Last week, GitHub quietly stopped accepting new Copilot subscriptions.
This is not system maintenance or a temporary glitch. GitHub's official explanation is straightforward: agentic workflows—long-running, parallel AI tasks—consume far more computing resources than the original subscription plans were designed to support.
Affected plans include Pro, Pro+, and Student, all individual user plans that have now suspended new user registrations.
What exactly changed
According to GitHub's official announcement published on April 20 (updated April 21), the changes are as follows:
Suspended new subscriptions:
- Copilot Pro
- Copilot Pro+
- Copilot Student
Feature changes:
- The Opus 4.7 model has been removed from the Pro plan, now available only in Pro+
- Two new quota mechanisms introduced: per-session limit + 7-day rolling token limit
- Pro+ quota is more than 5 times that of Pro
- Existing subscribers can cancel and request a refund before May 20
GitHub's VP of Product stated:
Long-running, parallelized sessions now regularly consume far more resources than the original plan structure was built to support.
In short: the pricing model designed for code completion is now being used for Agent tasks, which operate on an entirely different scale of consumption. The gap between pricing and actual costs is growing and has become unsustainable.
How this problem emerged
The early business model for AI coding tools was simple: charge each user $0–20 per month for code completion, code explanation, and Q&A features. The computing costs of this model were relatively predictable—each user request consumed a modest number of tokens, response times were short, and concurrency was limited.
But agentic workflows completely shattered this assumption. A typical Agent task might:
- Run continuously for minutes to tens of minutes
- Execute multiple sub-tasks in parallel (refactoring a module, generating tests simultaneously, modifying multiple files)
- Consume large context windows at every step
After tools like Claude Code and Cursor successively launched background Agent features, users began offloading heavy tasks—code refactoring, cross-file modifications, test suite generation—entirely to AI. The computing power that a $20 monthly subscription used to buy could now be exhausted in a single afternoon's Agent session.
For the platform, the result is simple: actual computing costs far exceed the expectations set when pricing was originally designed.
Why Opus was removed from the Pro plan
This detail is the most telling.
Claude Opus 4.7 is one of the most powerful programming reasoning models available, but also one of the most expensive to run. In intensive agentic tasks, Opus's token consumption is an order of magnitude higher than Sonnet- or Haiku-level models.
Removing Opus from the $10/month Pro plan sends a clear message: at current usage depths, $10 simply cannot cover the actual cost of running Opus.
This is not just GitHub's problem—it's a wall that every AI tool offering monthly subscriptions with high-performance models will hit. Monthly pricing was designed based on tool usage frequency, but Agent tasks continuously consume computing resources. The two models are fundamentally incompatible.
The bigger problem
This adjustment exposes a fundamental contradiction in the AI coding tool industry:
- Users want subscription fees as low as possible
- AI companies have been marketing stronger Agent capabilities as a selling point
- But stronger Agent capabilities = higher computing costs per use
These three goals cannot be satisfied simultaneously. Monthly pricing must either cap computing costs or compromise on capabilities.
GitHub chose to protect computing costs first—even at the cost of suspending new subscriptions. This itself shows that monthly subscriptions with unlimited agentic tasks simply don't add up mathematically.
The path forward offers a few options: raise monthly fees, charge by usage, bill agentic features separately, or downgrade models for different task intensities. GitHub's current move is a temporary patch—pushing heavy users toward Pro+ and removing the most expensive model from the low-cost plan.
But this is a band-aid, not a cure. As Agent tasks become more common in daily development, this pricing logic needs a fundamental redesign.
Every company building AI coding tools will eventually face this problem. GitHub is just the first to be forced by reality to show its hand.
What's worth watching: whether a new pricing system specifically for agentic workflows will emerge—charging by task duration, by token consumption, or by task complexity. Each direction has its own issues, but all are at least logically sounder than the current unlimited monthly usage model.
Source: Changes to GitHub Copilot Individual Plans (GitHub Blog, CocoLoop, April 20, 2026)