Skip to main content
Routing Rules are CEL-based conditions that intercept requests to the AI Gateway and redirect them to a different model when matched. Rules are evaluated in priority order and the first matching rule wins. No further rules are evaluated after a match.

​
Use cases

Routing rules are most useful when traffic needs to be redirected or distributed at the gateway level based on request attributes, without modifying any calling application.
Intercepts support bot requests addressed to one model and silently redirects them to a cheaper one, so the calling application needs no changes to benefit from the cost saving.
Two rules at the same priority split document traffic by complexity: simple documents go to a lighter model, complex ones go to a more capable model, all based on metadata attached to each request.
Matches requests that include a file attachment and routes them through an ordered list of models that all support native file input, so the request succeeds even if the first model is unavailable.
A low-priority catch-all rule that redirects traffic to an alternative set of models whenever primary providers are unavailable, keeping requests flowing without manual intervention.
Distributes requests in round-robin mode across multiple endpoints serving the same model, balancing inference load across providers without any changes required on the caller side.

​
How routing rules work

When a request arrives at the AI Gateway, the AI Gateway evaluates all active routing rules in ascending priority order. The first rule whose CEL expression matches the request determines the target model. If no rule matches, the model from the original request payload is used. Example: A request arrives with model: "openai/gpt-5.6-sol". A routing rule with condition model.contains("gpt-5.6-sol") and target openai/gpt-5.4-mini is the highest-priority matching rule. The AI Gateway redirects the request to gpt-5.4-mini, regardless of what the caller specified. In CEL expressions, model refers to the model value from the request payload.

See also: Request-level load balancing

To distribute traffic across providers at the request level without organization-wide rules, use the load_balancer parameter directly in your API calls.

​
Visibility

  • Visible to workspace administrators only.

​
Creating a routing rule

From the Routing Rules list, click Routing Rule. The form opens as a full page with a back control and an unsaved-changes confirmation. It carries five sections, and the submit button reads Create Rule, or Save Changes once the rule exists.

​
General

​
Load Balancer

Optional. Defines the target model or models to route matching requests to. Leave this section empty to build a rule that exists only to apply a plugin or a cache to matched traffic, with no model routing at all. Load balancing strategy sets how the targets are used:
  • Fallback: Route to the primary model. Try the next model when it fails with status 408, 429, 500, 502, 503, 504, or 529 by default. Once the rule has a target model, Fallback status codes appears and overrides this list; leave it empty to use the default. The API field is models_config.fallback_on_codes, accepting up to 20 unique integers from 400 through 599. This field is available only in Fallback mode.
  • Latency: Route to the model with the lowest recently observed latency. See Latency-based routing for the full selection algorithm.
  • Weighted: Split traffic across models by percentage weights.
  • Round Robin: Rotate evenly across all configured models.
Choose Latency when minimizing response time matters more than a fixed traffic split. Choose Weighted or Round Robin when the split itself, such as cost control or A/B testing, is the goal. Click Models to add a target model. The menu groups models by provider. Each target has a remove control. Fallback also gives each target a drag handle for reordering, and Weighted adds a Traffic weight slider to set that model’s share, which must add up to 100%.
When a request matches this routing rule, its Load Balancer configuration completely replaces the request’s own model, load_balancer, fallbacks, and fallback_on_codes values; they are never merged. A request-level load_balancer parameter has no effect once a matching routing rule with configured models applies.

​
Cache

Optional. Give the rule its own response cache for matching requests. Exact-match requests reuse a cached response instead of hitting a model. Leaving this section disabled adds no cache of its own, so matched requests keep using their request-level cache setting. Set Enable cache to turn the cache on, then choose Time to live to control how long matching responses stay cached, from 5 minutes up to 3 days. The default is 1 hour. When a rule has cache enabled, a matching request uses the rule’s cache configuration instead of the request’s own cache parameter; the two are not merged. See LLM response cache for how caching works.

​
Plugins

Optional. Attach plugins that run on traffic matching this rule. Plugins can transform the request, the response, or stored traces. Click Plugin to attach one of the following:

​
Conditions

Define when this routing rule should apply to a request. The Rule Builder builds the CEL expression shown read-only in the CEL Expression Preview; leaving it empty applies the rule to every request. Select Add rule to choose a condition type: Each Add rule creates a condition group titled Rule 1 · N conditions. Inside a group, select Condition to add another condition, and to remove one. Drag a condition by its handle to reorder it. Conditions in a group are joined with and by default; select the AND pill between two conditions to switch the group to or. Delete removes the whole group. Rule groups are joined with or by default; select the pill between two groups to switch them to and.

​
Editing and deleting a routing rule

Click a rule’s name in the Routing Rules list to open it in the same form. An existing rule exposes Delete Rule in the form footer, next to Save Changes. Deleting asks for confirmation first. The list’s row actions menu offers Edit, Enable or Disable, and Delete.