Microsoft replaces OpenAI and Anthropic in Office apps

Even Microsoft is starting to look at its AI credit card bill with horror.

For the past few years, Redmond has aggressively pushed Copilot across its empire, all while secretly writing massive checks to OpenAI and Anthropic every time you asked Outlook to summarize an email or Excel to crunch a spreadsheet.

Now, the tech giant is starting to cut those cords.

Ditching the third-party middleman

According to a report from Bloomberg, Microsoft has begun swapping out high-end models from OpenAI and Anthropic inside Office apps like Excel, Outlook, and Word, quietly replacing them with its own homegrown Microsoft AI (MAI) models.

A source familiar with the plan said tens of thousands of prompts every week are already being diverted to MAI models to trim costs. While that is still a tiny slice of Copilot's weekly volume of millions of prompts, the trajectory is clear.

The blunt reason is cash. Microsoft AI chief Mustafa Suleyman told Bloomberg last month that Anthropic is extremely expensive, noting that Microsoft is paying huge sums and is actively working to reduce and ultimately eliminate that bill.

Take a look at the pricing board: Anthropic's restricted Fable 5 runs $10 per million input tokens and $50 per million output tokens. OpenAI charges $5 input and $30 output for GPT-5.5, and even with partner discounts, running those at corporate scale is bleeding cash. Plus, Microsoft's landmark deal with OpenAI expires in 2032.

The "good enough" MAI lineup

To fill the vacuum, Microsoft rolled out seven in-house MAI models last month at Build. The headliner is MAI-Thinking 1, a 35-billion active parameter reasoning model with a 256k context window that reportedly matches Claude Opus 4.6 in coding blind tests at a tiny fraction of the token cost.

Microsoft has admitted that MAI might not beat the bleeding edge of external frontier systems. But to format a table or draft a reply, you do not need an overpriced supercomputer.

This fits right into a broader Silicon Valley budget reality check. Following an early-year token-spending frenzy, giants like Amazon, Meta, Uber, and Accenture have all started slashing their AI tabs. Some US firms are even turning to China's DeepSeek V4-Pro, which sells input tokens for just 43.5 cents per million, despite looming data security questions.

Why it matters

For regular Office workers, this transition will probably happen without you ever noticing. If Microsoft pulls it off, responses might even feel snappier. But we think the real pressure is on output quality: if cheaper internal models start hallucinating inside corporate financial spreadsheets, saving a few cents per query will look like a terrible trade-off.

For the AI ecosystem, this marks the official end of the blank-check era. When your biggest distributor and investor starts ditching your API because you charge too much, you have a structural pricing problem. Companies like Anthropic are now squeezed between lean internal enterprise models on one side and ultra-cheap alternatives on the other.

In our view, Microsoft is making the obvious, inevitable play here. Selling AI features is a money-losing trap if you are just an interface sitting on top of someone else's expensive API. By building its own tiered engine room, Microsoft is preparing for life after 2032, and forcing everyone else to prove why their frontier models are worth the markup.