Price and performance
Guardrails are classification. A small trained classifier is the production tool, not a chat model asked to judge every request.
Why a classifier
Prompting a generative model to moderate or route text can be a useful probe when a policy is new and you have no examples yet. It is the wrong production default. Every request pays for token generation, waits on a completion, and can change its mind on the same input.
A purpose-built classifier, typically a small encoder-style model, scores labels in one pass. That is cheaper, faster, and more stable once the policy is real enough to train on.
What you actually buy
- Latency. A classification forward pass, not a generated verdict.
- Cost. You pay for items classified, not for pages of chain-of-thought.
- Stability. The same string should not flip labels because a decoder sampled a different token.
A prompted decoder can still be the right first look at a brand-new policy. It is not "always worse." It is the wrong long-term engine for a guardrail you will call on every request.
What to do next
Call a managed model today, or train a guardrail from a policy. Pricing is usage-based either way.