Microsoft Copilot’s Multi-Model Router: OpenAI, Claude, Usage Billing and FinOps Controls

According to an October 2026 inside report by Aaron Holmes in The Information, later analyzed by tech analyst Michael Parekh in his AI: Reset to Zero newsletter, Microsoft executive vice president Charles Lamanna has spent the past year steering Microsoft Copilot away from what he calls "model monogamy". The new vision? Turning Microsoft into "the Costco of AI"—a high-volume, bundled platform that dynamically routes everyday office tasks to whichever underlying AI model can get the job done with the best balance of quality and cost.

For Windows and Microsoft 365 administrators, this shift is far more than boardroom theater. Alongside Microsoft’s late-September 2026 Copilot platform overhaul, it fundamentally alters how Copilot works under the hood, how enterprise tasks are routed across competing frontier models, and how IT departments will have to budget for a hybrid world of flat-rate licenses and metered AI agents.

Breaking Up with "Model Monogamy"​

For the first few years of the generative AI wave, Microsoft’s strategy looked straightforward: pour billions into OpenAI, enjoy exclusive access to its frontier weights at no extra licensing cost under their partnership agreement, and embed those models across Windows 11, GitHub, and Microsoft 365.

By the summer of 2025, however, that single-vendor dependency was showing cracks. As reported by The Information, enterprise customers were complaining that Copilot was buggy and struggled to reliably automate complex work inside Excel spreadsheets and PowerPoint presentations.

Lamanna, then a 38-year-old Microsoft president who had risen through Scott Guthrie’s cloud division to lead Dynamics 365 and Power Apps, brought Nadella a proposal that he himself described as "heresy" at the time: pay rival lab Anthropic to use its Claude models inside Copilot. Because Microsoft could tap OpenAI’s models at no additional licensing cost while Anthropic required hard cash, Microsoft commercial chief Judson Althoff reportedly pushed back on the expense. Lamanna won Nadella over with live demonstrations showing that Anthropic’s models were sharper at Excel math and generated cleaner, better-formatted PowerPoint slides.

In Lamanna’s view, relying exclusively on a single lab carried structural risk. He warned that building Copilot around "model monogamy" would bake in too much dependence on a single model's quirks and behaviors. By September 2025, Anthropic models were officially integrated into Office workflows, followed later by specialized bundles such as a $99 Copilot Cowork tier incorporating Claude Cowork and a "Critique" feature in Copilot Researcher that used Claude to double-check OpenAI’s output.

Section Summary:

• The Problem: Customer pushback in mid-2025 highlighted weaknesses in Excel math and PowerPoint automation when relying solely on OpenAI models.
• The Pivot: Charles Lamanna convinced CEO Satya Nadella to bring Anthropic’s Claude models into Copilot despite internal pushback over paying an outside vendor.
• The Philosophy: Ending "model monogamy" prevented Microsoft’s application layer from becoming overly dependent on a single AI lab's training behaviors.

Inside the "Costco of AI" and the Model Router​

Bringing in a second frontier model lab solved a quality problem, but it created a massive margin problem. According to The Information, Microsoft is on track to pay Anthropic more than $2 billion a year to help power Copilot—a figure that underscores why every prompt routed to a third-party model hits Microsoft's bottom line differently than an OpenAI call.

To manage those economics without degrading the user experience, Lamanna’s engineering organization built an intelligent model router powered by a homegrown large language model. When an enterprise worker types a prompt or triggers an automation, the router evaluates the request and selects the least expensive model capable of handling the task without sacrificing quality.

Here is how that routing dynamic plays out in practice today:

1. Default "Auto" Routing: Most Office users leave Copilot in its default "Auto" mode. As Microsoft’s official September 25, 2026 announcement by Chief Marketing Officer for AI at Work Jared Spataro confirmed, the Auto setting weighs accuracy, speed, and cost on each request to pick the right model for the job.
2. Agent Mode Adjustments: Copilot’s "agent mode," which directs models from both OpenAI and Anthropic to spend more time reasoning through multi-step Office tasks, is now the default experience. As OpenAI’s models caught back up to Anthropic’s performance in recent months, Lamanna’s team adjusted the router to lean more heavily on OpenAI for agent mode—taking advantage of Microsoft’s ability to run those models without paying per-token licensing fees to an outside rival.
3. The Wholesale Pitch: Lamanna tells his staff that Microsoft needs to be "the Costco of AI," arguing that the company should not act like a high-priced boutique reserved only for elite coders and R&D scientists, but should instead drive prices down so mainstream workers can use it at scale.

What Changed in the New Copilot App: Home, Code, and Autopilot​

The "Costco" analogy is about bundling and volume, and Microsoft’s September 25, 2026 product overhaul shows exactly what is going onto the warehouse shelves. Microsoft has consolidated its experience into a unified Copilot application built around three core pillars—Home, Code, and Autopilot—grounded in an enterprise intelligence layer called Microsoft IQ.

For IT pros and everyday Windows users, several technical details in Microsoft's official documentation stand out:

• Office in Copilot: Rather than generating static text blocks that break formatting when pasted elsewhere, asking Copilot from the Home tab to model a budget or draft a launch brief creates or updates a live, editable Word document, Excel workbook, or PowerPoint deck. Teams can co-edit side by side and @mention colleagues—or @Copilot directly inside Teams channels—while keeping version history intact.
• Code and Managed Runtime: Aimed at making app building as routine as writing a memo, Code lets non-developers spin up interactive dashboards or persistent desktop widgets using natural language. To keep shadow IT from spiraling out of control, Microsoft introduced Copilot Managed Runtime (currently in preview), a sandboxed hosting infrastructure inside the customer's Microsoft 365 tenant governed by IT.
• Autopilot (formerly Scout): Reportedly inspired by OpenClaw—and shaped by daily discussions between Lamanna and Autopilot lead Omar Shahine about rival agents like Meta’s Muse and Instinct—Autopilot is an always-on digital teammate. Because it is cloud-hosted within the organization's tenant with its own identity, memory, and virtual workspace, it can monitor Teams channels, chase down supplier review updates, and resume multi-day projects while the user is offline.

Section Summary:

• Unified Architecture: The revamped Copilot app merges conversational Chat, delegated Cowork tasks, natural-language app building (Code), and persistent background agents (Autopilot).
• Deep Data Grounding: Fabric IQ connects Copilot to over 20 million Power BI semantic models, while public previews are adding live context from Dynamics 365 and Power Platform workflows.

The Catch for IT Admins: "Costco" Doesn't Mean All-You-Can-Eat​

Here is where veteran systems administrators need to read the fine print. Calling Copilot the "Costco of AI" makes it sound like a single annual membership fee unlocks unlimited access to everything in the store. In reality, Microsoft is splitting enterprise AI into a two-tier financial model:

1. Fixed-Cost Everyday AI (USL): A standard User Subscription License (around $30 per user per month before bulk discounts) covers standard Copilot Chat, integration across Word, Excel, PowerPoint, Outlook, and Teams, and the Auto model router.
2. Metered Agentic Work (UBB): Heavy-duty features—specifically Cowork, Code, Autopilot, and access to specific frontier models like Astra and Fable—run on Usage-Based Billing (UBB).

To prevent an enthusiastic department from accidentally burning through the quarterly IT budget by leaving Autopilot agents running all weekend, Microsoft is rolling out "FinOps for AI" controls across Agent 365, Insights, and the Copilot admin center. Administrators can now:

• Enforce granular spending policies at scale via API.
• Route end-user credit requests through existing corporate approval workflows.
• Restrict which model families specific user groups are allowed to access—a policy setting that also constrains which models the Auto router can pick for those users.
• Allow end users to view their own credit burn rate, remaining balances, and usage history directly inside the Copilot interface.

Adoption Realities and the Succession Subplot​

Has the multi-model pivot worked? The numbers show real momentum, though with important caveats. According to The Information, Copilot’s corporate users have more than doubled since the beginning of 2026 to surpass 30 million. AT&T, for instance, now deploys roughly 18,000 Copilot seats and credits the platform with handling the "lion's share" of its personal productivity automation.

However, enterprise analysts are careful to distinguish between total corporate users (which can include freemium or bundled chat seats) and paid Microsoft 365 Copilot seats. Microsoft disclosed 15 million paid seats in January 2026 and topped 20 million paying users by May 2026 —meaning that against a total installed base of more than 450 million commercial Microsoft 365 users, well over 90% of the Office customer base has yet to convert to full paid licenses.

Furthermore, Microsoft faces fierce competition from the very labs supplying its models. OpenAI sells ChatGPT Enterprise and Anthropic sells Claude Code and Cowork directly to the same corporate buyers. Changing entrenched habits isn't easy, either: the CIO of National Life Group told The Information that while Copilot has improved markedly since being "useless" in 2023, most of his employees still reach for ChatGPT out of habit.

Still, orchestrating the Copilot turnaround has rapidly elevated Charles Lamanna inside Redmond. After Microsoft's March 2026 reorganization placed him at the helm of the Copilot platform alongside Ryan Roslansky and Perry Clarke—with Jacob Andreou leading the consumer/commercial experience and Mustafa Suleyman heading first-party model development—Lamanna added all of Office to his portfolio in October 2026 following the retirement of longtime executive Rajesh Jha.

His organization has more than tripled over the past year to roughly 25,000 employees spanning Office, Teams, and Copilot engineering. Along with Andreou, Xbox CEO Asha Sharma, security chief Hayete Gallot, and Arun Ulag, Lamanna is one of a select group of executives Nadella has been bringing to quarterly board meetings—a move some insiders interpret as early succession grooming, even if Nadella shows no signs of stepping down anytime soon.

For enterprise customers, the takeaway is clear: Microsoft is no longer betting its kingdom on a single AI lab. By positioning Copilot as an agnostic routing layer and enterprise runtime, Redmond is betting that governance, Office file integration, and wholesale economics will ultimately beat standalone chatbots. Just make sure your FinOps spending limits are configured before you hand out the warehouse club cards.