In December last year, Microsoft told thousands of its engineers, product managers and designers that they could use Claude Code, Anthropic’s command-line coding agent, with company money.
By spring, the tool had spread far beyond engineering: into the kind of non-technical roles that, in previous waves of enterprise software, would have waited years to fill a spot. Within Microsoft, the launch was seen as a learning exercise. Outside of it, the surface signal was simpler.
The world’s largest software company, the one with its own core models and its own coding assistant, had just paid a competitor to introduce a rival product to its workforce.
Six months later, that experiment is coming to an end. According reports in Windows Central and other media following The Verge’s original scoop, Microsoft is canceling most of Claude Code’s direct licenses within its Experiences and Devices group, the division that builds Windows, Microsoft 365, Outlook, Teams and Surface.
Affected engineers have been asked to migrate to GitHub Copilot CLI by June 30, the last day of Microsoft’s fiscal year. The official reason is the unification of the tool chain. The unofficial reason is in the calendar.
The đź’ś of EU technology
The latest rumors from the EU tech scene, a story from our wise founder Boris and some questionable AI art. It’s free, every week, in your inbox. Register now!
Claude’s pullback is the most credible sign yet that the unit economics of enterprise AI coding doesn’t work, at current token prices. Not because the tools are bad. Quite the opposite: They’re good enough for engineers to use constantly, and constant use is what breaks the math.
The clearest evidence is in Uber, which is not Microsoft and does not have the financial cushion of Microsoft. Praveen Neppalli Naga, Uber’s chief technology officer, told The Information in April that the company had exhausted its entire planned 2026 AI coding budget in four months.
For March, Naga’s own figures. grew Claude Code usage from 32 percent to 84 percent of its organization of approximately 5,000 engineers. Individual engineers spent between $500 and $2,000 a month on tokens. About 70 percent of committed code at Uber now originates with AI, and about one in ten live backend updates are sent by an agent with no human involved.
“I’m back to the drawing board.” Naga said, “Because the budget I thought I would need is already gone.”
That sentence is the whole story in miniature. The forecast was incorrect because the variable being forecast, token consumption, behaves nothing like the licenses and positions that finance teams know how to model. A traditional enterprise software agreement is called users.
A symbolic price deal is called how much the model has to think about. Agent coding makes the model think a lot. Sessions last hours, spawn parallel threads, and generate volumes of context that are nothing like the autocomplete interactions that shaped the original pricing structure.
we have been tracing this fracture for months. In November, GitHub stopped new registrations for Copilot Pro and Pro+ because paying customers’ agent workloads were generating costs that exceeded the price of their monthly plan.
The cost structures created for light support, the company admitted, no longer hold.
This is not an Uber or Microsoft problem. It is a condition of the industry. Bryan Catanzaro, vice president of applied deep learning at Nvidia, told Axios in April that, for his team, the cost of computing now far exceeds the cost of the employees who use it.
The chip company says so. Fortune followed up in May by reporting that token-based AI toolswhen used heavily, it can cost more per task than the human engineer was supposed to augment.
A 2024 MIT analysis that has since been widely circulated in financial circles suggests that at current prices, AI automation is cheaper than human labor for about a quarter of the jobs people thought it would replace.
Compare it with spending forecasts. Gartner expects global spending on AI to reach $2.5 trillion this year, up 69 percent from 2025.
The same company now places generative AI squarely at what it calls the nadir of disillusionment, predicting in a May press release that 25 percent of the AI ​​budget planned for 2026 will be moved to 2027 as proofs of concept die in the procurement process.
A separate Gartner reading from April found that only 28 percent of AI infrastructure projects fully meet their business case. That is not the curve of a technology that is going through an uncomfortable adolescence. That is the curve of a market that revalues.
Microsoft’s withdrawal occurs within this price review, and not by chance. There are two ways to read the play. The first is what Microsoft has reported: that Copilot CLI is the strategic destination, that engineers will still have access to Claude’s models within Copilot, and that the company simply wants a product that it can shape directly with GitHub. That story is true.
It’s also a story that Microsoft could have told at any time in the last six months and chose not to. What changed was not the strategic logic. What changed was the bill.
The second reading is more difficult to rule out. Microsoft is in a unique position to know how much it actually costs to use Claude on an enterprise scale, because its own engineers were the most common users outside of Anthropic’s customer base. Within Experiences and Devices, Claude Code had become, according to several reports, the preferred tool.
If the math had improved with scaling, this would be the time when Microsoft would strike a multi-year deal on favorable terms. Rather, it is unfolding the experiment in a window that conveniently closes the books for a fiscal year.
When the company with the most influence in the room moves away from a supplier whose product its own staff prefers, the signal is not about preference.
Whether this constitutes a bubble depends on the definitions. Prices at the token level will fall, as they have fallen by approximately a factor of ten every eighteen months for the past three years. The more interesting question is whether the token consumption per task falls faster than the cost per token.
So far the evidence goes in the opposite direction. Each generation of agent system, by design, consumes more tokens per unit of work, because it reasons longer, plans more elaborately, and verifies itself against the world.
Anthropic’s own infrastructure team has spoken publicly about reasoning workloads that generate an order of magnitude more compute per query than per chat. That is the bet that will be made in the next twelve months of model launches. It is also the bet that put Uber back on the drawing board.
There is a worked example in TNW’s own coverage. In April, Anthropic banned a popular open source agent framework called on OpenClaw to not run on Claude consumer subscriptions, after discovering that individual instances could consume the equivalent of $1,000 to $5,000 in API costs in a day of autonomous operation. The framework was running on a $200/month Max plan.
The economic transfer was so egregious that Anthropic had to write a new clause in its terms of service. Multiply that pattern across a Fortune 500 engineering organization and you get the Uber budget memo.
The counterargument is real and worth stating. The cost of a working AI coding agent compared to the cost of an additional senior engineer is, even at current prices, often favorable on a per-feature basis. The increase in productivity is documented; the substitution is happening. What is breaking is not the value proposition.
It is the acquisition model. Companies that signed up for a productivity tool are finding that they signed up for a metered utility, and the meter runs when no one is looking. The solution may be simple: limited budgets per engineer, tiered access for high-leverage roles, agent runtime quotas.
Many of the biggest buyers are already there. But the implication is that the era of “give every employee a Claude Code role” is coming to an end, and what replaces it will look more like AWS billing than Office licenses.
That’s what Microsoft’s quiet email to its Windows and Surface PCs is really announcing. It’s not the end of AI coding. Not even the end of Anthropic at Microsoft, as Claude models will still be accessible through the Copilot CLI.
It heralds the end of the experimental phase, the phase in which the world’s largest software companies were willing to absorb arbitrary token costs in exchange for learning. The learning is done.
What comes next is the hardest part. Companies will continue to purchase AI coding tools because the productivity is real and the competitive pressure is relentless. But they will buy them the same way they buy electricity, with usage limits, with shadow meters, with a finance team present.
Somewhere in a Microsoft conference room earlier this spring, someone looked at a Claude Code invoice, did the arithmetic with a Copilot CLI roadmap, and made a decision.
Now the same arithmetic is being done in all the CFO offices that agreed to the December 2025 launch. The withdrawal will not be noisy. They will be a series of end-of-fiscal-year emails, sent on a deadline that no one noticed until the budget was already over.
