The AI coding bill just came due

TechCrunch reported that Uber spent part of last year ranking its engineers on internal leaderboards by how much they used AI coding tools. By June, the company had burned through its entire 2026 AI coding budget in four months, with some engineers generating monthly bills between $500 and $2,000 on tools like Claude Code and Cursor. The leaderboard worked. That was the problem.

This is the week the AI coding bill came due. Three stories landed in quick succession that tell a single story about price discovery: the moment an industry finds out what something actually costs when you remove the artificial ceiling.

Microsoft is pulling Claude Code licences across its Experiences + Devices division, forcing engineers back onto GitHub Copilot CLI by 30 June. The timing is not subtle: that's the last day of Microsoft's fiscal year, and engineers in the division reportedly described a significant feature gap between Claude Code and Copilot CLI. Microsoft's response wasn't to close the gap. It was to remove the option.

Meanwhile, power users of GitHub Copilot are reporting cost increases of 10x to 50x after its switch to token-based billing on 1 June — monthly bills jumping from $29 to $750, and from $50 to $3,000, specifically for the agentic coding workflows that GitHub spent the last year encouraging developers to adopt. Each plan now includes AI credits equal to its subscription price, with overages billed by token. The backlash from developers has been swift.

The conventional reading is that companies are tightening belts. That's true but boring. The more interesting reading: flat-rate pricing for AI coding tools was a subsidy, and subsidies always end.

This pattern has a direct precedent in cloud computing. Amazon, Google, and Microsoft spent the early 2010s offering generous free tiers and predictable pricing to drive adoption. Once workloads were entrenched, the billing models grew more granular, more usage-based, and for heavy users, more expensive. The playbook worked because switching costs were high. The question for AI coding tools is whether the same lock-in applies.

I think the answer is: not yet, and that's what makes this moment unstable. Uber capped spending at $1,500 per tool per engineer per month. Microsoft didn't cap; it chose sides, betting that its own tool would improve fast enough that engineers would stop noticing what they lost. GitHub moved to usage-based billing, which is honest about costs but punishing for the heaviest users, precisely the developers who evangelised the product.

The practical takeaway for anyone building or budgeting: agentic AI workflows consume tokens at a fundamentally different rate than code completions, and most pricing and procurement frameworks haven't caught up. It's the autonomous, multi-step sessions where a model reads your codebase, plans changes, and executes across files that blow through token budgets. The industry priced AI coding tools for the autocomplete era and is now discovering what the agentic era actually costs.

For developers, the immediate risk isn't losing access to these tools. It's losing the best versions of them. Microsoft engineers didn't get told to stop using AI. They got told to use the one their employer controls. Uber engineers didn't get cut off. They got metered. The pattern is companies moving from "use as much as you want" to "use what we can afford to give you," and those two things produce very different outcomes.

The companies that figure out how to budget for agentic AI without throttling it will ship faster than the ones that panic at the invoice. Right now, most are panicking.

---