Keep your Copilot bill predictable after the Sept 1 credit drop

You open the Copilot dashboard on a weekend morning and see the same assistant, now with a meter. Code completions and next edit suggestions stayed free and unlimited for paid Copilot plans after the June 1, 2026 billing change. On paid plans, they are not billed as credits. The agent work that reads more context is the line item to watch. Most side projects stay side projects, and that is fine. A predictable bill keeps the weekend habit from becoming friction.

The cliff is September 1, not the June switch

On April 27, 2026, GitHub's CPO said all Copilot plans would leave Premium Request Units and start using AI Credits on June 1, 2026. GitHub Copilot replaced premium request units with usage-based AI Credits on that date, and one credit equals a cent. Each paid plan gets a monthly credit allowance matching its subscription price, and token usage is charged at the model's published API rate. The June switch was the setup, and September 1 is the cliff.

For Copilot Business and Enterprise organizations, the promotional included-credit allowance ends August 31, 2026. Promotional credits ran from June through August 2026, adding $30 per user per month for Business plans and $70 for Enterprise plans; the extra credits disappear on September 1, 2026. The standard included allowance becomes $19 per seat per month for Copilot Business and $39 per seat per month for Copilot Enterprise, down from $30 and $70 during the promotion. With 100 Copilot Business seats, the standard monthly pool falls from 300,000 promotional credits to 190,000, a cut of 110,000 credits per month. A solo side project may not feel the seat math, but the same credit meter applies.

The quiet parts of the tool do not move the total. The noisy parts ask the model to read more context, run longer tasks, or keep a conversation going, and those are the parts that spend credits. A side project may need them only when debugging a stubborn bug or wiring up a new feature. That is the usage worth auditing before the promotional window closes.

Three steps keep the bill predictable

The update also introduces controls for capping spend, reviewing usage, and choosing models. By default, GitHub keeps billing after AI credits run out rather than blocking usage, so a customer must enable a spending cap. A weekend builder needs a short written policy, not a vague intention.

A written policy does not need to be long. It needs to answer two questions: what happens when the included allowance is gone, and what model will carry the work if the answer is stop. If you cannot answer both, the dashboard is not finished with you.

Keep this 3-step guardrail next to your terminal. Each step is small enough to finish in a quiet stretch.

• Baseline usage: open the usage dashboard in August and note the current credit total.
• Choose bill-vs-block: set a spending cap at a number you can defend, or leave the default billing on if you can absorb overage.
• Price the fallback: choose a model you will actually use when premium features stop, test it on a real task before the promotional window closes, and note which features you keep paying for.

The dashboard makes the bill visible, and a cap turns the default from surprise into a boundary. Pricing the fallback keeps the next session from starting with an unknown cost. Do the audit in August, while the promotional credits are still in the dashboard.

Price the fallback before you need it

Copilot will not automatically switch to a lower-cost model when an organization's credit budget is exhausted. The pre-June 1, 2026 automatic downgrade to a cheaper model is gone, so depleted AI credits pause premium features until the next billing cycle or a top-up. Sonnet 5 has an introductory Copilot rate that is a discount, and it later matches Sonnet 4.6's rate card.

Price the post-discount rate before you rely on a model your side project leans on. A fallback that is too slow for your workflow calls for a higher cap or smaller usage. If it is good enough, keep the cap low. A weekend builder can keep the scope cheerful: the app still runs, the completions still appear, and the bill stays predictable. It also keeps the side project from asking for more hours than you have. Then close the dashboard and go back to the project.