> ## Documentation Index
> Fetch the complete documentation index at: https://to11.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Weighted Routing

> Split LLM traffic across multiple targets using weighted random selection.

To distribute traffic across several targets for the same model, create a routing rule with the **Weighted split** strategy. For each request, to11 picks a target at random, with the probability of each target proportional to its weight.

## Use cases

* **Cost control** — send most traffic to a cheaper target and the rest to a premium one.
* **Gradual migration** — shift a small share of traffic to a new provider and raise it over time.
* **Load distribution** — spread requests across multiple credentials for the same model to ease rate limits.

## Before you start

* Two or more targets for the same model — different providers, or different stored credentials for the same provider.
* Each provider connected under **Project → AI Gateway → Providers**, with the model in its model list. See [Connecting a provider](/docs/deploy/concepts/connect-provider-drawer).

## Create a Weighted split rule

1. Open **Project → AI Gateway → Routing**.
2. Start a new rule with the create-rule flow.
3. Choose the **Weighted split** strategy.
4. Set the model the rule applies to (for example `gpt-4o`).
5. Add each target — its provider and stored credential — and give it a weight as a percentage. For a 70/30 split, set the primary to 70 and the secondary to 30.
6. Save the rule.

## How weights work

Selection is random, proportional to the weights. A target weighted 70 receives roughly 70% of requests and one weighted 30 receives roughly 30% — a 70/30 split behaves the same as 7/3. The split is statistical, so it converges on the configured proportions over a large number of requests rather than alternating on every call. If you leave weights equal, traffic is divided evenly.

## Send a request

Call the model by name, exactly as you would for a single target. to11 selects a target per request:

```bash theme={null}
curl https://gw.to11.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "x-to11-authorization: Bearer $TO11_API_KEY" \
  -H "x-to11-project-id: $TO11_PROJECT_ID" \
  -d '{
    "model": "gpt-4o",
    "messages": [{"role": "user", "content": "Hello"}]
  }'
```

To see which target served each request, open **Projects → Traces** in [Observe](/docs/observe/traces). Each trace records the resolved target and the strategy that selected it, so you can confirm the distribution over time.

## Limitations

<Warning>
  Weighted split is not health-aware. A target that is returning errors keeps receiving its proportional share of traffic. When you need automatic failover, use a [Fallback chain](/docs/deploy/routing/fallback) instead, or combine the two.
</Warning>

## Next steps

<CardGroup cols={2}>
  <Card title="Fallback chain" icon="rotate" href="/docs/deploy/routing/fallback">
    Fail over to the next target on error.
  </Card>

  <Card title="Direct routing" icon="key" href="/docs/deploy/routing/simple">
    Route a model to a single managed target.
  </Card>

  <Card title="Routing overview" icon="route" href="/docs/deploy/routing/overview">
    The three-layer model and top-down resolution.
  </Card>
</CardGroup>