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

# Routing and fallbacks

> Fallback models, provider preferences, reasoning, plugins, and other advanced options.

These are advanced options the pre.dev AI API supports on top of the standard chat completion fields. This page lists them and how they interact with billing.

## Fallback models

Send `models` instead of a single `model`. The gateway tries each in order and the response reports which one answered.

```json theme={null}
{
  "models": ["anthropic/claude-sonnet-5", "openai/gpt-5.6-luna", "deepseek/deepseek-v4.1-flash"],
  "route": "fallback",
  "messages": [{ "role": "user", "content": "Summarize the plot of Dune in two sentences." }]
}
```

You pay for the model that actually served the call, at its own price. Read `model` in the response, or the `by_model` rows in [usage](/ai-gateway/api/usage), to see where traffic landed.

## Provider preferences

The `provider` object controls which providers serve a model:

| Key               | Purpose                                                |
| ----------------- | ------------------------------------------------------ |
| `order`           | Providers to try first                                 |
| `sort`            | Sort candidates by `price`, `throughput`, or `latency` |
| `ignore`, `only`  | Exclude or restrict providers                          |
| `max_price`       | Per-token price ceiling                                |
| `data_collection` | Allow or deny providers that retain data               |

```json theme={null}
{
  "model": "deepseek/deepseek-v4.1-flash",
  "provider": { "sort": "throughput", "data_collection": "deny" },
  "messages": [{ "role": "user", "content": "ping" }]
}
```

`usage.cost` reflects the chosen provider's price, so `max_price` bounds your credit spend per token. The `:nitro` variant is shorthand for throughput sorting.

## Reasoning

`reasoning` (an object) and `reasoning_effort` (a string) request thinking on models that support it. Reasoning tokens are output tokens for billing.

```json theme={null}
{
  "model": "anthropic/claude-sonnet-5",
  "reasoning": { "effort": "medium" },
  "messages": [{ "role": "user", "content": "Is 9973 prime? Show your reasoning briefly." }]
}
```

## Plugins and web search

`plugins` accepts `web`, `file-parser`, `response-healing`, and `context-compression`. `web_search_options` tunes the web plugin. Plugin costs are included in `usage.cost`.

```json theme={null}
{
  "model": "openai/gpt-5.6-luna",
  "plugins": [{ "id": "web" }],
  "web_search_options": { "search_context_size": "low" },
  "messages": [{ "role": "user", "content": "What changed in the latest Node.js LTS release?" }]
}
```

## Other pass-through fields

| Field        | Purpose                                                            |
| ------------ | ------------------------------------------------------------------ |
| `transforms` | Prompt transforms such as middle-out compression                   |
| `prediction` | Predicted outputs                                                  |
| `verbosity`  | Response length hint                                               |
| `session_id` | Group related requests under one session                           |
| `user`       | End-user identifier; the gateway prefixes it with a workspace hash |

If you send `user`, keep it stable per end user so abuse reports can be attributed. The gateway sets it when absent.