Microsoft Gave Its Engineers the Best AI Coding Tool on the Market. Then It Took It Away Because the Bill Got Too High.

Microsoft opened access to Anthropic's Claude Code across its Experiences and Devices division last December. That's the group behind Windows, Microsoft 365, Outlook, Teams, and Surface. Developers got it. Project managers got it. Designers got it. Everyone was encouraged to experiment. TechRadar reported that the tool became popular fast. Too fast, as it turned out, for the budget that was supposed to fund it.

Microsoft is now canceling most internal Claude Code licenses with a June 30 deadline, steering engineers toward its own GitHub Copilot CLI instead. The timing aligns with the end of Microsoft's fiscal year. It is not a complicated decision. The company watched token-based billing eat through the AI budget and decided to route engineers toward the tool it owns outright rather than the one it pays Anthropic for by the token.

Rajesh Jha, Microsoft's VP of the Experiences and Devices group, told The Verge that the original goal was to benchmark both tools in real engineering workflows and understand what best supported their teams. The benchmarking is over. The conclusion: the tool that costs Microsoft nothing to license beats the tool that got more expensive the more people used it.

That sentence is the whole story. Not just for Microsoft. For every enterprise running AI tools at scale right now.

The math that broke the budget

Every prompt, code review, debugging session, and generation request burns through tokens. Tokens are the base unit of computation that AI models process. At the individual level, the numbers are manageable. Anthropic's own documentation puts Claude Code usage at roughly $6 per developer per day on average, with 90% of users staying below $12.

But "per developer per day" is not how enterprise finance teams think. They think in thousands of developers, every day, for months. And at that scale, the numbers compound into something nobody modeled for. Fortune reported that heavy usage can push individual engineer API bills from a flat fee to anywhere between $500 and $2,000 per month.

Bryan Catanzaro, Nvidia's VP of applied deep learning, put it plainly in an interview with Axios: "For my team, the cost of compute is far beyond the costs of the employees."

That quote came from the company selling the infrastructure. The picks-and-shovels supplier just said the shovels cost more than the miners.

Uber is the cautionary tale that proves the pattern

Uber rolled out Claude Code to 5,000 engineers. The company incentivized adoption through internal leaderboards ranking teams by AI tool usage. It worked. By April 2026, monthly usage rates had climbed to between 84% and 95% across the engineering org. Per-engineer API costs hit $500 to $2,000 per month.

The result: Uber burned through its entire $3.4 billion 2026 AI budget in four months. Not half. Not most of it. The entire annual allocation. Gone before summer.

CTO Praveen Neppalli Naga told The Information he was "back to the drawing board, because the budget I thought I would need is blown away already."

Uber is not an edge case. It is the preview.

Cheaper tokens do not mean cheaper AI

Here is where it gets worse for every CFO building budget projections on the assumption that AI costs will come down.

Goldman Sachs recently forecast that agentic AI could drive a 24-fold increase in token consumption by 2030, reaching roughly 120 quadrillion tokens per month. The structural problem underneath that number is what actually matters.

A Gartner report found that inference costs on a one-trillion-parameter model will likely drop nearly 90% by 2030 compared to 2025. That sounds like relief until you read the rest: agentic models consume far more tokens per task than standard models, consumption growth can outpace falling unit costs, and AI providers are unlikely to pass all savings through to customers.

Gartner senior director analyst Will Sommer warned that product leaders "should not confuse the deflation of commodity tokens with the democratization of frontier reasoning."

That sentence belongs on a wall in every enterprise procurement office in the country.

Agentic workflows are the reason. A standard prompt to a language model burns through a predictable amount of tokens. An agentic workflow where the AI is planning tasks, calling tools, debugging its own output, and iterating through multi-step processes can consume a thousand times more tokens to accomplish a single instruction. The cost curve is nonlinear. Enterprise budgets are not designed for nonlinear.

The pricing model that created the problem is being rebuilt in real time

GitHub is moving Copilot from flat-rate subscriptions to usage-based billing starting June 1, 2026, replacing premium request units with credits tied to token consumption. One developer calculated that their projected monthly cost would jump from roughly €67 to about €966 under the new model.

Read that again. The tool Microsoft is funneling its own engineers toward is itself moving toward the exact pricing structure that caused the problem in the first place.

Zylo's 2026 data shows that organizations are now spending an average of $1.2 million on AI-native applications, a 108% increase from 2025. And 78% of IT leaders surveyed reported unexpected charges tied to consumption-based AI pricing. These are not billing errors. These are structural surprises baked into how vendors charge for AI, and finance teams are getting blindsided.

For nearly two decades, enterprise software was bought by the seat. You knew the number. You planned around it. AI broke that model. Tokens do not scale like headcount. They scale with usage, and usage in agentic workflows multiplies geometrically.

What this means for the enterprise AI bet

Microsoft's pullback is not a rejection of AI coding tools. This is a company choosing the tool it owns outright over the one it has to pay a competitor for on a per-token basis. The math is obvious.

But zoom out and the picture gets worse. The biggest, most well-resourced technology company on the planet gave its engineers an AI tool, watched them love it, and then took it away because the bill got too high. Microsoft staffers reportedly preferred Claude Code over Copilot because of the feature gap between the two products. The engineers are being moved to the cheaper option. Not the better one.

If Microsoft cannot make the economics work at their scale, the rest of the industry should be asking what chance they have.

Jensen Huang has said he envisions 100 AI agents working alongside every employee at Nvidia. If that is where the industry is heading, and it is, the invoices are going to look like nothing any CFO has seen before.

The companies that survive this phase will be the ones that stopped measuring adoption and started measuring return before signing off on the next wave of agent deployments. Most organizations have not done that yet. Most still believe more usage equals more value.

The invoices are starting to tell a different story. And the company that writes 30% of its own code with AI just admitted it can't afford the one its engineers actually wanted to use.