The Pricing Emails Are Landing (Part 1 of 3)

The Email I Was Not Expecting To Read Yet

Three weeks ago I wrote that building data centres in space smelled like a bubble. I expected the bubble to deflate from the top first. Orbital satellites, multi-billion-dollar capex, the showy stuff. That was the obvious place to look.

It deflated from the bottom first.

On the morning of April 28th, before my coffee was finished, I had two emails open on the screen. One was a forwarded link to a PCWorld piece about GitHub Copilot's pricing change. The other was a thread from a developer in Tokyo who had been hitting Claude Code limits two days into his weekly quota. Neither of these was a satellite story. Both of them were a bubble story.

This is the first of three posts. In this one I am going to lay out what just happened in the pricing layer, in the order it happened. The next post will be about why this was always going to happen. The post after that will be about what I think small businesses should actually do.

GitHub Copilot Blinks First

GitHub announced that on June 1, 2026, Copilot's flat-rate plans are moving to credit-based billing. The headline price does not change. Copilot Pro stays at $10 a month and Copilot Pro+ stays at $39. What changes is what those numbers buy you.

Today, Copilot Pro gives you a fixed allowance of "premium request units." From June 1, those go away. In their place GitHub is giving you a pool of AI Credits valued at the dollar amount of your plan. So $10 of credits on Pro, $39 on Pro+. Code completion, the thing most users actually use Copilot for, stays free. The agentic features, including Copilot's code review agent, will now consume credits.

GitHub's own framing is the part worth reading carefully. The company said the current model is no longer sustainable as the more capable, more agentic features have grown more expensive to run. That is a polite way of saying we mispriced this and we want our margins back.

It is easy to dismiss this as a Microsoft problem. It is not. It is the first of a category. Every flat-rate AI plan on the market today was built when the dominant use case was lightweight assistance. Tab completion. A short reply. A few hundred tokens at a time. Now the same plans are being asked to subsidise multi-step agents that run for minutes, generate thousands of tokens per turn, and call out to other tools while they are at it. The arithmetic does not survive the change in workload.

So GitHub blinked. They did the most polite version of repricing they could engineer. The headline number stays. The unit changes underneath it. The user who only ever taps Tab will not notice. The user who runs Copilot's review agent on every pull request will notice on day one.

It Is Not Just GitHub

Anthropic has been doing a less polite version of the same thing for over a month. The pattern is worth laying out in dates, because the dates are the story.

March 26, 2026. Anthropic confirmed it had been adjusting Claude's 5-hour session limits during weekday peak hours, roughly 5am to 11am Pacific time. The company said about 7% of Pro users were hitting those caps earlier than they had before. The change had been in production for a while before it was acknowledged. Users had been noticing.
April 4, 2026. Anthropic blocked third-party agentic harnesses from using Pro and Max subscriptions. If you had wired Claude into your own tooling, your own runner, your own scaffolding around the chat API, that route was closed. The official channels remained open. The unofficial ones, the ones that let people get more value out of a Pro subscription than the pricing model assumed, did not.
April 21, 2026. Claude Code briefly disappeared from the Pro plan in the comparison table on Anthropic's pricing page. Only the Max 5x ($100) and Max 20x ($200) tiers showed it. Hacker News noticed within hours. Four hundred and some comments piled up in less than a day. Anthropic reversed the change inside 24 hours. Their Head of Growth said it had been a small test on roughly 2% of new prosumer signups. Whether that is the whole story or not is something each reader can decide. The signal it sent was not small.
April 23, 2026. Anthropic published a postmortem after Max users had burned through weekly quotas in a day or two. A cache bug had been inflating token costs by 10 to 20 times for some customers. Anthropic reset all subscriber limits and apologised. Importantly, they also acknowledged that current flat-rate plans, in their own words, were not built for the kind of usage Claude Code is now seeing.

Behind the four dates is some context that ties them together. Anthropic announced a 25 billion dollar agreement with Amazon in April for 5 gigawatts of compute. That compute is not online yet. Then in mid-April, Anthropic released Opus 4.7, which is excellent and which, on Claude Code, runs sessions roughly three times longer than Opus 4.6 did. Better model, more thinking, more tokens, same flat-rate plan. The product itself was making the bill bigger.

If you only knew about one of the four April events, you might call it a hiccup. Stitched together, the pattern looks different. Peak-hour throttling. There was the third-party harness blocks, and then a quiet pricing-page experiment that backfired. A postmortem with the words "weren't built for this" in it. I think that is the shape of a pricing model under pressure.

A separate writeup of the April 21 incident by Italian developer Pasquale Pillitteri is worth a read for the hour-by-hour reaction on Hacker News. The reason it spread so fast is that the Pro tier is where most of Anthropic's individual paying users live. If Claude Code stops being part of that tier, even hypothetically, it changes a lot of household budgets. People felt that immediately.

What This Means In My Inbox

I run a small AI company in Tokyo. We use a lot of AI for actual work, every day, to build and to run for our own product features at Kafkai. I am not an outsider commenting on someone else's bill. I am the one paying it.

So when both Anthropic and GitHub move on pricing in the same six-week window, my reaction is not panic. It is recognition. The shape of this is familiar. We even changed our pricing for Kafkai last year too because the value that we're giving is too expensive for what we're asking for.

The flat-rate AI subscription, in the form we all have been used it for the last two years, was a user-acquisition product. It was priced to bring people in, not to balance the books on a per-user basis. That is fine. Many products start that way. The problem is the moment when somebody internally has to make the books balance.

That moment seems to have arrived.

I want to be careful here. None of what I have laid out is the end of Anthropic, or of GitHub Copilot, or of any of these tools. The companies are not in trouble. Anthropic raised more capital in February than most countries' tech industries are worth. The tools themselves are getting better. The models keep improving. What is in trouble is the specific commercial structure of an unlimited-feeling subscription priced at a flat number that does not reflect what frontier inference actually costs in 2026.

That structure is being repriced in front of us. GitHub is doing it on a scheduled date. Anthropic is doing it through a series of small, partly walked-back, partly normalised adjustments. Both companies will end up at the same place. Usage-based billing for the agentic features, free or near-free for the lightweight ones, and a flat fee on top that buys you a credit pool, not a buffet.

The First Crack

The headline at the top of this post is the one I want readers to take away. The bubble I called out three weeks ago is starting to deflate. It is not deflating where the cameras are pointed. It is not satellites coming back down. It is a series of pricing emails landing in the inboxes of developers and small businesses who built workflows around plans that were not built for those workflows.

This is the first crack, not the whole earthquake. The stock prices have not moved. The capex announcements keep coming. The orbital filings keep getting submitted. None of that is about to disappear. But the place where the bubble touches your monthly statement is the place it deflates first.

In the next post in this series I want to do the part I did not do here. I want to walk through why this was always going to happen. The numbers behind it are not subtle. OpenAI's inference spend, the total US AI capital expenditure projected for 2026 and 2027, the amount American consumers actually pay for AI per year. When you put those side by side, the flat-rate subscription becomes the obvious place for the pressure to show up.

Until then, I would suggest reading your invoices. The next surprise in your inbox is not going to be from GitHub.