GitHub Copilot’s usage-based switch: what it costs to run an agent now
One engineer on the team had taken the “let the agent run” pitch literally. Within nine working days, they had burned through a full month of GitHub Copilot credits. Long autonomous sessions: read the repo, plan, edit, run tests, read failures, retry. The agent did exactly what it was designed to do. The bill just finally reflected what that actually costs.
On June 1, 2026, GitHub moved every Copilot plan off flat subscriptions and onto usage-based billing. It was not a surprise to GitHub. It was a surprise to most teams using the product.
️ What Changed on June 1, 2026
GitHub replaced its old premium request unit (PRU) system with GitHub AI Credits. One credit equals $0.01. Every interaction that burns tokens now bills against those credits at each model’s published API rate.
Base plan prices stayed put:
- Copilot Pro: $10/month
- Copilot Pro+: $39/month
- Copilot Business: $19/user/month
- Copilot Enterprise: $39/user/month
The catch is that each plan’s included credits are pegged to the same dollar figure as the plan price. Use more than the plan’s worth of compute, and overages bill on top. Code completions and Next Edit Suggestions remain unlimited and free. Nearly everything else, including chat, agent runs, and code review, is now metered.
A few other details worth tracking: annual plans are being retired, the old “fall back to a cheaper model when you run out” safety valve is gone, and Copilot code review now also consumes GitHub Actions minutes in addition to credits. GitHub launched a preview-bill experience in May 2026 so teams could see projected costs before the switch. The previews surprised people. The bills surprised them more.
Why Agents Broke Flat Pricing
A flat subscription is a bet that average usage stays bounded. For a human typing in an editor, it does. You can only read, think, and accept suggestions so fast. That biological ceiling is what made “unlimited” pricing safe for years.
An autonomous coding agent has no such ceiling. Give it a goal and it will decompose the task, read files for context, draft a change, run the test suite, parse the failures, revise, and loop, for as long as the problem takes or the budget allows. The cost is a function of token volume, and token volume is a function of wall-clock time the agent spends thinking, not of headcount.
GitHub acknowledged this directly on April 27, 2026, stating that “a quick chat question and a multi-hour autonomous coding session can cost the user the same amount” under the old model, and that Copilot “is not the same product it was a year ago.” Once running the agent became the dominant usage pattern, the flat subscription’s math stopped working.
There is a second pressure point: frontier models are getting more expensive, not cheaper. Claude Fable 5 lists at $10 per million input tokens and $50 per million output tokens. An agent that defaults to the most capable model at every step bills like one.
What AI Coding Actually Costs Per Month Now
Independent cost analyses published around the June 2026 transition put the numbers in stark terms:
| User Type | Estimated Monthly Cost (Published Rates) | Old Flat Plan |
|---|---|---|
| Light user (a few chats, small edits) | ~$36/month | $20 |
| Daily pro user | ~$178/month | $100 |
| Full-day agent user | ~$594/month | $200 |
The flat tiers were not generous pricing. They were a subsidy, and the heaviest users were the most subsidized. That subsidy is what is being withdrawn.
Within days of June 1, developer forums filled with screenshots of projected overage bills running from hundreds to thousands of dollars. TechCrunch on May 30 quoted developers calling the new token-based billing “a joke.” The Register on June 2 reported users threatening to quit the product outright.
The shift was not limited to GitHub. Cursor, Windsurf/Devin, and the Anthropic API all repriced in June 2026 as well. This is a category-wide change, not a single vendor decision.
One Detail That Changes Everything
Under usage-based billing, if you do not set an additional spending budget, your agent simply stops when the credits run out. It does not overcharge your card. It dies. Your tool stops mid-sprint because someone forgot to watch the meter.
For three years, the worst case of an AI coding tool was a wrong suggestion. Now the worst case is a hard stop in the middle of a sprint. The cloud bill, two decades after it came for infrastructure, has arrived for the dev team.
Three Habits That Keep the Bill Predictable
Model selection is now a per-task cost decision, not a default you set once. These three habits are the difference between calm July invoices and chaotic ones.
1. Scope the agent before you unleash it
An agent told to “fix the bug” with no constraints will read half the repo to orient itself. An agent given two relevant files and a clear acceptance test does a fraction of the token work. Tight context and a checkable definition of done make agents both more reliable and cheaper to run.
2. Route by task, not by default
Reserve the frontier model for steps that genuinely need hard reasoning or complex refactors. Use a cheaper model for boilerplate, search, and first drafts. With per-token billing, the model picker is a cost dial. Leaving it pinned to “most capable” is the single most expensive default in your workflow.
3. Make the meter visible before the bill arrives
Set budgets at the user and team level. Watch the first full billing cycle the way you would watch a spiking error rate. Treat a spend spike as a signal that something in the workflow is misbehaving. The preview bill and budget controls GitHub shipped exist for exactly this.
The Catch Nobody Advertises
Even after setting budgets and routing tasks by cost, predicting a given week’s spend remains hard. An agent’s token consumption depends on the shape of the problem it encounters mid-run. A refactor that looks small can cascade into a dozen files once the agent discovers a tangle, and the cost balloons before any budget alert fires.
Per-task estimates can still be off by a wide margin on the messy, exploratory work that agents are most useful for. Metered pricing made the cost visible. It did not make it predictable. Those are not the same thing.
The One Question Worth Asking Your Team Today
If your most enthusiastic agent user burned a month of credits in nine days, would you know on day two or day nine? If the honest answer is day nine, you have the exact blind spot that turned a routine billing change into a category-wide backlash this month.
The flat subscription is gone. Scope the agent, route by cost, and watch the meter before the meter watches you.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
