GitHub announced this week that it is pausing new sign-ups for Copilot Pro, Pro+ and Student plans, tightening usage limits, and pulling Opus model access from Pro tier subscribers. The stated reason is straightforward: agentic workflows have fundamentally changed how much compute Copilot actually consumes, and the original plan structure was not built for it.
That last part deserves a moment's attention, because GitHub is saying something fairly candid here. The pricing model it sold developers was designed around a world where Copilot completed lines of code and answered the occasional question. Agents are a different proposition entirely. Long-running, parallelised sessions that spin up subagents, manage tasks, and work through extended problem sequences consume orders of magnitude more tokens than a conventional autocomplete interaction. GitHub acknowledges in its announcement that it is now common for a handful of requests to exceed the entire monthly plan cost in compute terms. That is not a pricing edge case. That is a structural problem.
The immediate practical impact for existing individual subscribers is a tighter set of weekly token limits, which are now displayed in VS Code and the Copilot CLI to give some advance warning before they hit. Pro users who need more headroom are pointed toward Pro+, which offers more than five times the limits of the lower tier. Opus models, meanwhile, are gone from Pro entirely, with only Pro+ retaining access to Opus 4.7 (and even then, Opus 4.5 and 4.6 are being removed from that tier too).
GitHub is offering refunds for April to anyone who cancels between now and 20 May, which at least suggests an awareness that springing this on existing subscribers mid-month is not ideal.
The wider point is one that applies across the AI tooling market. Flat-rate subscriptions made sense when AI assistance was essentially stateless and cheap. Agentic workflows are neither. Every vendor selling a monthly flat rate for AI development tools is making a bet on average usage staying within comfortable bounds. As agents become the default mode of working rather than a niche power-user feature, that bet gets harder to sustain. GitHub has hit the limit of it first. Others will follow.
