I watched a product manager mute a Copilot cost alert and keep talking like it didn’t matter. You could see the math in her eyes: every feature, every query, a tiny invoice. I told her we were watching one of the industry’s quiet bank runs.
At a gray conference room table, an Excel sheet was open to a column of invoices.
I want you to imagine what that spreadsheet felt like to the people who pay the bills. Bloomberg reported that Microsoft has quietly started routing tens of thousands of prompts each week in Excel and Outlook to its own MAI models, shifting work away from OpenAI and Anthropic. Microsoft declined to comment on the story, but the signal is clear: costs are steering engineering choices.
AI tokens have become toll booths on a billion-dollar highway. Those tokens — the units that measure compute work — are suddenly a material line item for even the biggest cloud budgets.
That change isn’t sweeping yet. Copilot, Microsoft’s workplace assistant, still consumes massive token volumes. But the company rolled out seven new in-house MAI models in June 2026, including MAI-Thinking-1, a 35 billion active-parameter reasoning model with a 256K context window that Microsoft said matched Anthropic’s Claude Opus 4.6 on coding in blind tests. There are also new image, transcription, voice, and coding models aimed at lower token costs.
Why is Anthropic so expensive?
Anthropic charges steeply for its flagship models: about $10 per million input tokens (€9) and $50 per million output tokens (€46) for its higher-end offerings. That pricing pushed Microsoft AI CEO Mustafa Suleyman to tell Bloomberg, “Anthropic is extremely expensive and I think many people are urgently looking for alternatives.” Internally, that kind of rate turns routine product features into recurring line-item risk.
At a vendor meeting, a finance lead pointed to a line labelled “Anthropic fees.”
Price comparisons are now a daily exercise. China’s DeepSeek made headlines with budget models that charge roughly $0.435 per million input tokens (€0.40) and $0.87 per million output tokens (€0.80) for its V4-Pro. OpenAI’s API for GPT-5.5 is cheaper than Anthropic at about $5 per million input tokens (€5) and $30 per million output tokens (€28), and Microsoft also enjoys a partnership discount on OpenAI use. Still, that discount runs on a clock — the current deal expires in 2032 — so the calculus is time-sensitive.
When every macro feature multiplies into millions of prompts, even a few euro difference per million tokens matters. That’s why an engineering decision — which model answers a user’s question inside Outlook — ripples straight into finance and product roadmaps.
How much does GPT-5.5 cost?
Using OpenAI’s API, GPT-5.5 is about $5 per million input tokens (€5) and $30 per million output tokens (€28). Microsoft’s internal deal lowers the effective rate for some of its uses, but the overall spend is still enormous given global scale and product breadth.
On a demo screen the Copilot usage meter blinked orange.
I’ll tell you plainly: Microsoft can replace some third-party calls with MAI models, but it can’t flip a switch and move everything overnight. Bloomberg’s reporting suggests the MAI substitutions so far represent a small fraction of total token usage. Large-scale features, product integrations, and enterprise SLAs mean migration is gradual and cautious.
Microsoft’s cost strategy is a tightrope across a canyon of vendor fees. The company needs in-house models that are precise enough for product needs and cheap enough to scale without breaking budgets.
Can Microsoft replace OpenAI in its apps?
In certain workloads — short prompts, standard formatting, image or transcription tasks — Microsoft can and is routing traffic to MAI. For high-end reasoning, coding assistance, and other token-hungry uses, Microsoft still leans on partners like OpenAI and Anthropic while it matures its internal stack. Suleyman has said Microsoft pays a lot to Anthropic and wants to cut that cost entirely, which explains the mixed strategy: keep the partnership, build redundancy, shave token bill shock where possible.
If you follow product teams, you’ll see two pressures: feature velocity from the business side and a quiet cost-control campaign from finance. That push-pull will shape which models live in Outlook and Excel, and which remain rented from third parties.
Pay attention: cheaper token economics are no longer a boutique engineering goal. They are a product design constraint, a procurement battle, and a strategic lever. Which vendor wins will determine whose models power the next generation of office features — and who gets paid more for doing less. Who do you think will come out on top?