Part of my job is to look out ahead. When everyone in AI is saying the same thing at the same time, that’s usually the signal to stop asking whether the consensus is right (it often is) and start asking where the advantage goes once everyone has acted on it. Right now everyone is saying the same thing: route to cheaper models. Fable 5 was the spark that lit this fire. At $10 per million input tokens and $50 per million output, double Opus 4.8, the newest frontier model showed up with a price that turned every leadership meeting into the same meeting, and now that meeting is spreading everywhere: do we have to pay for this, or can we route around it?
The answer to that question is arithmetic, and the arithmetic is finished. What interests me is the question forming behind it, the one almost nobody has started asking: once every company on earth has the same routing discipline (and that day is close, because the playbook is public and the consultants all sell the same one), where does advantage come from?
I think the answer is imagination. That word does a lot of lazy work in business writing, so let me pin it down: two specific, buildable capacities that determine the return on every model you run, cheap or frontier. One of them showed up in public a few weeks ago, attached to a $40 receipt.
This briefing covers:
Why a $1 model tied a $9 one — and why that’s a fact about the task, not the models. The convergence everyone is celebrating is the same convergence that locks you into everyone else’s results at everyone else’s price.
The $40 job no routing table would ever have assigned. Mitchell Hashimoto ran a second experiment almost nobody noticed, and it shows where advantage forms once execution goes cheap.
The two-layer setup — the cheapest capable model as your engine, frontier models as your steering. Optimize the engine hard, open-weights where you can so no vendor owns your costs. The leverage is in the steering, and it compounds as the engine gets cheaper.
A one-question diagnostic you can run on your own operation. It takes ten seconds to answer, and it names the real constraint on the returns you get from AI — which turns out not to be the price of any model.
Let me start with that receipt, and work back to what it’s really evidence of.
