# AI Coding Was Never Cheap. You Were Just Being Subsidized. > Source: > Published: 2026-06-24 03:52:38+00:00 On June 1, GitHub Copilot switched to token-based billing, and developers did the math out loud in public. A plan that cost $29 a month could, under heavy use, run past $750. The replies were the part worth reading — not the outrage, the fear underneath it. People who had quietly built their whole workflow around an AI assistant suddenly didn't know what next month would cost. If you've used Claude Code seriously, you already felt the edge of this. One developer reported burning roughly $1,850 of usage in thirty days — on a plan that looked, on paper, like it cost a fraction of that. The agent reads more files, plans more, retries more. Every one of those tokens is real money. You just weren't the one paying for them. Here's the uncomfortable part, and almost nobody is saying it plainly: the cheap era of AI coding was never real. It was a subsidy. And the subsidy is ending. **The price you paid was never the price it cost** For two years, the AI coding tools competed on one thing: getting you hooked. Flat-rate plans. "Unlimited" tiers. Twenty dollars a month for a tool that was clearly costing far more than twenty dollars a month to run against your codebase. That gap didn't come from efficiency. It came from venture capital, paying down your bill to buy your habit. It worked. A generation of developers — solo builders especially — internalized a dangerous reflex: building is basically free now, so just build. Why validate, why scope, why hesitate, when you can spin up the whole thing this weekend for the price of a sandwich? But the unit economics were always sitting there, waiting. Running a frontier model across a real repository, turn after turn, is expensive. The companies absorbing that cost were never going to absorb it forever. So when Copilot moves to token billing and Claude Code usage starts looking like an AWS invoice, that's not a betrayal. It's the meter finally being switched on. The number was always there. Now you can see it. **What this actually breaks** The honest version of this story isn't "tools got greedy." It's worse, and more useful to hear. A lot of indie SaaS economics that looked like they worked were quietly leaning on subsidized tooling. The math on a small product — build cost near zero, run cost near zero, so even a few customers means profit — only held because the most expensive part of the stack was being paid by someone else. Move that cost onto the builder and a chunk of "viable" side projects stop being viable. And the psychological damage is bigger than the financial one. When building feels free, you build carelessly. You skip the boring question — does anyone actually want this? — because skipping it costs nothing. The graveyard of 2026 isn't full of people who couldn't code. It's full of people who could build anything, instantly, and so they built things nobody wanted, faster than ever, for free. Now it won't even be free. That's the anxiety humming under the Copilot replies. Not "this is too expensive." It's "I'm not sure I know what my own building habit is worth anymore." **The skill nobody was selling** Every indie thread for the last year has told you the bottleneck moved from building to distribution. That's already becoming a cliché. Here's the one underneath it that the new bills are forcing into view: the skill quietly becoming valuable in 2026 isn't prompting, and it isn't even marketing. It's cost discipline. It looks unglamorous. Treating tokens like the real expense they are. Scoping a task tightly enough that the agent doesn't wander through forty files to fix one. Routing the cheap, dumb work to a cheap, dumb model instead of asking your most expensive agent to rename a variable. Measuring what you actually burn instead of trusting a flat-rate plan to hide it. Builders who do this seriously report cutting their costs by 77–91% — which is the difference between a project that survives the subsidy ending and one that doesn't. None of that is exciting. None of it goes viral. But it's the boring competence that separates the builders who keep shipping in 2026 from the ones who quietly stop when the bill arrives. **The meter is on now** The free lunch wasn't a lie, exactly. It was a trial. And the trial is ending across every tool at once — not by coincidence, but because the economics finally caught up with the marketing. The builders who panic will read this as the end of indie software. It isn't. It's the end of careless indie software. The ones who survive won't be the ones with the best prompts or the loudest build-in-public threads. They'll be the ones who learned, before they were forced to, how to read the bill. Building was never free. You just hadn't gotten the invoice yet. It's here now. Open it. Are you seeing this in your own usage yet — has the billing shift changed what you're willing to start building? Curious where other people land on this.