DEV Community

Cover image for Microsoft Just Killed Claude Code Internally. Their Own Devs Loved It.
Phil Rentier Digital
Phil Rentier Digital

Posted on • Originally published at rentierdigital.xyz

Microsoft Just Killed Claude Code Internally. Their Own Devs Loved It.

#ai #technology #claudecode #microsoft

Microsoft just spent 6 months proving Claude Code is better than their own tool. Then they cancelled it. That sentence is the whole story.

The Experiences + Devices division (the one running Windows, 365, Outlook, Teams, Surface) had rolled out Claude Code alongside their in-house Copilot CLI back in December. 6 months of side-by-side benchmark, two tools, same engineers. On May 14, Rajesh Jha announced the Claude Code licences go dark on June 30. The official reason leaked to The Verge: "very popular, perhaps a little too popular". The good tool got killed because it worked too well, which is the kind of plot twist that only happens at companies big enough to call it a strategy.

TLDR: Microsoft benchmarked Claude Code against Copilot CLI for 6 months. Their own devs preferred Claude Code. Microsoft killed the winner. The reason has nothing to do with the model and everything to do with who owns the harness, and the mechanism has a name from 2001 that explains why this concerns you too.

(lol, I really don't like m$ 😂. This news makes my day.)

This is not a tech decision. It's a confession. A 3-trillion-dollar company ran the test, its own engineers voted with their hands, and management shot the winner. To protect the mediocre one it owns. There's a name for this, an ex-Microsoft engineer coined it 25 years ago, and the mechanism is fractal. It plays out at every scale of the AI ecosystem right now, including yours.

The Press Got Half The Story. It Was Never About The Model.

Every outlet ran with the same framing. Vendor lock-in. Cost cutting. Consolidation. Fine, all of that is technically true. None of it is the actual point.

Here's the piece nobody zoomed on. The official quote from Rajesh Jha, EVP Experiences + Devices: "Copilot CLI has given us something especially important: a product we can help shape directly with GitHub for Microsoft's repos, workflows, security expectations, and engineering needs." Translation: we keep the one we own.

The June 30 deadline lines up exactly with Microsoft's fiscal year end. So yes, there's a cost-cutting angle baked in. m$ being m$, no surprise there.

But the real tell is what stays. Anthropic's models remain accessible through Copilot CLI. The Foundry agreement with Anthropic is untouched. Microsoft engineers will still talk to Claude. They just won't talk to it through Anthropic's interface.

Microsoft did not reject the model. Microsoft rejected the harness.

That distinction matters because it reframes the whole story. This is not about which AI is smarter. It's about which company gets to define the workflow shape around the AI. And in any healthy ecosystem, the answer would be complicated, because Anthropic has not exactly been a poster child of trust lately either. April was a mess. 3 distinct bugs in Claude Code, weeks of denial, blaming users, before a public postmortem finally admitted reasoning effort had been silently downgraded from high to medium on March 4. On dev trust, Anthropic is not in a strong position.

But on the internal Microsoft benchmark, Claude Code won. The engineers voted. Management overrode them anyway.

This Has A Name. An Ex-Microsoft Engineer Coined It In 2001.

Microsoft has form with this. Back in 2001, an ex-Microsoft engineer named Ben Slivka coined the term strategy tax to describe what happens when you cripple a product so it doesn't threaten another product in the same portfolio. At the time, Microsoft was holding Internet Explorer back so it wouldn't cannibalise Word.

Same playbook, 25 years later. Except this time it's applied to a third-party partner instead of an internal team. Claude Code gets killed so Copilot CLI doesn't look as bad next to it.

The name fits because the mechanism is identical. Different decade, different victim, same org chart logic.

The Pattern Is Fractal. I've Been Watching It At 2 Scales.

Strategy tax doesn't happen once, at the top. It plays out at every scale of the AI ecosystem at the same time, and once you've seen it once you can't unsee it. It's a raid boss mechanic that doesn't care about your gear score, only about who pulled aggro on the harness.

Scale 1: Anthropic vs anything that isn't Claude Code. The Terms of Service forbid channelling a Max subscription through anything other than Claude Code. OAuth bans for users who tried. Silent nerfs of reasoning effort in March 2026 (and Sonnet really struggles compared to Opus when this gets dialled down, on long context especially). Third-party agent layers that got too good at consuming Max tokens efficiently saw the cable pulled. Not out of malice, out of strategy tax. Anthropic protects its own harness against alternative harnesses, including the ones built by its own community. I lived through the time Anthropic killed my $200/month setup, and at the time I thought it was personal. It wasn't. It was structural.

Scale 2: Microsoft vs Anthropic. Same mechanism, but this time Anthropic is the one paying the tax instead of collecting it. The 3-trillion-dollar harness owner protecting its harness against the better harness, even when its own engineers prefer the better one.

Same pattern, 2 scales. Anthropic taxes its heavy users to protect Claude Code. Microsoft taxes Anthropic to protect Copilot CLI. The mechanism doesn't care about the size of the player. It only cares about who owns the harness.

A developer on the ResetEra thread covering the news put it bluntly: "Sometimes I feel like the harness is more important than the model these days, due to needing to manage external tool access and permissions." That's the whole game, said in one sentence, by someone who isn't selling anything.

The Model Is The Commodity. The Harness Is The Leverage.

You've seen the first half of this argument before, probably here. AI models have become substitutable commodities. Claude, GPT, Gemini, Llama, Kimi, Qwen, GLM, swappable in one environment variable.

The harness doesn't swap.

Your CLAUDE.md, your slash commands, your hooks, your MCP setup, your cached prompts, the muscle memory of your workflow, the CI integrations, the cron jobs that call your CLI at 4am, all of that is invested time. Vendors fight at the harness level because that's where the leverage sits. I went deep on why CLIs beat MCP for AI agents a while back if you want the technical version.

Here's the part I haven't written about yet, and the one Microsoft just illustrated for free.

Even when you own the harness in theory, you don't really own it if the conventions belong to a vendor. A CLAUDE.md is an Anthropic convention. Slash commands are Anthropic-shaped. Hooks fire on events that Anthropic's runtime defines. You can build your own CLI, your own agent layer, your own orchestration framework, and still be renting the harness through the back door. Because the people you're going to hire, the prompts you're going to share on Discord, the patterns you'll find on GitHub, all assume the conventions of the dominant vendor.

The model is one variable. The conventions are a year of muscle memory across a team.

I think this is the part most builders underestimate. You can swap the engine. You can't swap the steering wheel without making everyone learn to drive again.

Microsoft did the math at 3 trillion dollars. They concluded that owning a mediocre harness whose conventions they control was worth more than renting an excellent one whose conventions Anthropic dictates. If m$ can't afford to use the best harness on the market, your side project definitely can't.

Honestly not sure if this scales down cleanly to a solo builder. Maybe for a 1-person shop the convention lock-in is acceptable because the time saved beats the dependency cost. But the moment you have a team and a payroll, the calculation flips. Microsoft just showed where it flips.

Last week my kid asked me why we couldn't just use the best of everything. I tried to explain the strategy tax to a 9-year-old. He went back to his Lego after 30 seconds. Smart kid.

Karen from Accounting, the one who signed off on this kind of decision in 2001, signed off on it again in 2026. The job titles change. The reflex doesn't.

They Don't Have It Anymore. Not Because It Was Bad. Because It Was Too Good.

Next time a vendor hands you a free or subsidised harness (Cursor Pro, Copilot Free, Claude Pro with Code included, any "tier on the house"), ask yourself who's paying the strategy tax. If you can't tell, it's you. And you'll only find out the day the vendor decides your usage is "perhaps a little too popular".

Microsoft engineers preferred Claude Code. They don't have it anymore. Not because the tool was bad.

Because it was too good for them to be allowed to use it.

Sources

This post may contain affiliate links. If you click them, I might earn a small commission, costs you nothing, and helps me keep shipping quality articles every day for your reading pleasure.

Top comments (0)