AI credits
AI credits, explained by what actually spends them.
Short answer
GitHub Copilot AI credits are the metered half of Copilot: consumption charged against token usage at the rate of whichever model served the request. Code completions consume none. Chat, code review and the coding agent do. Each plan includes an allowance, and usage beyond it is billed as overage — which is why the model default matters so much.
Free surfaces and metered surfaces
The most important fact about AI credits is also the most routinely missed: the thing developers do most often is free. Code completions and next-edit suggestions — ghost text you accept with a tab — are included in every plan and consume no credits at any volume. A developer who lives in completions all day costs you their subscription and nothing more.
Credits are consumed where a model does substantial work on your behalf: chat, code review, and the coding agent. The economics differ sharply between those three. A chat exchange is bounded. A code review scales with the size of the diff. An agent task is unbounded by nature — it reads, plans, edits, runs and re-reads until it is done.
The consequence is an inversion that catches people out: your credit burn correlates poorly with how much your developers like Copilot.
Tokens are the unit, and input counts
Credits are consumed against token usage — input, output and cached tokens, each at the published rate for the model that handled the request.
Two things follow. First, input costs money before anything is generated. A prompt that drags in a large file, a long conversation history, or a wide swathe of repository context has already spent before the model writes a line. Verbose context is a cost decision, not only a quality one.
Second, the model is a price multiplier. Rates differ between models by a large factor, so the same task costs a large multiple more on one model than another. In most organizations model choice is set by individual preference with no guidance at all, which makes it the single most direct lever available and the one least likely to have been pulled.
Allowances, overage and the shape of the month
Each Copilot plan bundles a monthly allowance of metered usage; consumption beyond it is charged as overage. That structure means credit spend is not a smooth line — it is free until a threshold and then linear, and the only thing that determines which side of the threshold you land on is the pace of the first three weeks.
Watch burn rate against the calendar, not cumulative totals. Sixty per cent of the allowance consumed by the tenth is an overage you can still prevent. The same sixty per cent on the twenty-sixth is a plan working exactly as intended.
The reporting problem nobody warns you about
The credit signal GitHub exposes is a per-user daily hint. It is genuinely useful — it is how you find the agent-heavy team — but it is not an invoice, it does not carry a model breakdown per request, and it does not survive indefinitely. Billed amounts come separately, from the billing API, and lag it.
GitHub gives you two different kinds of number and they are not interchangeable. Usage signals — a seat's last-activity timestamp, minutes consumed, a daily per-user credit hint — arrive fast and describe behaviour. Billed amounts arrive from the billing API, lag the signals by a few days by design, and are what finance will actually be invoiced. Merge the two and you get a dashboard that is confidently wrong. Keep them apart, anchor the signal to the billed total, and show the gap.
FinOpsAid keeps both, side by side, with the reconciliation between them shown rather than hidden — and labels every modeled figure as an estimate with the date it was computed.
FAQ
Frequently asked questions
What are GitHub Copilot AI credits?
They are the metered portion of Copilot. Where a Copilot seat is a fixed monthly subscription, credits are consumption charged against token usage — input, output and cached — at the published rate of the model serving each request. Every plan includes a monthly allowance, and usage beyond that allowance is billed as overage.
Which Copilot features use AI credits?
Copilot chat, code review and the coding agent consume credits. Code completions and next-edit suggestions do not — they are included in every plan and unmetered at any volume. The coding agent is by far the heaviest consumer because a single task reads, plans, edits and re-reads across many steps, each of which is tokens.
Why did our AI credit usage jump without more developers?
Almost always because a team adopted the coding agent, or a default model changed. Agent runs are unbounded in a way chat is not, and model rates differ by a large factor. Headcount is rarely the variable — surface mix and model mix are. Both are visible in per-user daily usage if you retain the history to compare against.
Can we cap AI credit spend?
GitHub's budget controls let you set spending limits and alerts on metered usage at the account level. They are a backstop, not a management strategy: a cap tells you that you have overspent, it does not tell you which team or which surface did it. Use caps together with per-team attribution, not instead of it.
Try it
See these numbers for your own org
Connect your organization read-only and FinOpsAid meters it nightly — seats, credits, minutes, Codespaces and storage, attributed to teams and reconciled against your billing API. Every modeled figure is labelled an estimate, with the date it was computed.