I’ve been a GitHub Copilot subscriber since the early access days. It was one of those tools that quietly became load-bearing infrastructure in my workflow — always there, always fast, always finishing my sentences. I renewed without thinking about it.
Then one morning in late April I hit a wall mid-session. Not a rate limit on the free tier. Not a temporary outage. A hard weekly usage cap — on a paid Pro subscription.
What actually changed
On April 20, 2026, GitHub published a blog post titled “Changes to GitHub Copilot Individual plans”. The headline items:
- New sign-ups for Pro, Pro+, and Student plans were paused
- Weekly token-based usage limits were tightened for all individual plans
- Session limits were introduced on top of the weekly ones
- Opus models were removed from Pro plans entirely
The justification was reasonable enough on its face: agentic workflows have gotten expensive, long-running sessions are burning through infrastructure costs, a handful of heavy users were making the service unreliable for everyone else. Fine. These are real problems.
But here’s the part that stings: I didn’t sign up for a product with session limits and weekly token caps. Those weren’t in the deal when I subscribed. They were added retroactively, silently, to a plan I was already paying for — and only explained after the fact in a blog post most users wouldn’t see.
The kicker is in the footnotes on the current pricing page:
Response times may vary during periods of high usage. Requests may be subject to rate limiting.
That footnote is doing a lot of work. It’s covering for a fundamentally different product than what was sold.
The usage-based billing pivot
A week later, GitHub announced another shoe dropping: Copilot is moving to usage-based billing on June 1, 2026. Premium request units are being replaced by “GitHub AI Credits,” consumed based on token usage — input, output, and cached tokens — at published API rates per model.
Pro stays at $10/month. But that $10 now buys $10 of AI Credits, not an effectively unlimited flat-rate subscription. If you’re using Claude Sonnet or GPT-5 for a long agentic session, that credit disappears fast.
The plan pricing didn’t change. The product did. Significantly.
I understand why they’re doing it. A multi-hour autonomous coding session obviously costs more than a quick chat question, and charging the same flat rate for both was never sustainable once agentic usage took off. GitHub absorbed that cost for a while, and now they can’t.
But there’s a difference between adjusting pricing for new customers and retroactively changing what existing subscribers get. Annual plan holders are being migrated off their plans at expiry with no grandfathering of the model they signed up for. Monthly subscribers are being automatically migrated on June 1 whether they opt in or not.
Why the limits felt arbitrary
The specific frustration with the new limits isn’t just that they exist — it’s that they’re opaque and inconsistent.
There are two separate constraint systems running in parallel: session limits (per-request windows, for infrastructure protection) and weekly limits (total token budget, for cost control). You can have premium requests remaining and still hit a usage limit. The two systems are completely independent. You can be blocked from using the model you want while the dashboard shows you still have requests left.
GitHub now displays a warning in VS Code when you approach 75% of your weekly limit. That’s an improvement. But being notified that you’re running low on a budget you didn’t know existed, for a subscription you signed up for before that budget was introduced, feels like the product getting worse and calling it a transparency feature.
What I moved to
I switched to Claude Code. It’s a terminal-first tool built on top of Anthropic’s Claude models, designed specifically for agentic development sessions — the kind of long, multi-step, repo-spanning work that was quietly breaking Copilot’s pricing model.
The billing is transparent upfront. You pay for what you use, at rates that are published and consistent. There’s no dual-limit system to reason about, no premium request concept that’s separate from a token budget, no footnotes explaining that what you’re paying for might be subject to rate limiting.
It also happens to be very good at the agentic stuff — not just completing inline suggestions but actually reasoning through multi-file changes, writing tests, doing refactors across a codebase. That’s what I was pushing Copilot into doing, and it’s what Claude Code is explicitly designed for.
The broader point
I’m not writing this to pile on GitHub. Copilot is genuinely impressive engineering, and the team is dealing with a real cost problem created by how fast agentic usage has grown. Moving to usage-based billing is probably the right call long-term.
But when a product changes its core terms mid-subscription — what you get, what limits apply, which models you can access — without proactively telling subscribers before it happens, trust erodes. I found out about my weekly token limit by hitting it, not by reading a changelog.
The developer tools space is competitive enough now that trust matters. If you tell me I’m on a premium plan, I should be able to rely on what that means staying stable for the period I paid for. When it doesn’t, I start looking at the alternatives — and right now, the alternatives are pretty good.