GitHub stopped accepting new signups for its paid Copilot plans on Monday, pulled Anthropic's Opus models from the cheaper Pro tier, and offered refunds to anyone who wants out by May 20. In a blog post, VP of product Joe Binder blamed agentic workloads, saying the compute costs have outrun what Copilot's per-request pricing was designed to cover.
The freeze applies to Pro, Pro+, and Student subscriptions. Copilot Free stays open. Existing subscribers can still switch between paid tiers, which is the only remaining path in for anyone who hadn't already bought a seat.
The per-request model finally broke
GitHub has been billing Copilot by the request, not by tokens. One request is one interaction with the model, with a multiplier attached per model to reflect relative compute cost. Pro gets 300 requests a month for $10. Pro+ gets 1,500 for $39. That structure worked fine when a "request" meant completing a line of code or explaining a stack trace. It works less well when a single request kicks off a forty-minute agent run that chains tool calls, re-reads files, and burns through a 1M-token context window.
"Long-running, parallelized sessions now regularly consume far more resources than the original plan structure was built to support," Binder wrote. Which is the polite version of: we priced this before we understood what agents would actually do.
The changelog is blunt about what happens next. Session limits get tightened to "balance reliability and demand." Weekly token caps, introduced to cut off runaway agent loops, stay in place. Pro+ will offer more than 5x the limits of Pro, which is another way of saying Pro just got squeezed.
The Opus 4.7 math
Here's where it gets interesting. Opus 4.6 was priced at a 3x request multiplier. Opus 4.7, which Anthropic launched on April 16, runs at 7.5x under a promotional rate that expires April 30. GitHub hasn't said what the multiplier becomes on May 1. Given that the entire point of this overhaul is to stop losing money per request, the honeymoon number is almost certainly not the permanent number.
And the 7.5x headline understates the damage. The new model ships with a new tokenizer that produces more tokens for the same output: Anthropic's own documentation puts it at up to 1.35x, and Simon Willison measured roughly 1.46x on text in practice. Per-token pricing is unchanged from Opus 4.6, so depending on whose number you trust, the same prompt now costs 20 to 40 percent more. That's before the multiplier.
Opus 4.5 and 4.6 are being removed from Pro+ entirely. Pro loses all Opus access.
If you're on the $10 plan and you were using Opus for anything, that's over.
The refund offer is the tell
Companies don't usually volunteer to unwind a month of charges unless they're expecting a revolt. GitHub gave Pro and Pro+ subscribers until May 20 to cancel and get April refunded. The community forum reads about how you'd expect. One commenter flatly called it "totally unacceptable." Another pointed out that Opus 4.7 running at Copilot's medium thinking setting is more prone to hallucinations (Anthropic itself recommends higher effort levels for serious work), which makes the premium-model pitch feel thinner than it should.
The bigger complaint is the one Microsoft hasn't really answered. If I paid $39 a month for a specific tier of Pro+ service, and you've quietly cut what that subscription can do without consulting me, why is my subscription still worth what I paid for it? "Cancel for a refund" is an answer of sorts. It isn't a great one.
Where this is heading
Where's Your Ed At, citing leaked internal documents, reports that Microsoft is preparing to move Copilot to token-based billing. GitHub hasn't confirmed that publicly, and I couldn't get independent verification. But it's the only pricing model that makes the arithmetic work once agents are routinely running multi-step workflows that drain enormous context windows in a single "request." Per-request billing is a fiction that held together while Copilot was mostly autocomplete. It stopped holding together when Copilot became an agent.
GitHub isn't alone here. Anthropic tightened its own Claude usage limits in late March. Google did the same for Gemini CLI and Code Assist. OpenAI rebalanced its own plans earlier in April. The industry spent 2025 selling agentic AI as a generational leap and is now, quietly, starting to price in what that leap costs to keep the servers on.
Three things to watch. Whether GitHub publishes the post-promo Opus 4.7 multiplier before May 1, or lets subscribers discover it on their bills. Whether token-based billing surfaces officially at Microsoft Build next month. And how many Pro+ subscribers actually take the refund. If that number is large enough to notice, the next announcement is a price hike.
