# Auto routing

Source: https://platform.lyrouter.com/ai_gateway/docs/auto/en
Last updated: 2026-10-22

Set the request's `model` to `lyrouter/auto`, and the gateway picks the model to call from the models enabled in the workspace, based on how difficult the request's task is. It suits scenarios that tolerate a changing model, such as chat and coding assistants. For applications with strict requirements on output structure or model behavior, name the model ID directly, or pin the model with a [virtual model](https://platform.lyrouter.com/ai_gateway/docs/presets/en).

## Calling Auto

Every workspace can call `lyrouter/auto` directly, and `GET /v1/models` lists this name too. Setting the environment variables is covered in [Your first model call](https://platform.lyrouter.com/ai_gateway/docs/quickstart/en):

```bash
curl "$LYROUTER_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer $LYROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"lyrouter/auto","messages":[{"role":"user","content":"Write a quicksort function in Python"}]}'
```

- **Supported endpoints**: `/v1/chat/completions`, `/v1/messages` and `/v1/responses`. Image, video and audio endpoints do not support Auto and return 403 `Model is not available in the model plaza`; use a specific model ID instead.
- **Time to first token**: the gateway identifies the task type before forwarding the request, so the time to first token is slightly longer than calling a model directly.

## How the model is picked

1. **Identify the task type**: the gateway judges the intent from the last user message, for example Build & debug or Transform content.
2. **Map to a difficulty tier**: each intent category maps to one difficulty tier.
3. **Sort the candidates**: the tier's models come first, in the priority order set by the platform; the other models in the candidate scope follow in alphabetical order of model ID. When no task type is identified, all candidates are sorted alphabetically.

| Difficulty tier | Intent categories | Preferred model type |
| --- | --- | --- |
| Hard (`hard`) | Build & debug, Analyze & calculate, Decide & plan, Automate workflows, Review & validate, Governance & compliance | Flagship reasoning and coding models, such as `deepseek/deepseek-v4-pro` and `qwen/qwen3.7-max` |
| Medium (`medium`) | Create content, Summarize & synthesize, Extract & structure, Creative media, Communication & cases, Learn & practice | Balanced mainstream models, such as `anthropic/claude-sonnet-4.6` and `qwen/qwen3.7-plus` |
| Easy (`easy`) | Ask & explain, Transform content | Fast, low-cost small models, such as `anthropic/claude-haiku-4.5` and `deepseek/deepseek-v4-flash` |

- The platform maintains the priority lists, and workspaces cannot change them. A model on a list is chosen only when it is in the candidate scope and enabled in this workspace.
- The gateway tries the candidate models in order and moves to the next model only after every channel of the current one fails; the workspace's [Default Model](https://platform.lyrouter.com/ai_gateway/docs/routing/en) comes last.
- Channels of the same model are not sorted by the multi-channel routing strategy; Prefer this Channel and Only use this BYOK in BYOK still apply.

## Setting the candidate scope

The current Account Owner clicks **Advanced settings** in the Auto smart routing row of the **Smart routing** page, sets the candidate scope in the drawer, and clicks **Save Auto settings**. Account Members do not see this page, but they can call `lyrouter/auto` directly.

- **All enabled models** (the default): every model enabled for this workspace in the model marketplace is a candidate.
- **Specific models**: check enabled models, or type a model ID or wildcard rule in the search box and press Enter to add it; for example, `anthropic/*` matches every model starting with `anthropic/`. Manually added entries carry a "manual" tag.

(Screenshot: The candidate scope in Auto advanced settings decides which models can be chosen; the cost / quality tradeoff and request override switch do not affect calls yet.)

- **Only enabled models qualify**: a model ID you add that is not enabled in this workspace never becomes a candidate. When the scope contains no usable model, the request returns 503 `all routed candidates exhausted`; if a Default Model is set, it answers instead.
- **Keep only chat models**: when the selected model does not support the current endpoint (for example, an embeddings or image model picked for a chat endpoint), the request returns 400 `unsupported_model_api` directly and does not move to the next candidate. If you use All enabled models and the workspace has non-chat models enabled, switch to Specific models.
- **Affects only `lyrouter/auto`**: calls that name a model ID are not limited by the candidate scope. To limit which models a workspace can call, disable models in the model marketplace or use [Model access](https://platform.lyrouter.com/ai_gateway/docs/authorization/en).

The Cost / Quality Tradeoff slider and the "Allow requests to override workspace routing" switch in the drawer are only saved and do not affect calls yet: Auto picks models by task difficulty alone, and requests have no parameter that overrides routing settings.

## Verify

After a call, open its record in the [request logs](https://platform.lyrouter.com/ai_gateway/docs/logs/en). In the route trace:

- Virtual model shows `lyrouter/auto`;
- Routing strategy shows `intent:hard`, `intent:medium` or `intent:easy`, or `intent` when no task type was identified;
- Model (actual) shows the model that answered.

Compare the actual model, latency and cost on a fixed set of test tasks before deciding whether to use Auto in production.