On May 5, in a post titled “We’re Not Building AI Features for the Money,” Zed wrote a sentence almost nobody else in this market will write: until it switched to token-based billing, it had been “selling tokens at a loss to most paying AI users.” It now passes LLM usage through “at the provider’s list price plus a 10% markup.” I want to sit on that sentence, because it is the most honest thing said about AI coding tool economics this year, and it closes the loop on everything else that happened this week.
What the sentence admits
There are two admissions, and the second is the load-bearing one. First: the prior plan lost money on heavy users. Second, by implication: most of the market still works that way and is not saying so. A flat $20 or $30 monthly subscription, set against an agent that can burn tokens for hours at a stretch, is a subsidy with a countdown on it. It is priced to acquire you, not to be sustainable, and the gap between those two prices is venture money making a bet that the number can be raised after the habit forms.
Put rough numbers on it. A developer running an agent through real multi-hour sessions can move tens of millions of tokens in a month; at provider list prices that is not a $20 cost line, it is a multiple of it. The flat plan does not make that cost disappear — it moves it onto the vendor’s balance sheet and bets that the average across all subscribers stays low enough, long enough, to raise prices before the heavy users sink the plan. You are not the customer in that model so much as the position the vendor is holding until the market clears. It is a comfortable position right up until the moment the vendor needs it not to be.
The 10% is the tell
The honest part is not that Zed charges. It is the structure. “Provider list price plus a named 10%” is a number you can verify against the provider’s own published rates and reason about before you commit. A flat subscription with an undisclosed token ceiling is not a price you can reason about; it is a number that works until you become the kind of user who makes it not work, at which point it changes. A margin you can see is a contract. A subsidy you cannot see is a countdown someone else is holding the timer on. Zed converted the second into the first, in public.
Why this closes this week’s loop
Everything else I wrote this week pushes token consumption up while treating the bill as someone else’s problem. Codex on your phone removes the friction that used to cap how often you reach for an agent. The 60% number is, read literally, a statement that the majority of an organization’s code now arrives through paid model calls. Hardening the supply chain adds its own overhead: more runs, more tooling, more review cycles that themselves invoke models. Every trend this week increases the number of tokens a serious developer burns per day, and almost every vendor is still pricing as though that number were flat. Zed is the one that put the meter where the customer can see it.
Stack the three together and the direction is unambiguous: lower friction to invoke, more of the codebase routed through the model, and a security posture that itself runs on extra model calls. None of those trends bend back down. A serious developer’s daily token burn a year from now is not going to look like this year’s, and the plans that read cheapest today are the ones that have priced in the least of that curve. The honest move is to price the curve, not the snapshot; almost nobody does, because the snapshot is what wins the comparison table.
Pass-through is where this ends
The stable end state is the one Zed jumped to early: usage passed through near cost, a margin that is named rather than hidden, and likely a per-seat platform fee for the collaboration layer on top. The “unlimited” plans on metered infrastructure are the ones to distrust, because unlimited is not a price — it is a bet that you will not use the thing you were sold, and that bet gets repriced the moment enough people use it. The tools converging toward transparent metered pricing — the trend I traced in AI tool pricing convergence — are not getting more expensive. They are getting honest, and honesty looks like a price increase only because the old number was fiction.
The steelman, which I mostly accept
Subsidizing acquisition is not a scandal. Every infrastructure category did it; cheap early pricing is how a new tool reaches enough hands to be worth building for, and recovering margin later is an ordinary business move, not a bait-and-switch. I accept most of that. The objection is narrow, and it is the one Zed’s post implicitly makes: the subsidy is defensible, the silence about it is not. A developer making a two-year bet on a tool’s pricing is entitled to know whether today’s number is the real one or a customer-acquisition loss with a raise already priced in. Zed told them. The developers who will be annoyed in two years are not the ones on Zed’s transparent meter; they are the ones who built a workflow on a number that was never real and got repriced without ever being told it was a subsidy. The question worth ending on is not why Zed disclosed it. It is why a disclosure this basic is rare enough to be remarkable — and what that rarity tells you about every plan that reads cheaper than Zed’s and will not explain how it can.
