How much does it cost to run a classifier on every trace?

Nothing, in platform terms. Tessary is open source and free to self-host, so there is no per-trace rate and no separate meter for classification. Every trace is checked, and one trace is one agent turn: the user’s message, your agent’s full response, and every tool call in between.

What you do pay for is model-provider spend. The five classifiers that run by default make no model call. The expensive LLM judgment only reads the findings they write, which is a small slice of traffic, so your bill tracks how often the cheap checks fire rather than how much you serve.

Frustration is billed to your own OpenRouter or TypeSafe key, per message it scores: about $0.04 per 1,000 messages by the product’s own estimate. It skips a conversation’s first two user messages and stops at the first flag, so not every message is a call.

Groundedness costs machine time instead. Its model runs on a GPU you provide. On a Mac with Apple silicon nothing is billed. On AWS you pay for the instance while it runs, a disk billed all month, and one stored secret, and switching the classifier off doesn’t stop those charges: deleting the model’s stack does.