← FinOpsAid blog

Your September Copilot bill is already decided. Most teams won't look until it arrives.

Copilot's promotional credits ran into August 2026, so September is the first unsubsidised invoice. The consumption that produces it is happening now, and it is visible now — if anyone is looking.

There is a version of the next five weeks where your September GitHub invoice arrives and contains a number nobody predicted, and you find out about it at the same moment your CFO does.

There is another version where you already know roughly what it says, because you looked at the consumption that produces it while it was still happening.

The difference between those two versions is not budget, headcount, or tooling spend. It is about a hundred minutes of attention, applied before the end of August rather than after.

The invoice is not a future event

This is the part worth sitting with, because it inverts how most people think about a bill.

Your September invoice is not generated in September. It is generated by consumption that is happening right now, in the current billing period, in your developers' editors. By the time the invoice arrives, every decision that produced it has already been made — every agent run, every model selection, every seat that sat idle.

An invoice is not a forecast. It is a record. And a record of something that is currently visible is not something you have to wait for.

That is the entire argument. Not "you should track your spend" — everyone nods at that and nothing changes. The specific claim is narrower and more uncomfortable: the number is already knowable, and most organizations are choosing not to know it.

Three facts that make September different

Not a trend piece. Three dated, checkable facts.

On 1 June 2026, Copilot moved to usage-based billing. Headline prices did not change — Business is still $19 per user per month, Enterprise still $39. What changed is what those prices buy. Each subscription now carries an allowance of AI credits equal to its price, and consumption past that allowance is a real, variable charge. A flat subscription became a floor with a meter on top. The full change is here.

Business and Enterprise accounts received promotional credits running into August 2026. So consumption happened, the ledger moved, and the invoice stayed flat. Two months of usage-based billing arrived looking exactly like the two months before it.

Which makes September the first genuinely unsubsidised bill for many organizations.

Put those together and you get a specific, unusual situation: a pricing model changed, the feedback signal was muted for its first two months, and the mute comes off next month.

Five questions. Most engineering organizations cannot answer any of them.

This is the honest test, and it takes about ninety seconds. Not "do you have a cost tool" — anyone can buy one. These are the questions the tool exists to answer.

1. What percentage of its credit allowance is your average developer using this month? Not total credits. The ratio. A team at 20% and a team at 180% need opposite conversations, and a single total hides both.
2. Which surfaces is that consumption coming from? Completions and Next Edit suggestions are free and unlimited. Chat, the coding agent and code review are metered. If you don't know your split, you don't know your exposure — and the free surface is the high-volume one.
3. What share of your credits went to frontier models versus cheap ones? Same work, same volume, different dropdown, very different bill.
4. Which teams changed shape in the last 60 days? Agent adoption doesn't announce itself. It shows up as a consumption curve that changes slope.
5. How many seats are you paying full subscription for that consumed nothing at all? Still the cheapest win available, and unrelated to everything above.

If you cannot answer these, the problem is not that you are overspending. You may not be. The problem is that you have no way to find out until an invoice tells you — and by then the period it describes is closed.

What "behind" actually costs

It is tempting to frame this as money. It usually isn't, and pretending otherwise is the kind of overstatement that gets a cost programme dismissed.

What it costs is position. There is a specific professional moment — most people reading this have had it — where a finance partner asks a reasonable question about your own budget and the honest answer is I'll have to find out. Nothing is technically wrong. Your credibility on every number you present afterwards is nonetheless lower than it was ninety seconds earlier.

Organizations that can answer those five questions aren't spending less by default. They are spending the same money on purpose, which is a completely different conversation to be in when someone asks about the AI budget. They get to say here is what it costs, here is why, and here is what we chose not to do. Everyone else gets to say the invoice went up.

The gap between those two positions is not a tooling gap. It is roughly one nightly sync and somebody having looked.

The gap is also getting harder to close

GitHub is shipping into this space fast. In July 2026 it added cost centers with pooled AI credits and budgets settable at enterprise, cost-center and per-user level. It also shipped repository-level Copilot usage metrics, and added Claude Opus 5, Gemini 3.6 Flash, three GPT-5.6 variants and Kimi K2.7 to the model picker — each with its own rate.

That is good news and a moving target at once. The organizations building the habit now inherit each of those capabilities as an upgrade. The ones starting in six months will be learning a pricing model, a set of native controls, and a model roster that has all changed twice since this post was written.

Compounding runs in both directions. Cost discipline compounds because each month of history makes the next month's anomaly legible. Its absence compounds too — every month without a baseline is a month you can't compare anything to.

What to actually do before the end of August

Concrete, roughly in order of return. None of this takes a project.

Pull credit consumption per user against allowance. One number, expressed as a ratio. This alone tells you whether September is a non-event or a conversation.

Split your consumption by surface. If it is mostly chat, your exposure is bounded. If the coding agent has been adopted somewhere, that is where your variable cost lives.

Look at model mix by consumption share, not request count. A handful of frontier-model agent runs outweighs thousands of cheap chat exchanges, so request counts will actively mislead you.

Count seats that consumed nothing. Full subscription, zero usage. Check the names before acting — leave and long-term absence look identical to idleness — but this is the one saving that needs no behaviour change from anyone.

Set one credit budget, not five. Denominated in credits rather than dollars, because a dollar budget on AI spend stays at zero until the allowance is exhausted and then warns you far too late. Credits climb from day one.

Write down this month's number. Even if you do nothing else. September is only unreadable if you have nothing to compare it to, and a single recorded baseline fixes that permanently.

The one sentence version

The consumption that decides your September invoice is happening now and is already visible. Reading it costs an afternoon. Not reading it costs the ability to say anything about your own budget except what the invoice says — and it says it a month too late to matter.

Frequently asked questions

Why will my Copilot bill jump in September 2026?

Because promotional credits issued to Business and Enterprise accounts ran into August 2026, cushioning the first months of usage-based billing. Consumption was metered and the ledger moved, but the invoice stayed flat. September is the first billing period for many organizations where actual consumption above the allowance is charged in full.

Our June and July Copilot bills didn't change. Does that mean we're unaffected?

Not necessarily. A flat invoice during the promotional period reflects credits absorbing the charge, not an absence of consumption. The reliable check is credit consumption against allowance per user for the current month, which is visible now and is what September's invoice will be based on.

How can I predict my September Copilot invoice?

Look at current credit consumption as a ratio of each user's allowance, then break it down by surface and by model. Consumption above allowance is the variable component of the bill, so a team consistently over 100% of allowance this month is what a higher invoice looks like next month.

Isn't it too late to change anything before September?

No. The largest levers act immediately and need no project: reclaiming seats that consumed nothing, publishing a default model per use case, and setting a credit-denominated budget so the warning arrives mid-cycle rather than after it. All three change consumption in the period they are applied.

Does using Copilot more always mean a bigger bill?

No, and this is the most commonly missed point. Code completions and Next Edit suggestions consume no credits on any plan, however heavily used. Credits are consumed by chat, the coding agent and code review. An organization can have very high Copilot adoption and near-zero credit consumption.

You can read your own consumption before the invoice does. Connect your GitHub org — read-only, nothing in your org is ever changed, and free during beta — or explore the demo dashboard first.