Microsoft replacing third-party models with its own internally developed models

In an aggressive pivot toward tighter cost control, Microsoft has quietly begun replacing third-party models from OpenAI and Anthropic with its own internally developed “MAI” (Microsoft AI) models across some of its highest-volume software products.

According to reports from Bloomberg, tens of thousands of customer prompts per week inside Excel, Word, and Outlook are now being routed directly to Microsoft’s in-house infrastructure. This shift transitions Microsoft’s proprietary AI out of the research phase and straight into production as a viable alternative to the models it has previously rented.

1. The Strategy: “Token Minimizing” and Tiered Routing

The transition highlights a major industry shift where tech giants are moving away from “tokenmaxxing” (using premium frontier models for every single query) toward high-efficiency cost reductions.

Microsoft is executing a disciplined, tiered routing strategy to protect its profit margins:

• Routine Tasks Go In-House: Standard, everyday user prompts—such as summarizing a routine email thread in Outlook or formatting basic formulas in Excel—are increasingly being routed to Microsoft’s lighter, hyper-optimized first-party models.
• Frontier Models for Complex Edge Cases: High-end reasoning models like OpenAI’s $o$-series or Anthropic’s Claude are being reserved for genuinely difficult, multi-step operations where the extra depth justifies the steep price premium.
• The Exclusivity Off-Ramp: While Microsoft’s multi-billion dollar partnership with OpenAI remains intact, a recent contract renegotiation ended Microsoft’s strict exclusivity. This legally freed CEO Satya Nadella and his team to deploy competing tech to prevent the company from becoming over-dependent on a single external partner.

2. Meet the First-Party “MAI” Engine Lineup

At its Build developer conference, Microsoft unveiled seven first-party MAI models, explicitly designed by Microsoft AI CEO Mustafa Suleyman’s team to undercut the pricing of top-tier AI labs.

The centerpiece of this cost-slashing initiative is MAI-Thinking-1, the company’s first proprietary reasoning model:

Model	Active Architecture	Stated Purpose & Pricing Impact
MAI-Thinking-1	35B parameter Mixture of Experts (MoE)	Built entirely on commercially licensed data. Positioned to run on Azure at roughly 40% below comparable OpenAI reasoning pricing.
MAI-Code-1-Flash	5B parameter agile coding engine	Directly integrated into GitHub Copilot and VS Code to handle rapid autocompletions for just $0.75 per million input tokens.
MAI-Transcribe-1.5 & Voice-2	Specialized audio-to-text matrices	Slated to replace third-party systems for real-time video captions and transcription inside Microsoft Teams over the coming months.

3. Financial and Industry Repercussions

The market responded positively to the cost-containment narrative, with Microsoft’s stock ticking up 1.75% following the initial leak—proving that investors currently value margin stability over raw model hype.

However, the move presents a distinct threat to pure-play AI labs:

The Anthropic Target: Mustafa Suleyman did not mince words regarding the financial motivation behind the shift, stating at the Build conference: “We pay a lot of money to Anthropic—so our goal is to reduce and ultimately eliminate that cost.”

By building out its own code-generation and processing capabilities that match older flagship models (like Claude Opus) at a fraction of the price, Microsoft is systematically pulling hundreds of millions of dollars out of the external API ecosystem, keeping both the traffic and the enterprise profits locked safely inside its own cloud borders.