On June 1, GitHub retired the currency it had used to bill for its AI coding assistant and replaced it with a new one: tokens. Every Copilot plan now works on a monthly allotment of GitHub AI Credits, consumed by the input, output and cached tokens each interaction uses, priced at the published rates of the underlying model. One credit costs a penny.
The change ends the era of the flat fee. Copilot Pro remains $10 a month, but the $10 now buys $10 worth of credits, and heavy users can buy more when the allotment runs out. Pro+ stays at $39, Business at $19 per user and Enterprise at $39 per user, each including a matching value of credits, with additional usage billed at the end of the month. Existing Business and Enterprise customers get promotional credit increases for June, July and August, $30 and $70 a month respectively, to smooth the transition.
The mechanics are designed to make usage visible. An interaction’s cost depends on the model and the token count: a quick question to a lightweight model costs fractions of a credit, while a long agentic session working across a large codebase with a frontier reasoning model costs much more. GitHub is giving administrators new controls to match: budgets at the enterprise, cost center and user level, pooled credits across organizations instead of isolated per-user buckets, and the option to cap additional spending entirely. The pooled structure means a hundred-user company draws from one shared reservoir, so power users can borrow against the unused allowance of light users.
The reasons for the change are financial and structural. GitHub’s parent, Microsoft, has been running Copilot as a fixed-price product with usage that varies wildly among customers, and the flat fee forced the company to gate heavy users, capping what they could do and frustrating the developers who depended on the tool. Usage differences among customers have reached a hundredfold, according to people familiar with the product’s economics, a spread no flat price can capture. Under the new model, light users pay less than they did, heavy users pay in proportion to what they consume, and the product stops being a fixed cost that the company must ration.
The change also signals where the product category is heading. A tool that moves to usage-based billing is a tool that has become infrastructure: metering arrives when a product is used daily, predictably and at scale, and when its consumption is measurable in the same way a cloud bill is measurable. Copilot, launched as a code completion feature in 2021, has evolved into a platform that edits files, reviews pull requests and runs agentic workflows, and the new pricing reflects the shift from a feature to a service.
The billing change lands in the middle of a broader repricing of AI products. Model makers have been cutting per-token prices while raising the capabilities of their frontier models, and usage-based pricing lets GitHub pass those economics through without renegotiating contracts. It also aligns the company’s incentives with its customers’: Microsoft no longer needs to discourage heavy use, because heavy use is now a source of revenue rather than a cost to absorb.
For developers, the change is a shift in habits. Copilot’s fallback mechanism, which quietly switched users to cheaper models when they exhausted their plan, is gone; when credits run out, usage stops unless the user or their employer approves more spending. The new system rewards model choice, since cheaper models extend the same credit budget, and GitHub is offering a 10% discount on model costs for customers using its automatic model selection.
The market has been here before. Cloud computing moved to consumption pricing, then to reserved capacity, and then back, as customers learned to manage both. Software companies learned that metering changes behavior: budgets make costs visible, and visible costs get managed. Copilot’s shift, applied to the fastest-growing developer tool in Microsoft’s portfolio, is likely to become the template for how AI software is sold across the industry.
For enterprises, the change lands alongside a broader shift in how AI budgets are governed. Finance teams that once wrote one check for a software license now manage variable cloud bills, model API costs and agent workloads as separate line items, and Copilot’s metering gives them the same granularity they have for compute. The new budgets, set at the enterprise, cost-center or user level, let an organization cap an AI tool the way it caps a cloud account, with notifications as thresholds approach. Procurement officers who had to accept Copilot as a take-it-or-leave-it subscription now have the controls that procurement departments expect.
The practical question for customers is what their bills will look like. A developer who uses Copilot for occasional completions will spend less than before. A team running agentic workflows across large codebases will spend more, and its finance department will be able to see exactly where the money went. That, in the end, is the point of the change: the era of the flat-fee AI assistant is over, and the era of the metered one has begun.
