When GitHub Copilot moved to usage-based billing, an IT services firm with several thousand developers saw its bill multiplied by ten. "The x10 on GitHub costs, that's everybody," its head of generative AI told us. Since June 1, 2026, the seat keeps its price, but every chat, agent or code review request is paid at the rate of the chosen model. In the estates we analyse, 5 % of developers consume half of the credits. Two default settings, which take five minutes to check, let the bill run.
What should you check today in GitHub Copilot billing?
Everything sits in GitHub's billing settings.
- Open the AI usage report. GitHub breaks down the credits consumed by model, by user, by organisation and by cost centre. The per-user export, covering 31 days at most, gives the model and the repository of every request.
- Check the two default settings. Usage beyond the allowance is switched on by default for organisations and enterprises, and the option that stops consumption when a budget is reached is switched off by default. As long as nobody touches them, an exceeded budget sends an alert and billing continues.
- Set budgets by team or by cost centre. A user budget stops that person outright, with no fallback to a cheaper model. The enterprise budget, which GitHub distinguishes from a total monthly budget, only caps what is paid beyond the pooled allowance.
- Decide which models each organisation may use. Their availability is set by policy. In GitHub's price list, a token of the most expensive model costs a hundred times one of the cheapest.
No field in the report says what task a request served or what it produced.
GitHub Copilot plans: which one does a company need?
A company has two routes. The usual one is Copilot Business or Copilot Enterprise: licences it assigns to its developers, and consumption drawn first from a shared allowance of credits, which GitHub calls GitHub AI Credits. Since October 28, 2026, some GitHub Enterprise Cloud customers can take Copilot access instead, for now through GitHub's account teams: no allowance, usage billed from the first credit, and no Copilot seat charge for users who already hold a GitHub Enterprise Cloud licence, although the seat still has to be assigned to them. An enterprise cannot mix the two routes. The individual plans, Free, Pro, Pro+ and Max, plus a Student plan for verified students, are personal subscriptions.
| Plan | For whom | What is included |
|---|---|---|
| Free | an individual getting started | 2,000 code completions a month, limited chat and agent usage |
| Pro, Pro+, Max | an individual, who pays for it | unlimited completions and a monthly allowance of credits, larger from one plan to the next |
| Business | an organisation | unlimited completions and, per licence, a monthly allowance of credits pooled across the company |
| Enterprise | an organisation | the same, with an allowance per licence about twice the size of Business's |
| Copilot access | a GitHub Enterprise Cloud customer, for now through GitHub's account teams | no allowance: usage billed from the first credit; no Copilot seat charge for users with a GitHub Enterprise Cloud licence, though their seat must still be assigned; a per-user Copilot Standalone charge for the others |
Seat prices are on GitHub's plans page, which we read on November 6, 2026. We do not copy them here: they change, and what moves a company's bill is what is consumed beyond the allowance.
How much does GitHub Copilot cost per developer?
With Copilot Business or Enterprise, three things add up for each developer: the seat, at the price on GitHub's plans page; the usage beyond the pooled allowance, paid per credit at the rate of each model; and, for code review, GitHub Actions minutes. The seat is the fixed part. The rest depends on who uses which model, and how much.
The allowance is shared across the billing entity: the credits a light user leaves unused cover part of a heavy user's consumption, until the pool runs out. It refills in full on the 1st of each month, and whatever is left from the previous month is lost. With Copilot access, there is no allowance to share: each developer's consumption is billed from the first credit. Either way, an average cost per developer describes almost nobody: the real one is read in the AI usage report, user by user and model by model.
Should you put quotas everywhere?
The head of generative AI at that IT services firm did: "I put quotas everywhere, yeah. I'm creating frustration. Never mind." He knows who those quotas hit: "the users who cost a lot are the users who adopted it well." He spends his days closing the taps, and he gave us this advice: "Be the ones who reopen them."
At a consulting firm equipped with Claude Code, a uniform cap blocked one of the most productive workstations, where a good share of the work ran automatically. Messages kept landing on the head of AI's Slack: "Please, remove the cap, please," "because otherwise it drives me mad." We recommended changing that workstation's plan. We see caps hit the most advanced users first, rarely the ones who waste.
The other method starts by reading, team by team, what the expensive requests produce, then cuts what runs for nothing and keeps the rest open. A large public transport group already works this way: a default credit for all its developers, and Claude Code opened case by case, with the manager's approval. A quota that is too tight has a side effect too: some developers go back to their personal subscription, which is shadow AI.
The right model for the right task shows in the usage
The head of AI products and services at that transport group keeps telling those who ask him for volumes: "tokens don't mean anything. From one model to the next, they don't mean anything." Claude Code's documentation points the same way: Sonnet handles most coding tasks well for less than Opus, which it reserves for architectural decisions and multi-step reasoning. At the consulting firm, the agents following the work in the background ran on the smallest model, for a fraction of the cost per turn. Most requests go to the default model, which nobody chose for the task.
Why does the coding assistant bill climb so fast?
GitHub explained it when announcing the switch: a quick chat question and a multi-hour autonomous coding session could cost the same, and that model was no longer sustainable. Since June, the long session is paid by its weight in tokens.
That weight depends first on the chosen model. It also depends on context: according to Claude Code's documentation, a one-line question in a session open all day draws usage for the whole conversation, and unexpected spend usually comes from long sessions never cleared or from Opus left as the default model. Agents, finally, multiply the calls: an agent team in plan mode uses about seven times more tokens than a standard session. At GitHub, code review also takes Actions minutes, on a model GitHub picks automatically and does not show.
Consumption concentrates on a few people. OpenAI measures it in number of messages in its enterprise AI report: for code, users at the 95th percentile send seventeen times more than the median user. When those people move to a more expensive model, the whole bill follows. In the defence entity of a large group, the CTO saw the cost of GitHub Copilot become "x5, x10 for us."
GitHub Copilot Business and Enterprise: what stays per seat, what moves to usage?
| Since June 1, 2026 | |
|---|---|
| Per seat | the licence, at the same price; code completions and next edit suggestions, unlimited and without credits; a monthly allowance of credits per licence |
| The allowance | pooled across the whole billing entity, reset on the 1st of the month, never carried over |
| Per usage | chat, the CLI, the cloud agent, Spaces, Spark and third-party coding agents, at the model's per-token rate (input, output, cache); code review also consumes Actions minutes |
| Beyond the allowance | billed per credit, on by default; or blocked until next month if the administrator turns off paid usage; no more fallback to a cheaper model |
An idle seat feeds the pooled allowance, but it is still paid for. Microsoft 365 Copilot, the office suite assistant, is a different product with its own mechanics: see our page on Microsoft 365 Copilot licences.
What happened to "premium requests"?
They belong to the old system. GitHub now files per-request billing, and its per-model multipliers, under a legacy regime that only applies to some individual annual subscriptions.
And Claude Code or Cursor, same mechanics?
Each tool has a share included in the seat, then consumption at the models' rates.
- Claude Team and Enterprise, per seat: an allowance per seat, renewed over a five-hour window and a weekly window for Claude Code. Beyond it, prepaid usage credits at API rates, under spend limits per organisation, group or member.
- Claude Enterprise, usage-based: no consumption included in the seat; every token is paid at API rates, and usage billing cannot be turned off.
- Cursor Teams: weekly usage of Cursor's models per seat, then a prepaid balance shared by the team, which stops when it is empty or when the monthly limit is reached.
When the tool offers a choice between seat and usage, everything depends on the person's profile: for a heavy user, the seat is often much cheaper than the same volume paid per token, and for an occasional user, the reverse holds. A partner at an industrial group told us about the trade-off the other way round: on his usage-billed platform, most developers cost less than a seat, and the heaviest consumer ended up getting a licence. The same gap exists for office assistants: see an active licence is not a used licence.
What reading actual usage adds to the bill
Developers' impressions do not settle it, and neither do the finance team's. In a randomised trial run by METR in early 2025, experienced developers took 19 % longer with AI, while they believed they had saved 20 %. Faros AI saw twice as many pull requests merged in teams using AI, with review time almost doubled and no clear link to company results.
Usage logs, read team by team, show which tasks go through AI, on which model, at what cost and with what result. The reading starts from existing logs, with nothing installed on workstations, and covers usage, not people. That is the work of our audit of real AI usage. To place coding assistants within the whole estate, see our guide to AI in the enterprise.
