Microsoft Cut a $1 Billion Claude Bill and Bet the House on One Copilot

By Grace Kim

Microsoft talked its employees down from a projected $1 billion bill for internal use of Anthropic's Claude, cutting that spending by more than a third, The Information's Aaron Holmes, Jyoti Mann and Kevin McLaughlin reported on October 5. The episode captures the company's larger pivot: after a decade of riding OpenAI's coattails, Microsoft spent the past year and a half rebuilding AI in its own image.

The numbers behind that effort are big. Microsoft 365 Copilot passed 30 million paid seats by the July earnings call, CEO Satya Nadella said — up from 20 million in April, and higher than the most recent paid enterprise seat counts reported by Google or OpenAI. Adoption is another matter: a February survey by Recon Analytics found only 18% of workers prefer Copilot when their employer also provides ChatGPT.

What changed with OpenAI?

Microsoft's prescient investment in OpenAI gave it an early lead in generative AI, but the partnership grew fraught and has since been revised to play a smaller role in Microsoft's AI future. Nadella warned against complacency back in 2023: "As far as I'm concerned, early leads in technology don't matter."

The answer was self-sufficiency. Mustafa Suleyman, the DeepMind and Inflection AI cofounder who joined as CEO of Microsoft AI in 2024, now runs frontier model development in-house. "Essentially the core mission is self-sufficiency for Microsoft, ensuring that we have absolute frontier models available that have been highly tuned to our products," he says. His forecast: "I think we're going to produce the best models in the world in 2027."

Why one Copilot instead of a hundred?

Experts have mapped at least 80 — possibly more than 100 — separate products called Copilot. Charles Lamanna, a 13-year veteran now Executive VP of Microsoft Copilot, Agents, and Platform (and flagged by The Information as a possible future CEO), says the new unified app "will be the end of 'Copilot-palooza.'"

The app merges consumer and business Copilot into one experience resembling ChatGPT or Claude. "There's one Copilot for your personal life and your work life," Lamanna says. "It may have different UIs on it, but it's all the same Copilot behind the scenes." Jacob Andreou, the former Snap executive who joined in February 2025 and now oversees the app, describes the rebuild plainly: "The last year and a half has been a story of stitching all of that together." The app is in private preview for Microsoft's Frontier program, with no general availability date announced.

How does a 200,000-person company share what it learns?

In early 2025, Nadella concluded that internal "learning loops" were too slow. The fix is an AI accelerator: a giant Teams meeting every Thursday at 1 p.m., drawing 130 to 170 staffers from Xbox, LinkedIn, GitHub, Azure, Office and Windows. A related Teams channel counts about 1,000 members.

It produces real outcomes. Omar Shahine, corporate VP for Word since the 1990s-era product line, experimented with OpenClaw last February, gave an accelerator talk, and within weeks was running a new team building Microsoft agents on OpenClaw-based infrastructure — internally nicknamed "chief claw officer." The first agent, renamed from Scout to Autopilot, will run 24/7 in the cloud. An early internal version hit 40,000 weekly active users with 77% week-over-week retention. "Some of the biggest adoption we've seen internally has actually been in nontechnical teams," Andreou says.

Not every structure survives: The Verge's Tom Warren reported last week that Microsoft's beloved hacking network, The Garage, is shutting down.

Where are the new bottlenecks?

CTO Kevin Scott says that at Microsoft, "for net-new things, all of that code is getting written by AI right now," though he declines company-wide statistics. One study says 42% of developers now use AI to generate at least half their code, and 71% work in Visual Studio Code per a Stack Overflow survey.

AI removed one bottleneck and created others. "Writing lines of code is just no longer the bottleneck," Andreou says. So Microsoft physically colocates people — 30 to 40 staffers sat in one room for eight weeks building Copilot Tasks, now Copilot Cowork. The next choke point is verification. "How do you verify the code and make sure it hits our trust and safety and security and compliance bars? The pain point is that code review bottleneck," says Caitie McCaffrey, a member of the technical staff. Tim Bozarth, corporate VP of engineering for Microsoft CoreAI, adds: "Right now, that's the place where a ton of energy has shifted."

On spending, the "tokenmaxxing" fever has broken. Security VP David Weston: "V12 Lamborghinis are great, but you don't need to use them to go pick up the kids from soccer." Microsoft says divisions should manage token use "thoughtfully and efficiently" and recommends its auto model picker. Scott notes employees "are increasingly choosing to use Microsoft AI models, which is awesome to see."

Can distribution beat model quality?

Lamanna's target is explicit: "I joke with my team that we want to be the Costco of AI... We want to get to 400 million, 500 million [users]. As many people as use Office today should be using Copilot."

Analysts split on whether that works. Forrester's Tracy Woo calls Office "crappy" despite its ubiquity and warns that "best-of-breed and enterprise traction aren't necessarily one and the same." TECHnalysis Research president Bob O'Donnell disagrees: "They certainly haven't lost the race, even though OpenAI and Anthropic get all the attention these days." Constellation Insights editor-in-chief Larry Dignan sees progress already: "Copilot seems like it's a lot better than it used to be, because they're not the middleman anymore between you and OpenAI."

GitHub COO Kyle Daigle, who added the title CMO, developer in January, frames the bet in one line: "The real question is, 'How many people can we enable to do something that they couldn't do before?'"