Microsoft Starts Replacing Costly OpenAI And Anthropic AI With Cheaper In-House MAI Models

Microsoft Is Quietly Swapping Out AI Models To Cut Costs

Microsoft has started shifting a portion of its everyday AI workload away from OpenAI and Anthropic, turning instead to its own in-house models to handle tasks inside Excel and Outlook.

According to Bloomberg, which spoke to a person familiar with the matter, tens of thousands of prompts a week in these two widely used apps are now being processed by Microsoft's own MAI models instead of being routed through third-party systems as before.

The person was not named, as they were not authorised to speak publicly about internal strategy.

Why Is Microsoft Moving Away From Anthropic And OpenAI

The driving force behind this shift is cost.

Running top-tier AI models at the scale Microsoft operates at is expensive, and the company is looking to trim that bill wherever it can.

Mustafa Suleyman, who leads Microsoft AI, was blunt about the situation when he spoke to Bloomberg last month.

"Anthropic is extremely expensive and I think many people are urgently looking for alternatives. We pay a lot of money to Anthropic, so our goal is to reduce and ultimately eliminate that cost."

He also pointed out just how much this is costing the wider business, noting that "many, many people in our organisation are spending millions of dollars" on AI tokens, the units used to measure how much computing power a model consumes.

How Big Is This Shift Really

Despite the headlines, this is still a small piece of Microsoft's total AI usage.

Copilot, the company's flagship workplace assistant, handles many millions of prompts every week, dwarfing the volume now being redirected to MAI models.

Even so, this marks the first time the scale of MAI adoption inside Excel and Outlook has been made public, and it signals that Microsoft's own models are maturing enough to take on real, everyday workloads rather than just existing as experimental side projects.

What Are The New MAI Models Capable Of

The move follows Microsoft's Build conference last month, where the company unveiled seven new MAI models covering reasoning, image generation, transcription, voice recognition and coding.

The standout was MAI-Thinking 1, Microsoft's first reasoning model, described as a mid-sized system with 35 billion active parameters and a 256,000-token context window.

Microsoft said it was built for strong performance at a low token cost, and in blind testing, it reportedly matched the coding ability of Anthropic's Claude Opus 4.6, a model that remains popular despite being from a previous generation.

Microsoft's MAI models are also being made available through GitHub Copilot, and Suleyman has said a Microsoft-built transcription model will soon be rolled out to Teams and other products.

How Does The Pricing Actually Compare

The numbers help explain why Microsoft is so focused on cutting costs.

Anthropic charges 10 dollars per million input tokens and 50 dollars per million output tokens for its most advanced Fable 5 model, which comes with heavy usage restrictions.

OpenAI is cheaper by comparison, charging 5 dollars per million input tokens and 30 dollars per million output tokens for GPT-5.5.

Microsoft does get a discount on OpenAI's pricing thanks to their long-standing partnership, but that arrangement will not last forever, as the deal is set to expire in 2032.

Chinese AI labs have pushed prices down even further.

DeepSeek's V4-Pro model, which made headlines earlier this year for its affordability, costs just 43.5 cents per million input tokens and 87 cents per million output tokens.

Some American firms have already begun switching to Chinese models to take advantage of the lower prices, even while acknowledging concerns about the security risks involved.

Is This Part Of A Bigger Industry Trend

Microsoft is far from alone in rethinking its AI spending.

Earlier this year, the industry went through a phase of what's been dubbed tokenmaxxing, where companies pushed AI usage to the limit with little regard for cost.

That enthusiasm has since cooled, and firms including Amazon, Accenture, Meta and Uber have all reportedly taken steps in recent weeks to bring their own AI bills under control.

Microsoft's pivot towards MAI models fits squarely into that broader pattern, one where efficiency and cost control are starting to matter just as much as raw model performance.