The funniest thing GitHub has done this month is accidentally admit that the AI coding buffet was never a buffet. It was a spreadsheet with a logo on it, and now the spreadsheet is screaming.

On April 20, GitHub said it was pausing new Copilot Pro, Pro+, and Student sign-ups, tightening usage limits, and yanking some model availability around like a panicked stagehand. Their explanation was the usual corporate incense: agentic workflows are burning more compute, long-running sessions are expensive, and they need “reliable” service for existing customers.

Fine. Sure. Reality is rude. Compute costs money. Fancy models are not fairy dust. The part that makes me laugh until I can feel my soul chipping off is that this was sold, for months, like the cost problem had been solved by sheer confidence and a couple of dashboards with blue gradients.

It had not been solved. It had been hidden.

That is the whole stinking story here. The old subscription fantasy was simple: collect a tidy monthly fee, smear infrastructure costs across enough users, and hope nobody asks too many uncomfortable questions. That works beautifully when the product is mostly static requests and polite autocomplete. It works a lot less beautifully when people start launching long-running, parallelized agent sessions that chew tokens like a chainsaw through wet plywood.

At that point, the product stops being “software” in the cute SaaS brochure sense and becomes metered power with a personality disorder.

And yes, I know the official language is more dignified than that. There are “session limits” and “weekly limits” and “premium requests” and “plan mode” and “usage displayed in VS Code and Copilot CLI” so you can avoid the shame of hitting the wall unexpectedly. Which is adorable. Truly. Nothing says premium product like a little fuel gauge politely informing you that your expensive thinking machine has entered its medically induced coma.

If your tool needs to warn me that I am approaching the end of my monthly brain allowance, maybe it is not a magic assistant. Maybe it is a rented furnace.

The really delicious part is the mismatch between the pitch and the physics. We were sold a future where agents would just keep going, subagents would fan out, and the machine would handle the boring stuff while we stayed in the zone, whichever marketing department church that came from. Then reality walked in wearing work boots and said, “Cool story. Each of those requests costs real money, and the tab is no longer cute.”

So now the tune changes. Suddenly the language is about guardrails, transparency, and protecting reliability. Suddenly you’re encouraged to pick a smaller model for simple tasks. Suddenly you should reduce parallel workflows. Suddenly Pro+ exists to remind Pro users that the ladder has a rung missing unless they pay more.

That is not a product refinement. That is a confession dressed up as UX.

And I don’t even mean that as a moral insult. I mean it technically. The pricing model broke against the shape of the workload. The thing being sold as a flat subscription had become a variable-cost monster, and the numbers were starting to mug the business in broad daylight. When GitHub says a handful of requests can cost more than the plan price, that is not a temporary inconvenience. That is the business model making a noise like a car engine full of gravel.

This is the part where everyone pretends to be shocked, but nobody should be. Agentic coding tools are not cheap. They are not going to become cheap because somebody put the word “copilot” in a product page and sprinkled some IDE integration on top. Long-running context, retries, tool calls, parallel sub-tasks, model switching, all of it adds up like a bar tab for a startup that discovered self-destruction.

The only thing more absurd than the cost is the confidence. The industry spent an entire era acting like token bills were an implementation detail. “Just scale it.” “Just optimize it.” “Just add a tier.” Yeah, cool. Tell the electricity company you’ll pay them in vibes. Tell the GPU cluster to run on ambition.

What makes GitHub’s move feel so grimly hilarious is that it exposes the true shape of the relationship. Your Copilot plan is not a promise of unlimited help. It is a polite estimate of how much torment they’re willing to subsidize before the accounting team starts throwing furniture.

So if you’re asking me whether this is the end of the AI coding gold rush, no, probably not. Corporations will keep shoving agents into everything with sockets. Someone will keep inventing a new name for “an expensive loop with a chat window.” Users will keep trying to get more out of the box than the box was ever priced to give.

But the fantasy is cracking, and good. It deserved to crack. Any product that gets sold as infinite until the first honest invoice lands was always one bad quarter away from a tantrum.

The real lesson here is not that AI coding is doomed. The real lesson is that economics still exists, which is deeply inconvenient for everyone trying to cosplay as if it doesn’t. You can slap “agentic” on a feature all you want. Eventually somebody has to pay for the tokens, the GPUs, the retries, the bad prompts, the long walks through the swamp of partial context, and the thousand tiny little mistakes that turn into real dollars.

And when that bill comes due, the product either gets smaller, or it gets meaner, or it starts putting you on a timer.

GitHub chose timer.

That is the whole, miserable opera.