Microsoft has cut the price of AI-powered speech transcription by 72 percent with the launch of MAI-Transcribe-2, a new model from its in-house Microsoft AI unit that the company says beats OpenAI, Google, and ElevenLabs on both accuracy and speed. The model, launched on September 3, 2026, is priced at just $0.10 per audio hour through December 31, 2026 — a steep drop from the $0.36 per hour Microsoft charged for its previous MAI-Transcribe-1.5 model in June.
The move is the clearest sign yet that Mustafa Suleyman's Microsoft AI division, built around the team Microsoft hired from Inflection AI in March 2024, intends to compete directly with the partners and rivals it once relied on. For developers tracking the fast-shifting economics of AI APIs, AI Buzz Wire covers every major AI tools launch as it happens.
The Numbers Behind the Price Cut
According to Microsoft's launch post and benchmark coverage from VentureBeat, MAI-Transcribe-2 ranks first on the FLEURS multilingual benchmark across 60 languages, with an average word error rate of 5.2 percent. On Artificial Analysis's independent non-streaming leaderboard, the model posts a 2.0 percent word error rate and ranks second overall for accuracy, behind only Alibaba's Fun-Realtime-ASR-preview.
Speed is where Microsoft claims the gap truly opens. Based on Artificial Analysis evaluations, the company says MAI-Transcribe-2 runs 10 times faster than OpenAI's GPT-Transcribe, 7 times faster than ElevenLabs' Scribe v2, and 5 times faster than Google's Gemini 3.5 Transcribe. Artificial Analysis currently lists the model with a median speed factor of 410.7, meaning an hour of audio can come back transcribed in roughly nine to ten seconds.
What the Model Can Do
Beyond raw speed and price, the feature list is aimed at production workloads rather than demos. Available through Microsoft Foundry, the MAI Playground, and OpenRouter, MAI-Transcribe-2 supports automatic language identification, code-switching between languages mid-sentence, speaker diarization, word-level timestamps, keyword biasing, and configurable transcript styles.
Those features decide whether a transcript is actually usable in the settings where transcription earns money: call centers routing customer complaints, clinics documenting visits, newsrooms processing interviews, and meeting rooms where multiple people talk over each other. Keyword biasing in particular lets developers teach the model domain-specific vocabulary — drug names, product SKUs, legal terms — without retraining anything.
A Deliberate Strategy From Suleyman's Shop
MAI-Transcribe-2 is Microsoft AI's second transcription model in three months. MAI-Transcribe-1 launched at $0.36 per hour, and MAI-Transcribe-1.5 followed in June with support for 43 languages and a third-place ranking on Artificial Analysis. The new model extends coverage to 60 languages while dropping the price by more than two-thirds.
The unit has previously shipped MAI-1 for text and MAI-Voice-1 for speech generation, along with recent cybersecurity and image models. What is conspicuously absent from the launch is any mention of Copilot or OpenAI — notable given Microsoft's multibillion-dollar investment in OpenAI and its history of routing AI features through that partnership. Instead, Microsoft is selling a Microsoft-branded model that competes head-to-head with OpenAI's own transcription product.
A Price War in Speech AI
The transcription market has quietly become one of the most competitive corners of applied AI, and the pricing pressure is real: Microsoft notes the $0.10 promotional rate expires on December 31, 2026, which functions as both a customer incentive and a warning to competitors about where prices are headed.
For startups in the space, the math is brutal. If a hyperscaler can deliver top-tier accuracy at a tenth of a dollar per audio hour while turning around an hour of audio in seconds, venture-backed transcription vendors must find value in vertical specialization, workflow integration, or privacy guarantees — or get compressed out of the market entirely.
For buyers, meanwhile, the message is simpler than it has been in years: state-of-the-art speech recognition now costs less than a dime an hour, and it is only getting cheaper.
Stay Ahead of AI
New AI tools launch every week, and the winners and losers shift fast. Follow AI Buzz Wire for independent coverage of the AI tools market.
Read the latest AI news →
