Microsoft's MAI-Transcribe-2 undercuts OpenAI and Google on speed and price

Microsoft AI launched MAI-Transcribe-2 on September 3, claiming the model is the fastest, most accurate, and cheapest speech recognition service on the market. At $0.10 per audio hour through the end of 2026 — a 72% price cut from its predecessor MAI-Transcribe-1.5 — it aggressively undercuts OpenAI's GPT-Transcribe, Google's Gemini 3.5 Transcribe, and ElevenLabs' Scribe v2.

The model ranks first on the FLEURS benchmark across 60 languages with a 5.2% average word error rate, and second on the independent Artificial Analysis leaderboard. In head-to-head speed comparisons, MAI-Transcribe-2 processes audio 10× faster than OpenAI's GPT-Transcribe, 7× faster than ElevenLabs' Scribe v2, and 5× faster than Google's Gemini 3.5 Transcribe.

Beyond raw performance, the model adds features enterprise teams have been waiting for: speaker diarization to attribute speech to individual participants, word-level timestamps, automatic language identification, and code-switching support for blended-language conversations like Hinglish or Spanglish. Keyword biasing lets developers weight domain-specific terminology so it is transcribed correctly even in noisy conditions. Output modes toggle between verbatim and clean transcripts depending on the use case.

The expansion to 60 languages — up from 43 in the June release of MAI-Transcribe-1.5 — is significant for global enterprise deployments. The model is available immediately on Microsoft Foundry, the MAI Playground, and Open Router.

The launch is a clear escalation in Microsoft's strategy to own the transcription infrastructure layer rather than rely on third-party providers. Five months ago, its first model in this line cost $0.36 per hour. The new $0.10 price point makes it the cheapest option in a competitive market where latency and accuracy are increasingly table stakes, as first reported by VentureBeat.