April's price trend inversion, with a frontier launch arriving more expensive per token, hit directly anyone who had planned on the continuous decline of the previous two years.
Product planning with AI used to embed a comfortable premise: unit cost would fall over time, so an application unviable today would become viable next year with no engineering effort.
With the premise inverted, viability returns to depending on work: reducing what you send the model, using a smaller model for the predictable part of the load, caching repeated results instead of recomputing.
That's the same set of practices large companies adopted in the period under another name, announcing cost-based routing between models and cheap triage layers before the expensive model.
The good news for small operations is that these techniques require neither scale nor special contracts. They require measuring where spend concentrates, which is usually a handful of repeated calls nobody has reviewed since they first worked.