More than 90% of Claude Cowork usage is unrelated to software development. That figure looks like a product statistic. It is also a clue to why Microsoft reportedly wants its own models inside Excel and Outlook.

The unit of competition changed with the unit of use

A model used for an occasional experiment can be judged mainly on capability. A model invoked throughout Excel or Outlook has a different economic shape. Usage repeats, the cost accumulates, billing terms determine how predictable that cost is, and the surrounding workflow makes substitution less trivial than changing a dropdown.

For the model provider, that embedded usage is recurring enterprise revenue. For the software owner, it is recurring cost and supplier exposure. The same integration appears as growth on one side of the contract and procurement risk on the other.

This is why Microsoft’s reported move to replace some OpenAI and Anthropic models with its own MAI models in Excel and Outlook matters. The stated motive is not a new benchmark lead. It is reducing AI costs. The model is being evaluated as an input to a high-frequency production system, not as a demonstration of technical prestige.

Microsoft’s swap began as an architecture decision

The builder-layer signal appeared before the cost rationale reached specific Office products. In March 2025, sources reported that Microsoft had completed training a family of models codenamed MAI. They also said the company was experimenting with replacing OpenAI models in Copilot with MAI.

The important continuity is substitution. Testing whether an application can move between model suppliers creates leverage before any replacement is finalized. It turns the model layer from a fixed dependency into a procurement surface.

Once model calls become a material recurring cost, an application owner with an internal alternative has a reason to test it, route work toward it, and use that option in supplier negotiations. The relevant capability is no longer just generating an answer. It is being substitutable enough for the task.

Credits and billing terms are becoming product features

The suppliers are responding to the same structure from the opposite direction. OpenAI, Anthropic, and other leading labs are seeking durable streams of enterprise revenue while using credits, promotions, and bonuses to secure recurring use. Those offers are not peripheral marketing. They alter the adoption economics of embedding a model into work.

A credit lowers the cost of establishing a workflow. A promotion gives repeated use time to become habit. A bonus rewards volume. Once employees rely on a model for a task, replacing it means changing behavior as well as software. The subsidy is temporary; the workflow friction it purchases can persist.

Anthropic’s recent product and billing moves make the mechanism unusually visible. Access to Claude Fable 5 was scheduled to shift to token-based billing on July 7, while Anthropic extended access for all paid plans through July 12. Claude Cowork also expanded from desktop to web and mobile, with beta access for Max subscribers.

of Claude Cowork usage is unrelated to software development, according to Anthropic

That figure matters because it places the revenue contest beyond coding. As model use spreads into general work, access terms and billing mechanics touch a larger set of recurring tasks. OpenAI is moving along the same work-software boundary: a ChatGPT update added easier email editing alongside controls for tone, formatting, headers, lists, and emojis. The labs are not merely improving answers. They are fitting themselves into repeatable work.

This is prediction-input decoupling in commercial form. Once useful model output becomes available across multiple providers, an architecture built around one scarce source of intelligence becomes harder to defend. Value shifts toward the system that can deploy that intelligence repeatedly, economically, and with enough integration friction to keep the usage attached.

Cost discipline does not mean capability stopped mattering

The Microsoft replacement remains sourced reporting and is explicitly unconfirmed. The account describes a move beginning inside selected products, not a completed company-wide migration. A report is not a rollout, and an experiment is not proof that the alternatives perform equally well.

Anthropic is also expanding rather than retrenching. It plans to double its New York City workforce to 1,000 employees this year and lease a 16-story building in Hudson Square. The capability-and-scale race remains active alongside billing discipline.

That is counter-evidence only if procurement is mistaken for commoditization. It should not be. A cheaper model that cannot clear the task threshold is not a substitute. But once several models clear that threshold, price, billing predictability, integration cost, and control over the stack become decision variables. Capability remains the admission ticket; deployment economics decide what happens after admission.

The workflow owns the margin

Excel and Outlook expose the structural stakes because they already contain the work. Microsoft does not need to persuade an enterprise to adopt a new destination before it can distribute a model there. OpenAI and Anthropic, meanwhile, have an incentive to use commercial terms to make their models the recurring choice inside those destinations.

That produces convergence: Microsoft develops and tests internal substitutes; labs subsidize adoption and adjust billing; assistants expand from coding into email and general work; recurring enterprise revenue becomes the prize. The convergence does not require coordination. The unit economics are enough.

The 90% outside coding is not merely evidence that AI has entered general work. It marks the point at which model use becomes a recurring line item—and the software already open can decide whose model gets paid.