Most of the agent cost threads here end up arguing about frameworks, but in practice the bill is set by two decisions that get made in the first hour and never looked at again. Sharing the numbers because they are easy to check against your own setup.
Model Choice Per Step, Not Per Agent
A yes-or-no classification on a small model runs about 1 credit per decision, roughly a tenth of a cent. Handing that same call to a premium reasoning model costs 15 to 30 credits and returns the same answer. Extraction and short generation sit in the middle at around 4 credits.
The mistake is picking one model for the whole agent. Classification, extraction, generation and reasoning have genuinely different requirements, and the cheap steps usually outnumber the expensive ones by a lot. There is a breakdown of the per-model cost per call and how to work out an agent monthly spend if you want the actual numbers, and a companion piece on matching models to task types.
Chain Small Agents Instead Of One Big One
The second decision is structure. One monolithic agent that researches, decides and acts has to run every step at the capability level of its hardest step. Split it into a chain and each stage runs on the cheapest model that can do that stage.
Sequential pipelines, fan-out, aggregation and conditional branches all fall out of this, and failure handling gets easier too, because you can retry one stage instead of replaying the whole run. The patterns for chaining agents and passing data between them cover the four common shapes.
What It Adds Up To
Once both are deliberate, most business agents land between three and fifty dollars a month depending on volume. The agents that blow past that are almost always a single large agent doing cheap work on an expensive model.
Worth auditing your own chain for the step that is quietly overprovisioned. It is usually the classifier.
