How to reduce GitHub Copilot costs: seats, credits and budgets
By Jiri Sifalda
To reduce GitHub Copilot costs, cut the two parts of the bill one at a time. First remove the seats nobody uses and put each person on the right plan. Then control the AI credits: set user-level budgets, limit the expensive models, and turn on the hard stop for metered spend.
This guide covers each lever, what it saves, and the traps in GitHub's defaults that can cost you money.
Key takeaways
- The seat fee is fixed per granted seat. Idle seats are the easiest saving, but a removed seat stays billed until the cycle ends.
- Inline code completions do not use AI credits. Chat, the CLI and agents do, at a price per token that depends on the model.
- Paid usage past the included pool is on by default for Copilot Business and Copilot Enterprise.
- "Stop usage when budget limit is reached" is off by default. Without it, a budget only sends alerts.
- An enterprise budget does not cap seat fees. It caps only metered charges.
Know the two parts of the bill
GitHub lists Copilot Business at $19 per granted seat per month and Copilot Enterprise at $39 (GitHub Copilot plans). You pay that price for each seat, used or not.
The second part is AI credits. One AI credit is $0.01. Each seat adds credits to a pool that the whole organization shares. A Business seat adds 1,900 credits per month and an Enterprise seat adds 3,900 (usage-based billing for organizations and enterprises). Usage draws from the pool first. After the pool is empty, you pay for each extra credit.
Credits do not carry over. The pool resets on the first day of each calendar month at 00:00 UTC, and unused credits are lost.
Cut the seat fee
Remove idle seats
A seat with no activity for a full billing cycle still costs the full seat price. Find these seats with the last_activity_at date in the seat list API or the activity report. The unused Copilot seats report guide shows how to build that list per team, and why an empty date does not always mean an unused seat.
Two timing rules apply:
- A removed seat goes to "pending cancellation". The user keeps Copilot, and you keep paying, until the end of the current billing cycle (revoking access).
- Removing a seat mid-cycle does not shrink the credit pool until the next cycle (usage-based billing).
So run the review a few days before the cycle ends, and remove the seats then.
Put each person on the right plan
An Enterprise seat costs $20 more per month than a Business seat. It also adds 2,000 more credits to the pool, and 2,000 credits are $20 at $0.01 each. So a downgrade saves money only when the pool does not need those credits.
Before you downgrade a group of people, check the pool. If the pool runs out each month, the lost credits come back as paid usage at the same price, and you save nothing.
Give each person one license path
GitHub warns that when a user gets licenses from more than one organization, it picks one organization at random each cycle to bill the seat (budgets for usage-based billing). Give each person one license through one organization, so you do not pay twice and each budget applies to the right people.
Control AI credit spend
Know what uses credits
These features use AI credits: Copilot Chat, Copilot CLI, the Copilot cloud agent, Copilot Spaces, Spark and third-party coding agents (usage-based billing). Inline code completions and next edit suggestions do not. They stay unlimited, even for a user with a $0 budget.
This means you can block every credit feature for a person and they still keep completions in the IDE.
Limit the expensive models
The cost of a request depends on the model and the number of tokens. The price gap between models is large. Per 1 million output tokens, GitHub lists Claude Haiku 5.5 at $0.50, Claude Sonnet 5.5 at $10 and Claude Fable 5.2 at $40 (models and pricing for GitHub Copilot). The same work on Fable 5.2 can cost four times as much as on Sonnet 5.5.
Organization owners can enable or disable each model under Settings, Copilot, then Models (managing policies and features for Copilot in your organization). New models can become available by default, so check this page after each model launch (default availability of Copilot features and models).
Do not block the powerful models for everyone. Keep them for the people whose work needs them, and give them a higher budget.
Pick the code review effort level
Copilot code review chooses its model by itself, so you cannot pick a cheaper one. You can pick the effort level. GitHub estimates an average of about $0.25 per review at Lite, $1.50 at Balanced and $10 at Max, plus GitHub Actions minutes (models and pricing for GitHub Copilot). On a repository with many small pull requests, a lower effort level is the cheaper default.
Set budgets that actually stop spend
GitHub has several budget types. They work together, and two of the defaults let spend continue (budgets for usage-based billing).
- Set a universal user-level budget. It caps each user's total credit use in a cycle, pool credits included. It is always a hard stop.
- Raise it only where needed. Use a cost center user-level budget for a whole team, or an individual budget for one power user. An individual budget can expire at the end of the cycle, so a short-term raise removes itself.
- Decide on paid usage. On Business and Enterprise, the AI credits paid usage policy is on by default. To stop all spend past the pool, an administrator must turn it off.
- Add spending limits for metered charges. Set an organization, cost center or enterprise budget for the usage past the pool.
- Turn on the hard stop. "Stop usage when budget limit is reached" is off by default. Without it, charges keep growing past the limit.
Two more details matter:
- The enterprise budget does not cap seat fees. In GitHub's example, 400 Business seats cost $7,600 per month, so an enforced $5,000 enterprise budget allows up to $12,600 in total.
- The budget with the least headroom blocks the user first. If the user-level budgets together allow more than the pool holds, the difference becomes metered spend.
Review the bill every month
Each lever above needs fresh data. A seat can go idle, a team can move to a more expensive model, and a budget can be too low or too high. Check three things each month: idle seats, credit use per person, and which people are near their budget.
GitHub shows seats, credits and budgets, but it does not show what each person costs in total, and it does not include other tools such as Claude. The per-developer Copilot cost guide shows how to add each person's credits to their seat price.
BurntAI is built to put each developer's Copilot and Claude spend in one row, rank Copilot seats by last activity, and forecast who goes over budget before month end. See pricing to get on the list.