# AI routing policies

> How the AI Gateway chooses between a cheaper and a stronger model for each request sent to chat-auto, how to force a choice, how to test a prompt, and what the AI Routing page on Governance → Policies is for.

Source: https://zop.dev/docs/zopnight/policies/ai-routing

---

The [AI Gateway](https://zop.dev/docs/zopnight/integrations/ai-gateway) can choose the model for you. Send a request to the model **`chat-auto`** (or its alias `zn-auto`) and the gateway sends it to one of two tiers: **Efficient**, a cheaper model for short chats, summaries, and classification, or **Frontier**, a stronger model for hard reasoning, code, and long content. Sending simple prompts to the cheaper model can lower your AI spend.

![AI routing policies page with traffic split, cache savings, and requests routed counters, the Cost-saver, Balanced, and Quality-first routing postures, and the Memory and cache settings](https://storage.googleapis.com/zopdev-blog-resources/1/files/originals/20261020/fa399559-84c5-49b5-9da5-03d9de496d8d-policiesairouting.png)

*Governance → Policies → AI Routing: routing posture, memory and cache, and where requests go.*

Routing only happens when you ask for it. A request that names a specific model goes to that model and is never re-routed.

## How the gateway chooses

For a request to `chat-auto` or `zn-auto`, the gateway checks these rules in order and uses the **first one that matches**:

| # | If the request… | It goes to |
|---|---|---|
| 1 | has an `X-ZN-Tier` header set to `strong` or `cheap` | Frontier for `strong`, Efficient for `cheap` |
| 2 | asks for a long answer: `max_tokens` of 2,000 or more | Frontier |
| 3 | is long: about 1,500 tokens or more (roughly 6,000 characters) across all its messages | Frontier |
| 4 | contains *analyse*, *analyze*, *prove*, *step by step*, or *reason* anywhere, in any message | Frontier |
| | matches none of these | Efficient |

A few things to know:

- **Every message counts**, including the system message and earlier turns of the chat, not just the latest question.
- **Rule 4 matches inside longer words**, so "improve", "approve", and "reasoning" also go to Frontier.
- **Any other header value is ignored**, and the other rules apply.
- **The choice is made by these fixed rules**, not by calling another model, so routing adds no extra cost.

### Choose the tier yourself

When you already know a request needs the stronger model, or never does, add the `X-ZN-Tier` header. It overrides the other rules.

```bash
curl "$GATEWAY_URL/chat/completions" \
  -H "Authorization: Bearer $INFERENCE_KEY" \
  -H "X-ZN-Tier: strong" \
  -H "Content-Type: application/json" \
  -d '{"model": "chat-auto", "messages": [{"role": "user", "content": "Review this contract clause"}]}'
```

Your gateway address and key are shown in the usage guide when you create an inference key under **AI Estate → Inference Keys**. Use them in place of `$GATEWAY_URL` and `$INFERENCE_KEY`.

### Test where a prompt would go

To see which tier a prompt would get without sending it to a model, use the dry run:

```bash
curl -X POST "$ZOPNIGHT_API/orgs/$ORG_ID/ai/routing/evaluations" \
  -H "Authorization: Bearer $ZOPNIGHT_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"prompt": "Summarise this ticket", "maxTokens": 300}'
```

It answers with a `tier` (`cheap` means Efficient, `strong` means Frontier) and the gateway's model name (`chat-cheap` or `chat-strong`). It looks only at the prompt and `maxTokens` you send. Add `"tier": "strong"` to see the effect of the header. `GET /orgs/$ORG_ID/ai/routing` shows the rules; they're the same for every organisation. Every built-in role can call both. See [Authentication](https://zop.dev/docs/zopnight/authentication) for tokens.

## The AI Routing page

**Governance → Policies → AI Routing** is a planning view for routing:

- **Routing posture**: **Cost-saver** (Efficient by default; only the hardest, image, or document requests reach Frontier), **Balanced** (route by how complex the task is; code, analysis, images, and documents go to Frontier), or **Quality-first** (Frontier by default; only simple chat and low-budget requests go to Efficient). All three also send requests with personal data to Frontier, and Balanced sends long answers (`max_tokens` of 2,000 or more) there too. Changing a routing condition or a tier's model switches the posture to **Custom**; changing Memory & cache keeps it.
- **Memory & cache**: reuse answers to similar questions, and connect a knowledge base.
- **Where requests go**: which model would serve Efficient and which would serve Frontier.
- **Conditions**: how complex the task is, a cost cap per request, requests that need a fast answer, images, documents, long answers, sensitive data, fitting the model's context window, falling back to another provider, and task types (coding, analysis, extraction, chit-chat).
- **Test routing**: type a prompt (optionally attach a file or mark it as containing personal data) and see which tier these settings would pick.

**Warning**

**The AI Routing page doesn't change live traffic yet.** Its settings are saved only in your browser, so they aren't shared with your team and don't follow you to another device. The **Traffic split** and **Requests routed** counters are real (Requests routed counts every gateway request, not only `chat-auto`), but the cache savings and the model lists are sample figures for planning. The gateway handles every `chat-auto` and `zn-auto` request with the rules in [How the gateway chooses](#how-the-gateway-chooses). **Test routing** is an estimate from the page's settings; use the [dry run](#test-where-a-prompt-would-go) to see what the gateway will actually do.

Only the built-in **Admin** role (or a custom role that can update AI models) can change the page's settings. Everyone else sees it read-only.

## Related

- [AI Gateway](https://zop.dev/docs/zopnight/integrations/ai-gateway): connect providers, issue inference keys, and see AI spend.
- [Policies overview](https://zop.dev/docs/zopnight/policies/overview): the other kinds of policy.