What got cheap is last quarter's frontier

OpenAI cut prices this week, hard. Eighty percent off the small model, twenty off the mid tier, per the announcement. The line people latched onto came from someone doing the arithmetic in public: a model scoring where March’s flagship scored now sells for about a thirteenth of what that flagship cost, roughly four months later.

Read the second half of that sentence twice. What collapsed in price is not the frontier. It is March’s frontier.

That distinction decides what you should actually do about it. If you priced a workflow six months ago and shelved it because the token bill did not work, the bill is a different number now and the shelf deserves another look. Running a checker over every generated file. Summarizing every ticket that comes in. Work that was absurd at flagship prices and is merely unglamorous at a fifth of them.

But if the job genuinely needs the current frontier, the judgment call, the long horizon, the hard step in the middle, you are still paying frontier prices. That number has not moved much. The cheap lane is capability that has aged a few months.

The newsletter reporting all this adds a caveat I appreciate: public benchmarks get trained on to some degree, so one score matching another score is softer evidence than it looks. OpenAI also says part of the savings came from the model optimizing its own serving stack and rewriting production kernels, which reads as impressive or circular depending on your mood that day.

The vault’s notes on the economics keep landing on the same question, and it is not which model is best. It is where the value is large enough to justify the cost and the operational risk. A price cut does not answer that. It moves the line and hands the question straight back to you.