Microsoft Unveils Three Proprietary AI Models Amid Copilot Terms Review

Microsoft formally announced the launch of three new, internally developed foundational Artificial Intelligence models on Thursday, April 2, 2026, signaling an aggressive pursuit of "AI self-sufficiency" to directly challenge established leaders like OpenAI and Google in core AI modalities. The new suite—MAI-Transcribe-1 for voice-to-text, MAI-Voice-1 for voice generation, and MAI-Image-2 for image creation—is immediately accessible via the Microsoft Foundry platform and a dedicated MAI Playground.

This strategic deployment represents the first tangible output from Microsoft's Superintelligence Team, which CEO of Microsoft AI, Mustafa Suleyman, established six months prior to focus on creating world-class, cost-efficient model lineages over the next five years. MAI-Transcribe-1 has been positioned as the headline release, setting a new accuracy standard in speech recognition across 25 major languages utilized by Microsoft products. The model achieved an average Word Error Rate (WER) of 3.8% on the FLEURS benchmark, outperforming OpenAI's Whisper-large-v3 across all 25 languages tested and surpassing Google's Gemini 3.1 Flash on 22 of those languages. Furthermore, Microsoft claims this model requires half the GPUs of state-of-the-art competition, with a batch transcription speed 2.5 times faster than the existing Microsoft Azure Fast offering, priced competitively at $0.36 per hour of audio through Foundry.

The generative models in the trio also showcase significant advancements, particularly in speed and cost structure. MAI-Voice-1, the text-to-speech engine, can generate a full minute of natural-sounding audio in just one second and supports custom voice creation from minimal audio samples, priced at $22 per 1 million characters. MAI-Image-2, the second-generation image model, has secured a top-three ranking on the Arena.ai leaderboard and offers at least double the generation speed of its predecessor, with pricing set at $5 per 1 million tokens for text input and $33 per 1 million tokens for image output. This aggressive pricing strategy is designed to undercut competitors like Google, whose Gemini 3 Pro image generation model is priced at $120 per million tokens, signaling Microsoft's intent to compete directly on cost and performance.

Concurrently with this technological unveiling, Microsoft is addressing public controversy surrounding the consumer version of Copilot's Terms of Use. The agreement, updated in October 2025, contains a clause stating that Copilot is “only for entertainment purposes” and explicitly warns users not to rely on it for important advice, instructing them to “Use Copilot at your own risk.” This legal disclaimer creates a contradiction with the company's marketing of Copilot as a crucial business productivity tool integrated across Microsoft 365 applications and Windows 11. A spokesperson for Microsoft acknowledged the conflict, labeling the phrase a “legacy term” that no longer reflects the product's current utility and confirming it is slated for removal in the next update. This specific phrasing reportedly dates back to Copilot's initial positioning as a search companion in Bing, originating from an agreement in February 2023.

This dual narrative—advancing proprietary, high-performance models while managing contradictory consumer liability language—highlights the complex landscape of enterprise AI deployment. Microsoft CEO Satya Nadella has previously emphasized a focus on delivering models that reduce Cost of Goods Sold (COGS) and advance the frontier, a mission Suleyman is now fully dedicated to, having been freed from direct oversight of the day-to-day Copilot experience. While competitors also employ liability disclaimers, none use the explicit 'entertainment purposes only' language, which raises questions about the alignment between Microsoft's legal compliance framework and its commercial strategy for its flagship AI assistant.