AINEWS
GitHub Copilot's New Token Math Hits Code Review Hardest
GitHub's June 1 shift to per token AI Credits makes Copilot Code Review the new bill's loudest line item, and admin budget controls are the governance lever teams need to pull.
GitHub's June 1 shift to per token AI Credits makes Copilot Code Review the new bill's loudest line item, and admin budget controls are the governance lever teams need to pull.
GitHub replaced Copilot's flat premium-request billing with per-token AI Credits on June 1, 2026, and code review is the workload where the new math is most visible. A single review pass pulls long files, full repository context, multi-iteration comments, and the cached tokens that GitHub now bills on top of input and output, all of which used to roll up into a single unmetered request under the old model.
The change was announced on April 27, 2026 by GitHub product lead Mario Rodriguez, framed as a swap from premium-request units to a monthly GitHub AI Credits allotment per plan tier. Plan prices did not move: Pro stays at $10, Pro+ at $39, Business at $19, and Enterprise at $39. What moved is the meter. Usage is now calculated on input, output, and cached tokens, billed at the published API rate of each model. The base Copilot features most developers use in the editor are insulated: code completions and Next Edit suggestions are still included and do not draw from the credit pool. Everything more ambitious, including Copilot Code Review, runs against AI Credits — and code review also consumes GitHub Actions minutes at the same per-minute rates as other workflow jobs, per GitHub's announcement.
That is the line item that explains the spike teams are now seeing six days into the new billing. Code review pushes unusually long context into the model. A typical review pass reads the changed file plus surrounding files, repository conventions, prior review history, and a long prompt asking the model to flag bugs, suggest fixes, and write tests. Each of those inputs costs tokens. The model's output, often several hundred lines of inline review comments, costs more. Under the old PRU model, the whole review counted as a single request. Under the new model, the same review counts in tens of thousands of tokens, with cached tokens now billable on top of input and output where they were previously invisible to the user.
The reaction in the first week was loud enough that TechCrunch reported on May 30, 2026 that developers had taken to GitHub's own community channels to call the change confusing, and the publication's own headline framed the rollout as "a joke." The official venue is GitHub Community Discussion 192948, where users and admins have been posting billing questions, edge cases, and worked examples since the transition. Artificial Intelligence News corroborated the price-hike reports with its own tally of users seeing higher bills.
The sharpest secondary effect is the removal of the automatic fallback to a cheaper model. Before June 1, a long review that would have exhausted a premium-tier model's per-request budget could fall back to a smaller, cheaper model and still complete. GitHub has removed that fallback and replaced it with admin-set budget controls. In practice, a runaway review session no longer degrades to a cheaper model to save money. It either runs to completion on the selected model, drawing whatever AI Credits it needs, or it fails when the budget cap is hit. For teams that have not yet set those caps, the first expensive run of the month can come as a surprise.
The constructive answer is governance, and GitHub is shipping the tools for it in the same release. Admins now have a preview bill experience in the Billing Overview page that shows projected AI Credit consumption per user, per repository, and per feature. The preview was available from early May 2026, which gave teams a roughly four-week window to model their actual code review volume before the live transition. The same release adds admin-set monthly budgets at the organization and user level, and the option to disable specific Copilot features, including code review, for seat groups that should not be running them.
For a team auditing itself before the next billing cycle, the order of operations is straightforward. First, look at the model tier being used for code review. The largest token draw is usually a flagship reasoning model; routing code review to a smaller, cheaper model on the GitHub rate card can drop per-review cost by an order of magnitude. Second, set a per-user monthly AI Credit cap. Runaway sessions now stop at the cap instead of consuming the team's entire allotment. Third, audit which repositories have code review enabled; if only a few services in the org need it, scoping the feature to those repositories prevents casual use from the rest of the engineering org. Fourth, watch the Billing Overview daily for the first month to see which users and which repositories are driving the bill, then adjust the caps and the model routing accordingly.
The sense that some teams have of being charged twice comes from the fact that the new model surfaces costs the old model hid. The old premium-request allowance is gone, replaced by the per-token AI Credit pool. Cached tokens, long input prompts, and full multi-file review passes are now visible line items where they used to be invisible. Code review, by design a long-context, high-output workload, is the feature where the new math is most visible. Teams that model their actual review volume using the preview bill, set monthly caps, and route heavy repositories to a smaller model will land close to their old cost. Teams that do none of that will see the credit pool drain fast, and the developer reaction of the past week is the early signal of that gap.