# Guardrail latency

Guardrail latency is the time a guardrail check adds to each request. It is paid on every input, output and tool call, so it decides whether checking in real time is practical for a chat or an agent.

## Where it comes from

A check that can block must finish first: the [input check](/glossary/input-and-output-guardrails) before the model call, the output check before the user sees the answer, the [tool call check](/glossary/tool-call-guardrails) before the tool runs. Each costs a network round trip plus the time the checking model takes.

## Why it matters

- **Chat.** The user feels the wait before the first word, and an input check adds to it directly. With streaming, an output check must hold text back or check it as it is shown.
- **Agents.** One task can take dozens of model calls and tool calls. Each is checked, so the delays add up.

A check too slow for the request path gets run after the fact, where it can report a violation but not stop it.

## How to keep it low

- **Small models.** A [guard model](/glossary/guard-models) trained for the one job answers much faster than a general model asked to judge.
- **Parallel checks.** Run every policy on a message at once, not in sequence. An input check can also run alongside the model call, with the answer held until it passes.

## With White Circle

[Control](/products/control) analyzes every input and output in real time. The overhead is minimal: under 100 ms on average.

