TL;DR:
- GitHub Copilot moved to an AI Credits billing model in June 2026 — basic completions are still included, but agent mode, premium models, and automated code reviews now consume credit budgets
- Teams on the Copilot Business plan need to evaluate expected agent usage before rolling out Copilot broadly to avoid surprise overage bills
- The credit system also unlocks per-agent model selection, letting teams use cheaper models for routine tasks and premium models only where the quality difference justifies the cost
If you have not revisited your Copilot plan settings since June, the billing model has changed underneath you. GitHub began rolling out the AI Credits system to Copilot Business and Enterprise accounts on June 1, 2026, and it fundamentally changes how the pricing works for anything beyond basic inline code completion.
The short version: using Copilot as an autocomplete tool works the same as before and remains included in your subscription. Using Copilot as an agent — autonomous multi-step task execution, repository-wide analysis, pull request generation, or premium model inference — now draws from a credit budget that varies by plan.
What the Credit System Covers
GitHub has divided Copilot’s feature set into two tiers under the new model.
Included with subscription:
- Inline code completions (all supported editors)
- Copilot Chat in IDE and GitHub.com (standard model)
- Code explanation and documentation assistance
- Basic test suggestions
Credit-consuming:
- Copilot agent mode (autonomous task execution across files)
- Code review on pull requests (automated review comments)
- Premium model inference (GPT-5 variants, Claude claude-opus-4-7, Gemini Ultra)
- Workspace-level queries (entire-repository context)
- Copilot Extensions that invoke third-party LLM calls
The credit costs are not flat. A routine agent task that opens a pull request to fix a simple bug costs considerably fewer credits than a workspace analysis session that involves planning a multi-file refactor. GitHub publishes a credit cost reference table in the documentation, but the practical implication is that agent-heavy workflows can consume budgets quickly if teams are not paying attention.
What Each Plan Gets
Monthly credit allowances by plan as of July 2026:
Copilot Individual — 1,000 credits/month. Enough for moderate agent use — a few dozen PR-opening tasks per month, or regular use of premium model responses.
Copilot Business — 1,500 credits per user per month. For a developer doing modest agent work (one or two automated PR tasks per day, occasional code review automation), this typically covers usage. Heavy agent users can exhaust a monthly budget in a week.
Copilot Enterprise — 2,500 credits per user per month, plus admin controls to set per-user limits and monitor consumption at the organisation level. The enterprise controls are where the credit system actually becomes manageable at scale.
Unused credits do not roll over to the following month. Overage credits can be purchased, but at a higher per-unit rate than the effective included cost.
How Teams Are Responding
The reaction from engineering teams has been mixed. For individual developers using Copilot as a productivity tool rather than an autonomous agent, the change is largely invisible. Completions keep working. Chat keeps working. The credit model only surfaces when you start using the more autonomous features.
The tension is with teams that were beginning to build agent-based workflows into their CI/CD pipelines — using Copilot to automatically review incoming PRs, generate draft summaries, or run recurring codebase analysis tasks. These automated, non-interactive uses can consume credits in bulk without anyone noticing until the monthly usage report arrives.
A few patterns that help:
Audit your integrations before the bill arrives. If you have Copilot connected to GitHub Actions, Webhooks, or third-party tools that trigger agent mode automatically, check the credit consumption rate now. GitHub’s usage dashboard in the admin console shows credit consumption broken down by feature and by user.
Set credit caps per user. Enterprise admins can cap individual user credit consumption to prevent runaway spending from any one team member’s experiments. Setting conservative initial caps and adjusting them up based on actual usage is safer than leaving the default uncapped.
Use model selection deliberately. The credit system exposes per-agent model selection — you can configure certain tasks to use a cheaper, faster model and reserve premium models for use cases where the quality difference matters. A PR description generator probably does not need GPT-5 Ultra. A complex architectural review might.
Treat agents as team tooling, not individual perks. If agent-mode tasks are being triggered by individuals for exploratory use, credit consumption will be high and variable. Centralising agent workflows as shared team tools with defined trigger conditions makes usage more predictable and easier to budget.
The Upside of the Credit System
The credit model is not purely a cost increase. It also unlocks capabilities that were not available under the old flat-rate pricing.
Multi-model workflows within a single workflow are now supported. You can instruct Copilot to use haiku-class models for classification and tagging tasks, standard models for routine generation, and premium models only for final synthesis steps. For teams that had been defaulting everything to premium models out of habit, this alone can reduce effective credit consumption while maintaining output quality on the tasks that actually benefit from it.
The usage analytics are also more granular than they were. Teams that previously had no visibility into how Copilot was being used now have per-feature, per-user, and per-repository breakdowns. That data is genuinely useful for making tool purchasing decisions.
What to Do This Week
If you are on a Copilot Business plan and have not looked at the admin usage dashboard since May, check it now. The June billing cycle will have produced the first month of credit usage data. Look specifically at any automated workflows or CI integrations that trigger agent mode, as these tend to produce the most surprising consumption numbers.
If agent usage is running high, the options are to increase the budget (buying additional credits), reduce automated trigger frequency, or switch automated review tasks to standard models rather than premium.