GitHub Copilot’s shift to token-based billing has made the cost of AI-assisted software work visible after years of subscription pricing that felt almost unlimited. The monthly prices remain unchanged, but each plan now provides credits consumed according to the selected model and amount of computation.
Developers began reporting unexpectedly rapid depletion after the June 1, 2026 change. The complaints show how difficult it is to predict spend when coding assistants move from short completions into code review, long context and multi-agent workflows.
A subscription now functions as a credit allowance
Copilot Pro remains $10 per month, Pro+ $39, Business $19 per user and Enterprise $39 per user. A typical credit is worth one cent, giving Business users 1,900 monthly credits and Enterprise users 3,900.
Models consume those credits according to their token prices. One example cited for ChatGPT-5.2 charges $1.75 per million input tokens, $14 for output and $0.175 for cached context.
IDE code completions and next-edit suggestions remain free, but activities such as Code Review draw from the allowance. Users may purchase additional credits after the included amount runs out.
This makes the displayed subscription price a starting budget rather than a fixed total. Teams cannot estimate cost from seat count alone; usage pattern, model choice and output length now matter.
Agentic workflows expose the true compute bill
Community reports describe credits disappearing faster than expected for small visible changes. The final lines of code may conceal extensive context loading, reasoning, tool calls and rejected attempts.
That gap makes cost per edited line a poor measure. A difficult review that prevents a defect may be valuable even when few lines change, while an elaborate agent loop can consume tokens without improving the product.
Organisations should measure completed outcomes: time to merge, defects caught, developer review effort and incidents avoided. Billing data should be connected to repositories and workflow types without turning individual usage into a gamified leaderboard.
Earlier encouragement to “token max” rewarded activity rather than return. Under pay-per-use pricing, performative consumption becomes a direct cost. Large reasoning models should be reserved for tasks that benefit from them, while cheaper models handle routine transformations.
Engineering leaders need an AI FinOps discipline
Teams can set budgets by role and workflow, monitor anomalies and test model-routing policies. They should also account for retries, context caching, code review and parallel agents when forecasting a project.
Alternative tools may offer temporary price relief but often rely on the same frontier providers and may adopt similar usage billing. Self-hosted open models can improve control for suitable tasks, though they require infrastructure and may lack advanced coding harnesses.
The best response is not to abandon assistants or give every developer an unrestricted budget. Organisations need transparency at the moment a model is selected and alerts before allowances disappear.
Shared dashboards should reveal team trends without rewarding waste or penalising legitimate complex work.
GitHub should help users estimate the cost of a task and explain which components consumed credits. Predictability will influence trust as much as raw model quality.
The pricing change reflects an economic reality: inference, long context and agent orchestration are not free. The earlier flat-rate experience likely subsidised adoption while providers learned how deeply developers would integrate AI.
Copilot can still produce strong returns, but those returns must now be demonstrated. Engineering leaders who connect token spend to delivery quality can choose the right model for each task and defend productive use.
The transition’s discomfort comes from replacing a simple seat licence with a variable computing resource. That complexity is permanent. Software teams must learn to manage AI capacity with the same attention they already give cloud infrastructure, performance and reliability.
