GitHub announced on April 28 that Copilot will shift to usage-based billing on June 1, replacing the existing "premium requests" allotment with "AI Credits" pegged to the per-token API rates of whatever model you're calling. Subscription dollar amounts don't change โ Pro+ stays $39/month and includes $39 in credits, Business stays $19, and Enterprise stays $39 โ but those credits convert to tokens at published API rates that vary roughly 6ร between cheap and high-end models. Monthly subscribers auto-migrate; annual Pro and Pro+ users keep their current plan until expiry, at which point they fall to Copilot Free. The trigger, per a leaked internal document Ed Zitron reported on last week, is that week-over-week Copilot costs had nearly doubled since January.
The mechanics matter. Code completion and Next Edit remain free of credit consumption โ those are the lightweight, cached-heavy operations Microsoft can still subsidize. Everything else โ chat, agent sessions, code reviews โ consumes either AI Credits or, in the case of code reviews, GitHub Actions minutes (a separate billing surface). The token math: at $30 per million output tokens for a top-tier model, $39 in monthly credits buys roughly 1.3 million output tokens, which a multi-step autonomous coding session can chew through in minutes. For organizations, GitHub is introducing pooled credits shared across the team rather than stranded in individual seats, and admins get new budget controls at enterprise, cost-center, and user levels. Existing Business and Enterprise customers get promotional credits for June, July, and August.
Copilot was the last major agentic coding assistant on flat-rate pricing. Cursor, Claude Code, Cody, Aider โ they all bill by token or session because their costs scale with model usage, and hiding that scaling under a $20-a-month subscription only works while users are mostly typing. Once "always-on multi-agent workflows" became the default, GitHub couldn't eat the difference. Ars Technica's coverage names "Openclaw" โ a clear stand-in for autonomous coding agents โ as the trigger; the structural read is that Microsoft has decided not to subsidize agentic users at the expense of margin. Pooled team credits are a genuinely good design choice, but the broader move tells you the "AI tool that saves developer time" pitch is now hard to distinguish from a metered cloud service.
For builders, this changes the comparison shopping. Copilot, Cursor, Claude Code, and the rest are now all metered โ you can run apples-to-apples cost-per-task comparisons instead of pretending the flat fee covered your usage. If you run autonomous agents heavily, the included $39 in credits will go fast and you will be on overage at API rates; if you mostly use code completion and Next Edit, your bill won't move. The "preview bill" tool GitHub is shipping before June 1 is the thing to actually look at โ it will tell you whether you are a $20-a-month user or a $200-a-month user under the new model. And if you are on an annual plan: you have runway until expiry, but the off-ramp lands you on Copilot Free, not a discount migration.