DEV Community

Cover image for GitHub Copilot Credit Cliff: What the Reset Changed
Rebeca
Rebeca

Posted on Originally published at olumia.dev

GitHub Copilot Credit Cliff: What the Reset Changed

On June 1, 2026, GitHub Copilot rewired how every enterprise pays for it. Premium request units are gone. In their place is a token-metered system called GitHub AI Credits, where one credit equals one cent and every interaction draws down a monthly pool based on the input, output, and cached tokens each model consumes. Seat prices did not move, Business is still 19 dollars per user per month and Enterprise still 39, but the way usage is billed changed completely. The transition shipped with a promotional allowance, and that allowance ran through September 1, 2026.

The promotional window has closed

To smooth the transition, GitHub handed out enhanced credit allowances through September 1, 2026. During that window, Business seats included 3,000 credits per user each month and Enterprise seats included 7,000. On September 2 those allowances snapped back to their standard levels: 1,900 credits for Business and 3,900 for Enterprise, which is where they sit today. No contract changed. No developer behaviour changed. But the usage covered by the seat you already pay for fell by roughly 37 percent on Business and 44 percent on Enterprise, overnight.

That was the trap, and it is worth naming plainly because this kind of increase does not announce itself. Teams spent the summer letting developers use Copilot freely, calibrated their sense of normal against the promotional pool, and then received a materially larger invoice for the exact same work. Because overages bill at published rates once the pool is exhausted, the gap never showed up as a price increase. It showed up as consumption the team was always doing, now landing above the line.

What the cliff actually cost

Consider an Enterprise organization with 500 developers whose typical usage ran near the promotional ceiling of 7,000 credits per user. Before September that sat inside the included pool, so the marginal cost was zero. Now only 3,900 credits per user are included. The remaining 3,100 credits become overage: at one cent each, that is 31 dollars per user per month, or 15,500 dollars a month across the team, about 186,000 dollars a year in new spend that nobody approved, because nobody bought anything. Your real number depends on your actual token mix; treat this as the shape of the risk, not a quote.

The mechanics reward a particular kind of discipline. Credits are pooled at the billing entity rather than stranded per seat, so one developer's light month can offset another's heavy sprint, but only if someone is watching the pool as a whole. Unused credits do not carry over: the pool resets to the full monthly amount at 00:00:00 UTC on the first day of each calendar month, and whatever is left is forfeited. Code completions and Next Edit suggestions remain free and never touch credits, so the drawdown is driven almost entirely by chat, agents, and premium-model calls. The heavier the model and the longer the context, the faster the pool empties.

How to tell whether the cliff landed on you

The first full billing cycle under the standard allowances has now closed, which makes the impact measurable rather than hypothetical. Three checks matter most.

  • Compare September against your July or August run-rate. Pull credit consumption per user and per team for the first full cycle on the standard allowances and set it beside a promotional-window month. If consumption held steady and the invoice moved, the cliff is the entire explanation. This is the same predictive discipline covered in our CFO's guide to forecasting AI coding spend.
  • Check whether paid overage is even switched on. Additional usage beyond the included pool is enabled by default, so for most organizations the meter simply kept running past the reset rather than stopping. What happens when the pool empties, and how to change it, is covered in what happens when you hit the Copilot credit limit and in how GitHub Copilot spending limits actually work.
  • Watch the token mix, because that is the real lever. Two teams with identical seat counts can post very different bills depending on which models they lean on and how long their contexts run. A single team quietly moving to a premium model, or an agent workflow looping on large files, is exactly the kind of anomaly that compounds silently until the pool tips over.

The broader shift this signals

Copilot was not an outlier. Cursor moved to usage-based, API-pegged billing and then began retiring its flat Auto rate in August 2026, and the industry has converged on the same model: seats get you in the door, tokens determine the bill. That is a structural change in how AI coding tools are priced, and it quietly moves the cost-control burden onto the buyer. Fixed per-seat pricing let finance forecast with a spreadsheet. Token-metered pricing means the bill is now a function of developer behaviour nobody can see in a contract, and it only reconciles thirty days after the fact, when the invoice arrives.

That is the gap Olumia was built to close. We connect to the AI coding assistants your teams use, such as Copilot, Cursor, Claude, OpenAI and more, through read-only APIs and forecast spend before it lands, so a change like the September credit reset is something you model in advance instead of explaining afterwards. The method itself is vendor-agnostic and written up in how to track AI code assistant spend across every vendor.

The takeaway

The promotional window closed on September 1. Included Copilot credits fell by up to 44 percent, seat prices did not move, and the difference now lands as overage on work teams were already doing. Pull the first full cycle under the standard allowances, compare it against a summer baseline, and decide the overage policy deliberately rather than discovering it on the next invoice. The question was never whether the reset was coming. It is whether it shows up in your forecast before it shows up in your invoice.

Top comments (0)