Copilot's bonus credits end August 31. What breaks in September?

GitHub Copilot's promotional AI credits stop on August 31, 2026. Nothing about how your team writes code changes on September 1. The invoice does.

Copilot Business drops from 3,000 included credits per user each month to 1,900, and Copilot Enterprise drops from 7,000 to 3,900, according to GitHub's billing docs. That is a 36.7% cut for Business and a 44.3% cut for Enterprise, computed from the credit allowances GitHub publishes and applied to teams whose usage has not moved at all.

Key takeaways

• 1 AI credit = $0.01 USD, GitHub's docs state, so the Business pool falls from $30 to $19 of monthly usage per seat and Enterprise falls from $70 to $39.
• Credits are consumed on token usage, not on request counts, so agent-mode and long-context work drains a pool far faster than inline completions.
• Set user and org budgets before September 1. GitHub's billing docs describe four separate budget layers, and none of them is on by default.

What actually expires on August 31

The promotion, not the plan. GitHub Copilot AI credits are the metered currency introduced when Copilot moved to usage-based billing on June 1, 2026. Each plan includes a dollar amount of credits, and usage beyond that is billed as overage.

The June change came with three months of extra credits for existing orgs. Business seats carried an additional $30 per user per month, and Enterprise seats an additional $70, GitHub said at launch. Those extras cover June, July and August 2026.

The base allowances never changed. Copilot Business remains $19 per user per month including $19 in credits. Copilot Enterprise remains $39 per user per month including $39 in credits.

The September invoice, in numbers

Copilot Business is GitHub's per-seat team plan, priced at $19 per granted seat per month. Here is the same team before and after, on a 50-seat Copilot Business plan.

Line	Through August 31, 2026	From September 1, 2026
Seat price	$19 per user per month	$19 per user per month
Included credits per user	3,000 ($30)	1,900 ($19)
Included usage, 50 seats	$1,500	$950
Seat cost, 50 seats	$950	$950
Usage covered before overage	$1,500	$950

A 50-seat Business org loses $550 a month of included usage. A 200-seat Enterprise org loses $6,200, because the Enterprise pool falls by $31 per seat.

Whether that shows up as a bill depends on one thing. Was your team already using more than the base allowance while the promo hid it? Teams that never touched the extra credits will see no change at all.

Why token-metered credits drain faster than you expect

Credits burn on tokens, not on actions. GitHub's billing concepts page states that input tokens, output tokens and cached tokens are each priced by model and then converted into credits.

Agent mode is the Copilot setting where the model plans a multi-step task and edits files on its own, rather than completing the line you are typing. Three habits dominate the credit number.

1. Agent mode on a large repository. An agent reads files before it edits them. Every file read is input tokens, charged at the model's rate.
2. Long-context refactors. A 200,000-token context window costs whether or not the answer needed all of it.
3. Premium model selection. Model choice, not prompt count, is the biggest single lever on a credit pool.

Inline completions are the cheap case. A developer who uses Copilot as autocomplete may never approach $19 a month. A developer running agentic tasks daily can pass it in a week.

How to find your real number before September

Do not estimate. Pull the data GitHub already has, described in Monitoring your GitHub AI Credits usage.

Since June 19, 2026, the Copilot usage metrics API also returns an ai_credits_used field for each user, in both the one-day and 28-day user-level reports. It is a usage signal, not a billed total. Still, it is the fastest way to see which seats sit above the base allowance.

1. Open the usage report for July 2026, a month fully inside the promotion.
2. Compare credit use per user against 1,900 credits, not 3,000.
3. Count how many seats exceeded 1,900. Those are the seats that will generate overage in September.
4. Multiply the excess credits by $0.01 to get the monthly overage in dollars.

That arithmetic takes ten minutes and turns a surprise into a budget line.

Set the budgets, because they are off by default

GitHub's usage-based billing guide describes four controls: user-level budgets, cost-center budgets, enterprise spending limits, and organization-level budgets. They stack. For example, a user budget can cap one developer at $25 a month. The org budget then caps what happens after the shared pool runs dry.

The useful pattern for most teams is two layers. Set an organization-level budget at the number your finance team approved. Then set user budgets a little above the median usage you measured in July. That stops one runaway agent loop from eating the pool.

A budget that triggers an alert is not the same as a budget that blocks spend. Configure which one you want before the first September invoice, not after it.

Is this the industry direction, or just GitHub?

Both, and the contrast is instructive. GitHub moved to usage-based billing and let a promotional cushion lapse. Anthropic went the other way on August 10, 2026, confirming that Claude Sonnet 5 keeps its $2 and $10 per million token introductory pricing instead of rising to $3 and $15 on September 1.

The two moves point at the same underlying fact. Model inference is getting cheaper per token, while agentic tools consume far more tokens per task. Vendors are choosing different places to absorb that. The cheapest AI API is not the cheapest to run, and a seat price with a metered pool behind it is the clearest example of the gap.

For a US or European team, the practical consequence is that seat count no longer predicts spend. Usage does. That is a different budgeting exercise from the one most engineering managers ran in 2024.

Should you switch tools over this?

Probably not, and the reason is arithmetic rather than loyalty. A Business seat still includes $19 of usage for $19, so the credit pool is effectively at cost. Consider what a switch actually buys: another vendor's metered pool, another integration, and a migration for every developer.

The better response is to measure, cap, and decide in October with two months of real data. Tools like ChatGPT Business Premium sit at a much higher seat price. Those comparisons only mean something once you know your own usage curve.

Switch if your measured overage exceeds the difference in seat price elsewhere. Otherwise, set the budgets and move on.

Frequently asked questions

When do GitHub Copilot's promotional AI credits expire?

The promotional credits run through August 31, 2026, covering June, July and August. From September 1, 2026, Copilot Business returns to 1,900 included credits per user per month and Copilot Enterprise returns to 3,900, per GitHub's billing docs.

How much is one GitHub AI credit worth?

One AI credit equals $0.01 USD, GitHub's docs state. So 1,900 credits is $19 of included usage per user each month on Copilot Business, and 3,900 credits is $39 on Copilot Enterprise.

Will my Copilot bill go up in September 2026?

Only if your team consumed more than the base allowance while the promotion was active. Pull July 2026 usage and compare each user against 1,900 credits for Business or 3,900 for Enterprise. Seats above that line generate overage at $0.01 per credit.

What uses the most Copilot credits?

Token volume, not request count. Agent-mode runs across large repositories, long-context refactors, and premium model selection consume the most, because input, output and cached tokens are all priced by model and converted into credits.

Can I cap Copilot spending per developer?

Yes. GitHub's billing docs describes user-level budgets, cost-center budgets, organization-level budgets, and enterprise spending limits. None of them is on by default. Set them up before the billing cycle you want to protect.