The deal with Copilot, until now, was simple. Pay $10 a month, get unlimited code completions, unlimited Chat, access to the cloud agent. No meter running. No end-of-month surprise. The subscription model meant developers could use Copilot freely without calculating cost per session.
That deal ends June 1.
GitHub is moving all Copilot plans to usage-based billing with AI Credits. The plan prices are unchanged. But the cost of using Copilot’s most powerful features — Chat, the cloud agent, code review in Actions — is now variable. It depends on how much you use it.
This matters more than the dollar amount suggests.
What the old model got right
Predictability is not a small thing. A flat $10/month subscription is acquirable by a single engineer without an expense report. It fits in a personal budget alongside other subscriptions. The finance question is: “is $10/month worth it?” — a question that requires no ongoing tracking.
Usage-based billing changes the question to: “how much did I use it this month?” The cognitive overhead is different. You start thinking before starting a long agent task: is this worth the Credits? You check your usage partway through the month. You set spending alerts. These are minor friction points individually, but they add up.
The flat subscription model also made Copilot easy to deploy at the team level. A $19/user/month subscription has a known budget line. Usage-based billing at the team level means the budget line is a range, not a number.
What “usage-based” actually means in practice
The Credits model is straightforward in principle. Each plan includes a monthly Credit allotment. Code completions don’t consume Credits — they stay effectively unlimited. Chat, the cloud agent, code review, and third-party extensions do consume Credits.
The cost per feature is a function of how much compute that feature uses. A single Chat message costs differently than a cloud agent session that runs for 20 minutes across several files. GitHub is publishing the Credit rates for each feature before June 1, which helps — but Credit rates can change, and “20 minutes of cloud agent” covers a very wide range of actual compute.
The hard part is that agentic tasks are variable by design. A quick refactor might take 3 minutes of agent time. A complex feature might take 45. You don’t know before starting.
Predictability beats price
I’ve watched this pattern before in cloud services. AWS’s early adopters weren’t scared off by on-demand pricing — they were scared off by bill shock. The first month’s surprise bill, or the mental overhead of estimating costs before every action, creates friction that flat pricing doesn’t.
GitHub Copilot’s most enthusiastic users are the developers doing the most agentic work. These are the users who will be most affected by per-usage billing and who will think hardest before starting a long agent run. The users who use Copilot lightly for Chat suggestions will barely notice the change.
This is a structural tension in the billing model. The users with the highest engagement are the ones being asked to shift from predictable to variable costs.
Is this an industry move?
Yes. OpenAI has always charged per-token for API access. Anthropic charges per-token. Cursor charges per-request for its premium models. The exception was Copilot’s flat subscription, which GitHub held onto longer than the others.
The argument for usage-based billing is that compute genuinely costs more for agentic tasks. A Chat message that triggers a 30-second multi-step tool use costs GitHub more than an autocomplete suggestion. Flat pricing meant the heavy agentic users were subsidized by light users. Credits make the cost more proportional to actual consumption.
That argument is correct, as far as it goes. The problem is that “proportional to consumption” is not the same as “predictable for the buyer.” These are different properties.
The real question
The question that matters is whether the agentic features are worth what they actually cost to use — not the monthly subscription rate, but the marginal cost of a specific task.
For most developers, the answer is probably yes for the features they use most. Copilot Chat is not expensive on a per-message basis. The cloud agent, for the tasks where it works well, delivers significant time savings per session.
The risk is chilling effect. If developers start calculating the Credit cost before starting an agent task, they’ll use the agent less. They’ll stick to the cheaper Chat or just write the code themselves. The features that look best in GitHub’s demos require willingness to run them freely — and usage-based billing makes that willingness harder to maintain.
GitHub’s bet is that once developers see the Credits costs, they’ll find them acceptable. That bet will be tested in June and July when teams get their first usage-based bills.
For specific steps on managing spend, see How to keep Copilot bills predictable after the June 1 switch.
