. "Microsoft is building their own AI and I don't think Microsoft will use OpenAI in the future," Salesforce chief executive Marc Benioff said recently, describing Microsoft's trajectory toward its own frontier models. MAI-Transcribe-2 is a narrow product in a crowded field, but it is also a data point in a much bigger question hanging over the OpenAI-Microsoft relationship: how much of the underlying AI stack Microsoft ultimately intends to own itself.
