Aug 2026 · The Watch
The free lunch is over.
There’s a reason you could ignore AI costs until recently, and it wasn’t laziness. It was rational. That reason just expired, and most teams haven’t noticed the strategy change.
The old strategy: wait.
Drew Breunig put the expiring era better than anyone:
“Prior to Fable, it felt silly to waste too much time improving your coding harness or context strategies. A new model would arrive at the same price (or cheaper!) and paper over most of your problems.”
That was genuinely good advice for years. Why spend a week engineering your prompts and context when the next release would be smarter and cost the same or less? Every quarter, the hardware of the situation improved underneath you. Optimization was work you could safely defer, because time was doing it for free.
What broke it.
The newest tier of models is genuinely excellent and genuinely expensive — expensive enough that “just use the best one for everything” stopped being a defensible default. Meanwhile the previous generation didn’t get worse. It stayed good enough for most work, at a fraction of the price.
So the curve that used to bail you out flattened. The next release is no longer guaranteed to be cheaper and better. Now capability and price can move in the same direction at once, which means the decision you were deferring — what work goes where — is finally yours to make.
Scruuge’s TLDR
You used to be able to wait out your AI cost problem. That was a real strategy, and it worked. It doesn’t any more. The work you were rationally deferring — routing, context discipline, cleaning the desk — just became the work.
The market already moved.
This isn’t theoretical. Billing data across roughly 70,000 companies shows the newest frontier model capturing only a few percent of spend a month after release, while the older, cheaper generation holds the overwhelming majority — the receipts are here. Teams are already choosing what goes where. The ones doing it deliberately are getting a much better deal than the ones doing it by accident.
What “doing the work” actually means.
It’s less than you fear, and it’s all things we’ve given away already:
- Route by task — best model to plan, good-enough to execute (the 68% piece).
- Keep the context clean — most of your bill is re-sent conversation, not what you typed (reused input).
- Compile the repeats — stop paying a model to re-derive the same procedure (compiled AI).
- Then check the ceiling — cutting cost is the floor; where you point the frontier is the actual strategy (floor vs ceiling).
The whole set lives in one place: the Cost Engineering Playbook. Free, as always.
The honest caveat
Prices could fall again — they have before, sometimes dramatically, and a serving breakthrough can reset the floor for everyone overnight. So “the free lunch is over” is a statement about today’s decision, not a permanent law. The difference is that betting on the next price cut is now a bet, where it used to be a near-certainty. Engineer for the bill you have.
Stop waiting. Start measuring.
Two minutes with the calculator tells you whether you’re on the wrong tier right now. The $999 Assessment does the routing work across your whole setup — the work that used to be safe to defer.
The quoted passage is Drew Breunig’s, surfaced via Simon Willison. Model-share figures from the Ramp AI Index (July 2026). Scruuge’s contribution is the strategic read — that a rational wait-and-see posture just stopped being rational — and the caveat that it could change again.