Microsoft Starts Eating Its Own Cooking: MAI Models Quietly Replace OpenAI and Anthropic Inside Excel and Outlook

What happened

On July 7, Bloomberg reported that Microsoft has begun replacing OpenAI and Anthropic models with its own in-house MAI models inside Excel and Outlook. Per that reporting, "tens of thousands of AI prompts in the widely used spreadsheet and email applications are now being completed each week" by MAI systems that Microsoft built itself.

There was no launch event. No blog post. No changelog entry naming the affected features. Microsoft declined to comment, and neither OpenAI nor Anthropic issued a statement about the transition. The story exists because a reporter found it, not because a company announced it.

That is precisely what makes it worth reading closely. For three years the defining fact of Microsoft's AI strategy was that it rented its intelligence — first from OpenAI under a partnership so tight the two firms were often discussed as one entity, later from Anthropic as well. The Bloomberg report is the first concrete evidence that the rental period has an end date, and that Microsoft has already started moving out of some rooms.

The quiet part: nothing changed on screen

Office Watch, reviewing the change on July 8, put the user-facing impact bluntly: there essentially isn't one. "The Copilot button looks identical and sits in the same place, but the machine behind it is increasingly one Microsoft built to cut costs." The publication noted that a model picker in the interface still surfaces choices like GPT-5.5, Claude Opus 4.7, and Claude Opus 4.8 alongside an "Auto" default — and it is the behavior of that "Auto" default, not the named options, where the substitution appears to be happening.

This is the substitution that matters and the substitution nobody consents to. A user who explicitly selects Claude Opus 4.8 gets Claude Opus 4.8. A user who leaves the setting on "Auto" — which is to say, nearly every user — gets whatever Microsoft's routing layer decides is cheapest for the task. The model picker's presence provides a kind of alibi: the frontier models are available, so nothing was taken away, even as the default silently migrates.

Office Watch was careful to flag the epistemic limits here, and so should we. The reporting traces to a single Bloomberg source. Microsoft has not documented which features moved, when they moved, or what fraction of traffic they represent. "Tens of thousands of prompts per week" across Excel and Outlook — two applications with hundreds of millions of users — is a rounding error. This is a direction of travel, not a completed migration.

Why Microsoft would do this

The motive is not subtle, because Microsoft's AI chief said it out loud a month ago. Mustafa Suleyman, speaking in June: "We pay a lot of money to Anthropic — so our goal is to reduce and ultimately eliminate that cost."

Read that sentence again, because it is remarkable. The head of AI at a company that is simultaneously one of Anthropic's largest customers and one of OpenAI's largest investors stated, on the record, that eliminating payments to a frontier lab is a corporate objective. Not diversifying. Not negotiating. Eliminating.

The scope is expanding accordingly. Per Bloomberg's account, MAI models are already available through GitHub Copilot, and Microsoft plans to deploy an internally built transcription model in Teams and other products in the coming months. At its June Build conference, the company unveiled seven new AI models — including one specifically designed to match the coding performance of Anthropic's Opus 4.6 at lower cost.

The logic is straightforward once you separate AI work into tiers. Drafting an email reply, summarizing a thread, generating a SUM formula — these are high-volume, low-complexity tasks where the marginal quality difference between a frontier model and a competent cheap one is invisible to the user. Routing that traffic to a model you own outright converts a variable third-party cost into a fixed internal one. The hard reasoning tasks can still go to OpenAI or Anthropic. Microsoft is not trying to win the frontier. It is trying to stop paying frontier prices for commodity work.

The other half of the story: the floor fell out of token prices

Microsoft's move looks less like a bold bet and more like an inevitability when you set it against a second piece of reporting from the same week.

Also on July 7, CNBC published an investigation — summarized by CIO and Forkast — finding that Chinese-origin AI models now account for more than 30% of the AI tokens routed weekly by US companies through OpenRouter, a platform that brokers requests across many providers. That share has held above 30% every week since February 8, 2026, peaking around 46%. The prior twelve-month average was 11%. In the first half of 2025 it was 4.5%.

The driver is price. Justin Summerville of OpenRouter told CNBC that "Chinese models are consistently 60% to 90% cheaper than the leading offerings from Anthropic and OpenAI." The concrete comparison in the reporting: DeepSeek V4 Flash at $0.14 per million input tokens against OpenAI's GPT-5.5 at $5.00 — a gap the report characterizes as ranging from 4x to 100x depending on the pairing.

By Forkast's account of the CNBC data, DeepSeek leads with 17.6% of routed tokens and Alibaba's Qwen follows at 13.9%, while Anthropic — the largest US provider on the platform — captures 14.8%.

Hype vs. real

Three cautions, in ascending order of importance.

First, the numbers deserve scrutiny. Forkast reports DeepSeek at 17.6% of routed tokens and cites 5.13 trillion tokens weekly, alongside Qwen at 13.9% and 2.77 trillion weekly. Those pairs do not reconcile against each other, nor cleanly against the platform total the same piece gives (OpenRouter growing from roughly 5 trillion tokens weekly in April 2025 to over 20 trillion by April 2026). Something has been garbled in the retelling. The share figures are the load-bearing claim and they appear consistently across outlets; treat the absolute token counts as unverified.

Second, OpenRouter is not the enterprise. It is a routing platform whose users skew toward cost-sensitive developers and hobbyists, exactly the population most likely to chase a 90% discount. Extrapolating "30-46% of US enterprise AI" from its traffic mix is a leap the underlying data does not support. Large regulated enterprises with procurement departments and data-residency requirements are not the marginal buyer here.

Third, and most substantively: MAI may not be good enough yet. The Decoder's assessment of Microsoft's MAI-Thinking 1 reasoning model is that it trails competitors meaningfully — performing "roughly on par with Deepseek V3.2" rather than with the current frontier, notwithstanding Microsoft's claims of comparable coding performance. That is a third-party evaluation, not a benchmark Microsoft published, and it should be weighted as such. But if it is even roughly right, the honest framing of the Excel and Outlook swap is not "Microsoft matched the frontier internally." It is "Microsoft decided a chunk of Copilot never needed the frontier, and is finding out whether users notice."

What to watch

The Decoder raises the question that will determine whether this becomes a story about margins or a story about trust: users may face reduced capability at unchanged prices. The publication notes that CEO Satya Nadella has hinted at a shift toward usage-based billing, which would make cheap MAI models the default while premium third-party models become a paid add-on.

That is the real fork in the road. Copilot's per-seat pricing was set in an era when every seat implied frontier-model inference. If the inference underneath quietly becomes cheaper while the price does not, the margin improvement accrues entirely to Microsoft. If Microsoft instead moves to metered billing and charges for Opus access on top, it has effectively repriced the product upward while telling a cost-savings story. Neither outcome is fraud. Both are worth naming in advance.

Watch also for the second-order effect on Anthropic and OpenAI. If the largest enterprise software distributor on earth systematically strips commodity inference out of its third-party spend, and simultaneously a cohort of price-sensitive developers migrates to DeepSeek and Qwen, the frontier labs are left selling into a narrowing band: the genuinely hard tasks, where nothing cheaper suffices. That band is real, defensible, and lucrative. It is also much smaller than the total-addressable-market slides implied.

The takeaway

The frontier and the floor are separating, and this week gave us a data point from each side.

Microsoft's Excel and Outlook swap is small in volume and large in signal: the company that did the most to establish frontier models as infrastructure has begun treating them as a line item to be optimized away. Suleyman said the goal is to eliminate the Anthropic cost, and the first tens of thousands of prompts have moved. Meanwhile, on the open platforms, Chinese open-weight models have gone from 4.5% of routed US traffic to a sustained 30%-plus in roughly eighteen months, on the strength of nothing more exotic than being 60% to 90% cheaper.

What connects them is a realization that took the industry three years to reach: most AI work is not hard. Summarizing an email thread does not require a model that can prove theorems. The 2023-2025 era priced every token as though it might. That era is ending, and it is ending from the bottom up — not because the frontier stopped mattering, but because everything below the frontier turned out to be a commodity, and commodities find their price.

The frontier labs' response will define the next year. Nobody has announced one yet.