On June 1, GitHub flipped the switch.
GitHub Copilot moved from a flat monthly fee to usage-based billing: every token you consume, input and output and cached, now runs through a meter. The base subscription prices did not change. What changed is that the ceiling disappeared.
Developers noticed immediately. Reddit threads filled up with reported cost jumps from $29 to $750 per month. From $50 to $3,000. The official announcement drew nearly 900 downvotes. Someone on X called it "a joke." GitHub's community thread ran past 400 comments inside 48 hours.
The anger is understandable. The diagnosis is wrong.
How it worked before
The old model gave you a monthly allotment of Premium Request Units. When you burned through them, Copilot gracefully fell back to a lighter base model. You could keep working. The ceiling held.
That design was a product decision, not an infrastructure reality. Running Claude Opus 4.8 or GPT-5 at frontier quality, across agentic loops, across full-file context windows, across every autocomplete suggestion in your IDE, costs real money. It cost real money before June 1 too. GitHub was eating the difference.
The official rationale, almost word for word from the announcement: "Copilot is not the same product it was a year ago." Which is true. A year ago you weren't running agentic workflows through it. A year ago it wasn't routing your requests to frontier models on demand.
GitHub's version of that sentence frames it as a product growing up. The more accurate version: the subsidy model finally buckled under the weight of what users had been encouraged to do with it.
The subsidy was the strategy
This is not unique to GitHub.
OpenAI reportedly spends roughly $1.35 for every dollar it earns, losses driven primarily by inference costs. Google moved its AI offerings toward metered consumption. Across the industry, today's token prices are subsidised by venture capital and hyperscaler cross-subsidy, buying adoption, buying habit, buying the kind of dependency that makes switching feel impossible.
That is not a conspiracy. It is a standard customer acquisition playbook. You price below cost to build the installed base, then you adjust once the base is locked in.
GitHub told you to go agentic. They ran campaigns around it. The docs leaned into it. The prompts encouraged it. And the economics of doing so, at the flat rate you were paying, never made sense for them, only for you.
June 1 is not the rug-pull. The rug-pull was the original pricing. The generous flat rate was the sample. You were being acquired.
The coupon test
Here is the question worth sitting with.
Does your AI-assisted workflow survive at unsubsidised prices?
Not "does it feel slower without Copilot." Not "would I miss the autocomplete." The actual economic question: if the tool costs what the infrastructure actually costs to run, does the workflow still justify itself?
If 10x pricing kills your workflow, then your workflow was a habit the discount paid for. Not a productivity gain. Not a compounding capability. A heavily subsidised habit that looked like signal because it was cheap enough to ignore the noise.
The developers reporting $750 months are not necessarily doing the wrong things. Some of them are running agentic loops that genuinely deliver value. But some of them, maybe many of them, accumulated token consumption the way you accumulate browser tabs: out of friction-free availability, not deliberate intent.
Tokenmaxxing did not require the meter to be running. It just waited for the bill to arrive.
This is an infrastructure dependency problem
The billing change is the visible symptom. The structural problem is older.
When your development workflow runs on someone else's frontier model, priced at someone else's discretion, you have built a dependency that you cannot audit, cannot cap without degrading capability, and cannot exit without rewriting how you work. That is vendor lock-in at the workflow layer, which is harder to escape than vendor lock-in at the API layer.
The token-saver tax post made a version of this point about plugin overhead: the cost is guaranteed and front-loaded, the savings are conditional and back-loaded. Usage-based Copilot billing is the same shape at larger scale.
The consumption is guaranteed. The value is conditional. And when the prices move, you find out which was which.
Code completions and Next Edit Suggestions, the original core features, remain unlimited and unmetered. That is not generosity. That is GitHub protecting the habit-forming layer while monetising the power-use layer.
The features that feel like magic stay free. The features that run up actual infrastructure bills now run up actual bills.
What to actually do
Set a spending cap. Today. The billing dashboard has one. Without it you have no ceiling, and "I forgot to set a limit" is now an expensive mistake, not a billing curiosity.
Measure value, not usage. Does agentic Copilot actually change how fast things ship? Does it change defect rates? If you cannot point to a number moving in the right direction, you are paying for a feeling.
Keep an exit. Local models have become genuinely capable for a large slice of everyday coding tasks. The completions tier remains unmetered. A workflow that only functions at frontier model quality, at unlimited scale, at someone else's pricing, is a workflow you do not control.
Price the workflow honestly. Take the new monthly bill, actual usage, and ask what outcome it is purchasing. Not "is this useful in general." The specific number, against specific evidence that it changes specific outcomes.
The meter is not the problem
The developers angry at GitHub are angry at the wrong thing.
The bill is not the surprise. The surprise is what the bill reveals: that two years of subsidised pricing built habits at a scale that the underlying economics never supported. That "go agentic" was advice given to you by the people who were absorbing the cost of the advice. That a product being cheap and a product being worth it are not the same claim.
The meter switched on June 1. What it is measuring was always running.
Now you get to find out what your workflow was actually worth all along.