Unit Price Is Falling—That's the Trend
Over the recent period, the per-token price of model APIs has broadly been in a downward channel: more efficient architectures, fiercer competition, plus distilled small models taking on large amounts of simple tasks, pushing the cost of "running one call" ever lower. For developers, this is good news—many scenarios that were once not worth it because they were too expensive are becoming viable.
But the Bill Doesn't Necessarily Fall With It
A recurring reality: unit price drops, yet the total bill doesn't necessarily fall, because usage is rising, context is getting longer, and agent-type applications go many rounds per task. What really keeps cost in check is usually not waiting for the vendor to cut prices, but engineering means—handing stable long prefixes to Prompt Caching for reuse, doing model-tiered routing by task difficulty, giving simple work to cheap models, and using a unified entry point to attribute usage to specific applications and teams.
Knowing How to Use It Matters More Than Cheap
The conclusion is plain: price cuts lower the floor, but whether you actually save depends on whether you know how to use it. See your usage clearly first (who's spending, on which model, and is it worth it), then apply caching, routing, and tiering—this usually beats simply eyeing the price list at bringing the bill down. The cheaper the model, the easier waste is to overlook—so fine-grained cost governance should actually move up the priority list.
via: Synthesis of vendor pricing and public cost practices