Microsoft told its engineers Claude Code was too expensive. Then started selling Anthropic's AI at a 65% markup.

In December 2025, Microsoft gave thousands of engineers in its Experiences and Devices division access to Claude Code. Windows, Microsoft 365, Outlook, Teams, Surface. All the products most people actually use every day. The engineers used it a lot. Then in May 2026, the licences got cancelled and they were told to switch to GitHub Copilot CLI by 30 June.

The official reason: toolchain unification. The real reason, which didn't require much reading between the lines: the tool had become, in someone's internal words, "perhaps a little too popular."

(I want to dwell on that phrase for a second. "Perhaps a little too popular." In twenty years of running a software business, I don't think I've ever told a client their tool was cancelled because their team liked it too much. But here we are.)

I use Claude Code myself. I've been watching the enterprise cost story build for months, and the Microsoft cancellation didn't surprise me. What I didn't see getting enough attention was the bit buried underneath the headline. Microsoft cancelled Claude Code for its own engineers on cost grounds. In the same week, the company was selling Anthropic's underlying agentic technology to enterprise customers through Microsoft 365 E7, white-labelled as "Copilot Cowork," at what Ben Thompson at Stratechery called a 65% premium over the prior tier.

Microsoft cancelled Claude Code for cost reasons. Then kept selling you Anthropic's AI at a 65 percent premium through Copilot Cowork. That's a business model. I'm not sure it's the one they meant to build.

What Microsoft Actually Did

Tom Warren at The Verge broke the story on 14 May 2026 through his Notepad subscriber newsletter. The engineers affected were in the Experiences and Devices division, covering Windows, Microsoft 365, Outlook, Teams, and Surface. Not all of Microsoft. Azure, the Developer Division, and the AI Platform group weren't named as affected. Worth noting, because a lot of the initial coverage framed this as a company-wide retreat from Claude Code, which it wasn't.

The timeline: Microsoft brought Claude Code in for the E+D division in December 2025. Over the following months, access expanded beyond engineers to product managers and designers. By May 2026, the bill had presumably reached a number that someone in finance found alarming enough to act on. Tom Warren's confirming tweet went up on the 14th, and the internal memo from Rajesh Jha (EVP, Experiences + Devices), as reported by Tom Warren, characterised Claude Code as "an important part of that learning."

Which is a generous way to describe a tool you're cancelling. I've written client emails that were less gracious about software that genuinely didn't work.

Jha's framing for why Copilot CLI was the better choice: "Copilot CLI has given us something especially important: a product we can help shape directly with GitHub." That's the polite corporate version of "we own GitHub, so we should probably be using the thing we own." Fair enough, actually. That logic does hold up. The cost logic holds up better, but this one's not wrong either.

Now, here's the context that makes the whole thing awkward.

A JetBrains Developer Ecosystem Survey from April 2026, covering more than 10,000 developers, puts overall usage at 29% for Copilot and 18% for Claude Code across all experience levels. Copilot leads on raw reach. But the Microsoft story isn't really about which tool more developers happen to have installed. It's about which tool Microsoft's own engineers chose to use, enthusiastically enough, persistently enough, and in large enough numbers that the usage created a budget problem that had to be solved by cancellation.

The internal characterisation ("perhaps a little too popular") does the work here. That's not a description of a tool people were using reluctantly. When your engineers vote with six months of token consumption, that's a specific category of internal product feedback. The kind that doesn't go in a press release.

The June 30 deadline isn't arbitrary either. It's the end of Microsoft's fiscal year. Budget consolidation, cost centres being squared off before the books close. That timing tells you something about which explanation is doing the heavy lifting.

The Contradiction

The same week Microsoft cancelled Claude Code for its engineers, it was selling Anthropic's agentic AI through Microsoft 365 E7 under the name "Copilot Cowork" at $30 per user per month. Ben Thompson at Stratechery's Major Matters newsletter called this a 65% premium over the prior E5 tier.

Let me state that plainly. Microsoft removed Anthropic's tool from its engineers' desks. Then it charged enterprise customers to put Anthropic's tool on their engineers' desks, at a markup.

That's not hypocrisy, technically. It's a business model. But the gap between "we can't afford this internally" and "you should absolutely buy this, rebranded, for $30 a seat" is wide enough to park a Surface Studio in.

The week of 14 May 2026 produced three separate Microsoft AI stories within three days, and they rhyme in a way that's hard to ignore. On 14 May: the Claude Code cancellation. On 16 May: Microsoft announced it was removing free Copilot access from consumer Office apps. The free tier, which had been part of the pitch that Microsoft was putting AI everywhere, was quietly walked back. On 17 May: Mat Velloso, former Technical Advisor to CEO Satya Nadella at Microsoft (now at Meta's Superintelligence Labs), said publicly that Microsoft had "missed the AI wave" the way it missed mobile and the internet.

Velloso also noted that only 3.3% of Microsoft's 450 million Microsoft 365 users had converted to paid Copilot seats. That's roughly 15 million people out of 450 million. Microsoft built Copilot, embedded it across every product in its portfolio, spent billions, and converted 3.3% of its user base to paid seats.

Three stories. One narrative: Microsoft spent billions building Copilot, its own senior engineers preferred a competitor's product, and the bills for that competitor's product had to be cancelled on cost grounds.

When your internal AI rollout gets described as "perhaps a little too popular," you've got an unusual problem. Most products hope to be too popular. Microsoft's response was to make it unavailable.

There's also a broader cost model story here. From 1 June 2026, GitHub moved Copilot Pro and Pro+ from flat-rate subscriptions to usage-based billing. That's not coincidental timing. The whole industry's grappling with the same arithmetic: the more useful an agentic AI coding tool is, the more tokens it consumes, and the more it costs. The pricing model that made sense for autocomplete doesn't scale to autonomous agents. GitHub changing its billing structure is the industry acknowledging this is a system-wide problem, not just a Microsoft one.

What the Developers Think

The Hacker News thread on the cancellation reached number two with 492 points and 465 comments. The most useful observation in the thread was this: "quantifying tokens used is analogous to quantifying sawdust generated on a construction site."

The point being that token consumption isn't the right thing to optimise. Productivity is. If an engineer spends $200 worth of tokens in a day and ships three weeks of work, that's not a cost problem. It's a cost model problem. The sawdust analogy is good because it gets at why the instinct to cut usage is the wrong instinct. You don't build less to generate less sawdust. You build a better sawdust budget.

The Uber numbers are the clearest external data point we have on what this actually looks like at scale. Uber deployed Claude Code to around 5,000 engineers. Usage rates hit 84 to 95 percent. Per-engineer monthly costs came in between $500 and $2,000. Uber's CTO, Praveen Neppalli Naga, said: "The budget I thought I would need is blown away already." The company exhausted its entire planned 2026 AI budget within four months.

I've written before about a related situation, where a major US enterprise discovered they'd spent $500 million on Claude API calls without a spending limit in place. (Yes, $500 million. Yes, no spending limit. Yes, that's a real thing that happened.)

Related Article9 min read

How do you accidentally spend $500M on AI?

An unnamed enterprise company spent $500 million on Claude AI in a single month because nobody configured a spending limit. Uber burned its 2026 AI...

Read full article

The pattern's consistent. Agentic AI coding tools, used enthusiastically by capable engineers, generate token costs that don't look like traditional software seat licences. Enterprise procurement wasn't built for this. The budgets weren't sized for it. The governance frameworks weren't designed for it. Everyone who predicted this would happen was right. The industry's now working out what to do about being right.

The developers in that Hacker News thread don't think the answer is mandating a less-preferred tool. They think the answer is better cost governance and usage limits. That seems right to me. Whether Microsoft's finance team agrees is a separate question, and finance teams are usually louder in the room.

What This Means If You're Running a Dev Team in Australia

I'm going to resist the urge to write five bold-header bullet points here (you know, the LinkedIn listicle format where every insight gets its own dramatic heading). This story is more of a narrative than a checklist. But there are a few genuinely useful things to take from it.

The first is that the best tool and the most affordable tool are often not the same thing, and you need to know which you're choosing before you commit. Overall, Copilot leads on reach (29% usage vs 18% for Claude Code across all developers, per JetBrains April 2026). But "more widely installed" and "more enthusiastically used" aren't the same thing, and it's the enthusiastic usage that generates the token bill. If your team genuinely likes the tool and uses it for real work, the costs scale with that enthusiasm. Before committing to a tool, run a realistic pilot at realistic usage levels. Not a careful, everyone's-on-their-best-behaviour ten-person trial. A proper pilot with real workloads. Then price out the bill at the scale of your actual team.

The second thing: "toolchain unification" is real, but it usually comes second. Microsoft's official reason for the cancellation is plausible. There are genuine advantages to engineers using tools the company directly controls. Tighter integration, better security posture, direct influence on the roadmap. Those things are real. But the cost driver came first, and the official reason arrived later. When a vendor gives you an official reason and a financial reason, the financial reason is usually doing more of the work.

Third, and probably the most practically useful: usage-based pricing needs different governance than seat licensing. Microsoft's Claude Code problem, Uber's Claude Code problem, and the $500 million mystery company problem are all the same problem wearing different clothes. Token-based pricing scales with how useful the tool is. That's the opposite of how enterprise software procurement was designed. Set usage limits before rollout, not after you get the invoice and have to explain it to someone.

And finally, the developer preference gap matters for productivity in ways that are hard to quantify. If your engineers prefer Tool A and you mandate Tool B, you don't save money on Tool A. You save money and lose something in output quality and morale. I've seen this play out with other tools at Webcoda over the years. We once migrated away from a tool the team genuinely liked because the licensing got expensive, and the productivity dip lasted longer than anyone expected. The savings looked clean on a spreadsheet. In practice, not so much.

For context on what proper AI permission governance looks like before you roll out any of these tools:

Related Article11 min read

Anthropic told its paying customers Claude Code was fine. It wasn't. For a month.

From 4 March to 20 April 2026, three separate bugs quietly degraded Claude Code's performance while developers were told they were imagining it....

Read full article

Closing

Microsoft built GitHub. It owns Copilot. It has Azure AI, Azure OpenAI Service, and whatever's left of Cortana (not much, to be fair, and I'm not going to pretend I miss it). It has more AI infrastructure than almost any company on earth. Its own engineers still picked Claude Code, often enough and enthusiastically enough that the licences had to be cancelled on cost grounds.

The toolchain unification announcement is technically a business decision and financially a necessary one. It's also, whether Microsoft intended it this way or not, a product verdict. Delivered not by a press release or a Gartner report, but by thousands of engineers voting with their token consumption for six months straight.

I'm not switching away from Claude Code based on one enterprise cancellation. But I did spend about ten minutes this week thinking through what we'd do if Anthropic's pricing changed significantly, or if the tool shifted in a direction that didn't work for us. That's not paranoia. That's just what Microsoft's experience reminded me to do.

The best tool today isn't always the accessible tool tomorrow. Have a backup. Know what it is. Don't wait for the invoice to tell you.

Job done.

Key Takeaways

• Microsoft cancelled Claude Code licences for engineers in its Experiences and Devices division in May 2026, with a migration deadline of 30 June 2026
• The official reason is toolchain unification. The real driver is token-based costs scaling faster than expected with enthusiastic usage
• Microsoft continues to sell Anthropic's agentic AI sold as "Copilot Cowork" through M365 E7 at $30 per user per month, described as a 65% premium over the prior E5 tier
• JetBrains survey (April 2026, 10,000+ developers): Copilot leads on overall usage (29% vs 18% for Claude Code), but Microsoft's engineers used Claude Code enthusiastically enough to create a budget problem. That's the real preference signal
• Uber's Claude Code deployment cost $500 to $2,000 per engineer per month and exhausted the company's full 2026 AI budget in four months
• GitHub moved Copilot Pro and Pro+ to usage-based billing from 1 June 2026, confirming this is an industry-wide cost model problem
• For Australian teams: pilot at realistic scale, set usage limits before rollout, and keep a backup tool in mind