On June 1, 2026, GitHub Copilot stopped being a subscription and became a meter.
The flat monthly fee that 4.7 million developers had budgeted for since 2021 was replaced with token-based billing. Every input, every output, every cached token in an agentic session now costs something. And the early numbers from developers who actually ran agentic workflows all day were not pretty.
One developer's monthly bill was projected to jump from $29 to $750. A three-person team running agent mode against a large monorepo estimated their combined cost rising from $117 to somewhere between $1,800 and $3,600 per month. Another user on Reddit reported burning through 8% of their monthly Pro+ allotment in two hours, which put them on track to exhaust the entire quota in under two days.
That last one is the most interesting data point. Not the dramatic projection. The quiet, accidental overage that happens mid-morning on a Tuesday while doing actual work.
What Actually Changed
The old model was simple. Copilot Pro cost $10/month and Copilot Pro+ cost $39/month. Both gave you a fixed allowance of premium requests, and when you ran out, Copilot fell back to a cheaper model so you could keep working. The ceiling was predictable. The floor was always there.
The new model replaced premium request units with GitHub AI Credits. One credit equals $0.01. Credits are consumed by every token an interaction processes, priced according to the listed API rates for each model. Code completions and Next Edit Suggestions are still unlimited. Everything else is metered.
Copilot Pro subscribers now get $10 worth of credits per month. Pro+ subscribers get $39 worth. A new $100/month Copilot Max plan exists for heavy users, bundling $200 in total credits ($100 base plus a $100 flex allotment).
The fallback model is gone. When credits run out, premium features stop. You can buy more, or wait until the next billing cycle.
There is also a subtle change that annual plan subscribers got hit hardest by. The multiplier for Anthropic's Opus 4.7 jumped from 7.5x to 27x for annual plan users, which effectively wiped out any savings those users had locked in.
Why This Happened
GitHub was losing money at scale. That part is not speculation. Internal documents reported by Ed Zitron's newsletter indicated that Copilot's operational costs were nearly doubling week-over-week since January 2026. The billing change was described internally as a financial emergency rather than a planned product evolution.
The reason the costs were spiraling is the same reason the bills are shocking developers now: agentic workflows changed the economics completely. When Copilot was an autocomplete tool that suggested the next line of code, the token cost per interaction was tiny. When it became an agent running multi-file refactors, writing test suites, reviewing pull requests, and running autonomously for hours against large codebases, the compute cost per developer per day multiplied by somewhere between 10 and 50.
GitHub built its pricing on the assumption that most developers would use Copilot like autocomplete. A significant chunk of them started using it like a junior developer who works all night. The economics of those two use cases are not even close.
This is also not the first structural pricing change in twelve months. Premium request billing was only introduced in June 2025, capping Pro users at 300 monthly premium requests with overages at $0.04 each. The June 2026 change is the second major adjustment in under a year.
The Reddit Reaction
The official GitHub community discussion thread got 904 downvotes against 22 upvotes across 435 comments. That is a historically lopsided ratio for a GitHub forum post.
The reaction on Reddit was consistent. "What a joke" was probably the cleanest summary. Users posted screenshots of projected bills, complained that GitHub had trained them to use the tool aggressively and was now penalizing them for it. One comment that kept getting upvoted: "The only one at fault here is Microsoft. They provided this billing method and kept making it easier and easier to burn through massive numbers of tokens."
There was also a counterargument that got some traction. A contingent of developers pushed back on the most extreme projections, arguing that users racking up $3,000/month bills were using Copilot badly: accepting bloated context, not cleaning up iteration loops, essentially vibe-coding without understanding what they were sending to the model. "The vast difference between some of us working all day and still barely having overage and then these screenshots. I struggle to believe it's complexity differences in the workload," one commenter wrote.
That argument has some validity. But it also misses something. GitHub spent three years marketing Copilot as a tool that handles the heavy lifting. The whole pitch was that you could offload thinking to it. Developers who bought that pitch and built workflows around it are now being told, implicitly, that those workflows were wrong.
One developer put it plainly: a session using Claude 4.8 to fix website issues consumed 1,180 credits, about 16% of their monthly Pro+ allowance, for results they described as only mediocre. The spend was hard to predict before it happened, and the output did not justify it after.
The Broader Pattern
GitHub Copilot is not the only product doing this.
OpenAI's ChatGPT has been moving toward usage-based billing for heavy users. Google's Gemini Spark, the persistent agentic version, requires the $100/month AI Ultra tier. The entire consumer AI subscription model is going through the same recalibration for the same reason: flat-rate pricing worked when usage was bounded by how fast a human could type. It breaks when an agent can burn through thousands of tokens autonomously in the background while the developer makes coffee.
The "gym membership" model of AI subscriptions was always subsidized. Pay $10-$39/month, use it as much as you want. That worked while usage stayed within a range the company could absorb. Agentic workflows blew past that range, and the subsidy became unsustainable.
This does not mean the tools are getting worse. It means the free lunch is over.
What Developers Are Actually Doing
The migration to alternatives started before June 1 and accelerated after.
Cursor, at $20/month, is still flat-rate. So is Windsurf at $15-$20/month. Both have been the primary beneficiaries of developers looking for predictable costs. Cursor's Supermaven autocomplete and Composer mode cover most of what Copilot's agentic features do, and the per-month price has not changed.
Cline, the open-source VS Code extension with bring-your-own-API-key, gives developers direct API billing from Anthropic, OpenAI, or Google. No markup, no monthly subscription premium, full visibility into what each session costs. For developers who can optimize their context and model selection, this approach can be cheaper than any subscription once you account for agentic overhead.
Claude Code at $17-$100/month offers flat costs within plan limits, with no per-token overage as long as you stay inside the session window. For developers who need extended, deep codebase sessions, the predictability is the selling point.
The common thread in all of these alternatives: you know what you are paying before you pay it.
The Harder Question
The billing change is the surface story. The more uncomfortable question underneath it is whether agentic AI development is actually generating returns that justify the cost at current prices.
The $10/month Copilot subscription was easy to justify. Even a modest productivity boost covered the cost in the first hour of the month. At $150-$300/month for a power user, the math requires a harder argument. You need to demonstrate that your output is genuinely multiplied, not just that the tool is busy generating things.
Some teams can make that argument. Enterprise teams running automated PR review pipelines, test generation at scale, and documentation workflows across large codebases have real productivity numbers to point to. The agentic cost is justified by the output volume.
Individual developers and small teams doing less systematic AI work are in a different position. If you are using agent mode to explore an idea, iterate on UI components, or debug something you half-understand, the token costs accumulate without predictable output. That is where the Reddit horror stories come from. Not from irresponsibility, but from using a powerful tool experimentally without a feedback loop on cost.
The billing change is forcing developers to be deliberate about what they ask AI to do. That is probably good, in the long run. But it is also an admission that the hyper-growth phase of AI-assisted development, where every developer could run agents all day on a flat fee, is over.
The tools got better. The bills caught up.
If You Are Still on Copilot
A few things worth knowing:
GitHub is offering promotional credits for June, July, and August to ease the transition. If you are on an annual plan, you stay on the old request-based pricing until your plan expires, at which point you roll off to Copilot Free or choose a monthly plan.
Organizations get pooled credit usage, so unused credits from lighter users in your team offset heavier users. For companies with mixed usage patterns, this can soften the impact.
You can set budget caps to prevent runaway costs. By default, all budgets are set to zero and premium requests above the plan allowance are rejected rather than billed. You have to explicitly enable paid overage.
For individual developers doing mostly code completions and Next Edit Suggestions with occasional chat, the new model is probably not much more expensive than the old one. Those features stay unlimited. The pain is concentrated in agentic workflows, long context sessions, and intensive model usage.
Know what you are actually using Copilot for before deciding whether to stay or switch.
