Why does a classifier cost less to run than an LLM judge?

Because one check does less work. A classifier tests a narrow property with a rule or a small trained model; an LLM judge reads the trace and generates a verdict. Training the classifier happens once, but running it on each new trace still uses compute. Hosted inference is billed while it runs, even when the per-trace cost is small.

A classifier can run across the stream cheaply, but it only checks the property its rule or training defined. An LLM judge costs more per event and can judge whether an answer’s reasoning holds up.