Copilot pauses sign-ups as agent workloads break plan math

GitHub’s Copilot change is a rare public look at the economics behind coding agents. In an April 20, 2026 blog post updated on April 21, GitHub said it is pausing new sign-ups for Copilot Pro, Copilot Pro+, and Student plans, tightening usage limits, and changing which models individual subscribers can use.

The reason is not a small pricing tweak. GitHub says agentic workflows have changed Copilot’s compute profile: long-running and parallelized sessions now consume far more resources than the original individual plan structure assumed. That matters because modern coding agents can run multi-step repository work, spawn parallel tasks, and keep context alive long enough for a few requests to carry very large token costs.

The concrete changes are direct. New individual sign-ups are paused so GitHub can protect service quality for existing customers. Usage limits are being tightened, with Pro+ offering more than 5X the limits of Pro. Opus models are no longer available in Pro plans; Opus 4.7 remains in Pro+, while Opus 4.5 and Opus 4.6 are being removed from Pro+ according to GitHub’s changelog.

The important detail for developers is that GitHub now describes two separate limit systems. Session limits protect the service during peak usage. Weekly limits cap total token consumption across a rolling 7-day period, especially for long-trajectory and parallel requests. These limits are separate from premium request entitlements, so a user can still have premium requests left while losing access to model choice after hitting a token-based cap.

GitHub is adding limit visibility in VS Code and Copilot CLI when users approach a cap. It also points users toward smaller-multiplier models, plan mode, fewer parallel workflows, or a Pro+ upgrade. For users who decide the change breaks their workflow, GitHub says Pro and Pro+ subscribers can cancel and request a refund for remaining time before May 20.

The broader signal is that coding-agent subscriptions are moving from simple request accounting toward compute-risk management. For teams building around Copilot, the operational question is no longer just which model is best. It is whether a workflow depends on parallel agents, long context, or high-multiplier models enough that limits become part of engineering planning.