What happened
GitHub Copilot — the coding assistant most engineering teams use — bills by usage as of 1 June 2026. Instead of counting "premium requests," each plan gets a monthly pool of AI credits charged by token count: input, output and cache, at the API rates of the given model. One credit is USD 0.01. When the pool runs out, Copilot stops — it no longer quietly switches to a weaker model — unless you deliberately enable additional spend. For the transition period, from June to September 2026, GitHub adds a temporary "flex" bonus; from October it shrinks toward the base pool. Code completion itself and edit suggestions remain unlimited and don't touch credits.
This isn't an ordinary price increase. It's a change in the billing model: the cost of the tool has stopped being a fixed line in a subscription and has become a variable cost that rises with how much the team actually generates.
Our thesis
It's not worth reading this as "Copilot got more expensive," because for some teams the bill won't change. It's worth reading it as a signal of a broader move: AI assistance in tools is shifting from a "per seat" model to "per usage" — and, we think, the same awaits further software categories (that's our forecast, not the vendors' announcement). Early reports after 1 June speak of bills several times higher for people who heavily use agentic modes — and that's the real lesson. The cost didn't jump for everyone equally; it jumped where usage was large and invisible.
The operational takeaway is simple: the surprise bill in the autumn will go not to the company that uses Copilot the most, but to the one that doesn't know who uses it and how much. Usage-based billing penalizes a lack of visibility, not intensity itself.
Why it matters
Private Equity
For a fund, this is the same vendor dependency we described in the note on AI due diligence, only now on the cost side, not the legal-risk side. If a portfolio company wrote a fixed cost of AI tools into its margin thesis, and those tools move to usage, part of the forecast has just become variable. It's worth flagging in the value map as an item to watch, not to ignore.
Enterprise
For a large organization, this is the moment for FinOps for AI — before the invoice comes, not after it. Three things worth having before the flex shrinks: visibility of usage broken down by team, a default limit that halts work instead of generating open-ended spend, and a decision on which models for which tasks — the more expensive one for the hard problem, the cheaper one for the rest. That's work at the layer of policy and cost architecture; we help put it in order as part of our services.
SMB / mid-market
For a smaller team the change can be felt sooner, because one intensive week can drain the pool. The good news: the default behavior is safe — without manually enabling additional spend, the worst that happens is a pause until the next month, not a bill from outer space. It's worth taking advantage of that and leaving the limit where it is.
One step you can take this week
Go into Copilot's billing settings — or those of any AI tool that charges by usage — and do two things: set a budget, even a zero one, meaning "stop at the enabled pool," and pull last month's usage broken down by people or teams. The first protects against open-ended spend; the second tells you whether September will be calm or not.
Describe your case
If you pay for some AI tool by usage and don't know how to set limits or how to allocate the cost across teams, bring one bill and a list of tools. We start from something concrete. Describe your case: mailto:[email protected]?subject=Rozmowa%20z%20Aurora%20AI.