Copilot's Ten-Day Cliff

On April 20, 2026, GitHub published a short changelog entry titled "Changes to GitHub Copilot plans for individuals." The post announced that new signups for Pro, Pro+, and Student plans were paused. It announced that Claude Opus was being removed from the Pro tier entirely, and that Opus 4.5 and 4.6 were being retired from Pro+. It announced a one-month window, from April 20 to May 20, during which paying customers could cancel their subscription and request a refund for April^[1].

Read in isolation, the post reads like a plan refresh. Read in the context of the ten days preceding it, it reads like a controlled evacuation.

The timeline

Sometime in March, GitHub discovered what it called a "token-counting bug." Newer models, including Claude Opus 4.6 and GPT-5.3, had been consuming more compute than the billing system recorded. When engineers corrected the accounting, a different problem became visible: the advanced models cost far more to serve than the monthly subscription price assumed. In the words of industry analyst Roman Kir of StratoAtlas, "the unit of sale had been decoupled from the unit of actual cost"^[2].

April 10. GitHub published a changelog titled "Enforcing new limits and retiring Opus 4.6 Fast from Copilot Pro+." The post introduced two new classes of rate limits, one for "service reliability" and one for "model capacity." It cited "high concurrency and intense usage" that "places significant strain on our shared infrastructure and operating resources." It advised customers to "distribute requests evenly over time rather than in concentrated waves." The retirement of Opus 4.6 Fast was framed as an ordinary product decision^[3].

April 15. The Register published "Customers revolt as GitHub Copilot 'fixes' rate limits." The piece collected user reports of rate-limit lockouts ranging from 44 hours to 181 hours, several dozen of them filed within two days of the April 10 changes. One user, John Clary, told the outlet he had been "frequently receiving short rate limits interrupting the request with no listed countdown." A community commenter, Delsin-Yu, put it plainly: "This feels less like a bug and more like a business decision"^[2].

April 16. GitHub made Claude Opus 4.7 generally available inside Copilot at a 7.5x premium request multiplier, with a promotional note that the multiplier would stay in place only until April 30. What the post-promotion multiplier would be, the changelog did not say^[4].

April 20. Signups for Pro, Pro+, and Student plans were paused. Opus was removed from Pro entirely. Opus 4.5 and 4.6 were slated for removal from Pro+. A refund window opened^[1].

That is the full arc. Ten days. No press release. No pricing page update. Each changelog treated in isolation as a routine operational note.

What the words mean

Language in GitHub's posts is consistent enough that it is worth translating.

"To ensure service reliability and a sustainable Copilot experience" is how the April 20 post opens. The phrase does not describe a product improvement. It describes a supply constraint. A service that is reliable at $10 per month for a customer using Opus 4.7 is a service that was priced below cost. A service that is sustainable is one that is not.

"Pausing new signups" is not a plan change. It is a capacity statement. The company has determined that the cost of serving an additional Pro customer exceeds what a Pro customer pays. The rational response, absent a price increase the market would reject, is to stop accepting new customers at the losing price. The refund window accomplishes the same thing for the customers already on the rolls: rather than raise prices mid-cycle and expose itself to consumer-protection pushback, GitHub is offering a door marked "exit" and inviting its heaviest users through it.

The model-tier cascade

The model-availability changes tell the same story in a second language. Opus 4.6 was removed from Pro+ Fast on April 10. Opus was removed from Pro entirely on April 20. Opus 4.5 and 4.6 are being removed from Pro+. Opus 4.7 carries a 7.5x request multiplier through the end of the month.

Each of these moves does the same thing: it takes the most expensive-to-serve model class and pushes it toward the tier least accessible to the median Copilot customer. The Pro user who signed up for "Copilot with Claude" is being quietly reassigned to Copilot-without-Claude. The Pro+ user who upgraded to keep Opus is being told the Opus inventory on that tier is being restocked with less capable versions. The Opus 4.7 user on any tier is being warned, without being told, that the premium multiplier on the best model is about to go up.

This is a capacity ration, managed at the model-availability layer because managing it at the price layer would require a public conversation.

Who pays

The cost does not disappear. It is being shifted.

It is being shifted from Microsoft, which contractually guaranteed per-seat Copilot pricing, onto Anthropic, whose Opus inventory is now being allocated to the tier that yields the highest dollar-per-token, and onto Copilot customers, who are either being offloaded, upsold, or rationed.

The refund window is the hedge. A one-month refund offer, published in a product changelog, is not a customer-service gesture. It is a filing cabinet: a paper trail that says customers were given the chance to leave before conditions changed. If a state attorney general or a class-action firm looks at the Copilot quota pattern six months from now, that paragraph is the first exhibit in GitHub's defense.

The thing that isn't being said

Copilot is not a standalone AI company. It is a distribution channel for models built by Anthropic, OpenAI, and others, wrapped in an editor integration. When Copilot tightens, the tightening reflects something upstream. The best public reading of the April 10 through April 20 sequence is that Microsoft, a major enterprise customer of Anthropic's frontier models, is being asked to take less than it was taking, and is passing that constraint through to its own paying developers.

That reading matches the rest of the visible AI economy. It matches Vercel's enterprise perimeter being compromised through a third-party AI startup. It matches the NSA running Anthropic's Mythos against a Pentagon blacklist^[5]. It matches a sector in which the scarcest resource is not talent or ideas but the raw compute to serve the models that every vendor has already sold twice.

Copilot's ten-day cliff is not a billing story. It is the moment a distribution channel started quietly rejecting load that the system upstream can no longer carry. The refund window is the courtesy.