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

# Models

> Every model Opper supports, with EU-hosted options marked.

Opper supports 700+ models from 50+ providers through one gateway and key. Browse the full, always-current catalog — with pricing, context windows, capabilities, and EU-hosted models marked — at [opper.ai/models](https://opper.ai/models).

> [card] [Browse all models](https://opper.ai/models)
>
> The complete model catalog. Copy a model id to use it from the SDK or any drop-in client.

Call generation models by their provider-prefixed name (`openai/gpt-5.5`, `anthropic/claude-sonnet-4-6`, `gemini/gemini-3.1-pro-preview`) through the compatible SDKs. See [Drop-in SDKs](https://docs.opper.ai/build/gateway/drop-in-sdks.md). Models for other tasks use their dedicated endpoints.

## Evaluation models

Decision models provide yes/no probabilities, category choices, and ordered scores. Discover them with `GET /v3/models?type=evaluation`, then call `POST /v3/compat/decisions` (OpenAI's Decisions format) or `POST /v3/compat/v1/systemone` (TypeSafe's System One format) with your project API key. Every decision model answers on both; see [Classification & scoring](https://docs.opper.ai/build/gateway/evaluations.md).

Evaluation models cannot be called through Chat Completions.

## Multimodal models

Beyond text, the catalog includes models that accept images, audio, and video, and models that generate images, speech, and video. Filter the [catalog](https://opper.ai/models) by capability — `vision`, `audio`, `video`, `image gen` — to see what each model supports.

For the modality-specific endpoints, a discovery call returns the usable models and their capabilities at runtime, so you don't have to hardcode a list:

| Endpoint | Models for |
| - | - |
| `GET /v3/images/models` | [Image generation](https://docs.opper.ai/build/multimodal/images.md) (`POST /v3/images`) |
| `GET /v3/audio/models` | [Speech & transcription](https://docs.opper.ai/build/multimodal/audio.md) (`POST /v3/audio/speech`, `/v3/audio/transcriptions`) |
| `GET /v3/videos/models` | [Video generation](https://docs.opper.ai/build/multimodal/video.md) (`POST /v3/videos`) |

For sending images and PDFs into a text model, see [Vision & PDFs](https://docs.opper.ai/build/multimodal/vision-pdfs.md).

## Fallbacks and backup chains

There are two ways to get a fallback chain.

**Per request, on Chat Completions.** Send the primary in `model` and the
rest in `models`, the same shape OpenRouter uses:

```json theme={null}
{"model": "openai/gpt-5.6-luna", "models": ["azure/gpt-5.6-luna", "deepseek-v4.1-flash"], "messages": [...]}
```

Entries are tried in order when the one before fails for any reason. An
entry can be a pinned id, a bare pool name (its providers are tried in turn)
or a legacy alias. A [compliance rule](https://docs.opper.ai/control-plane/rules/overview.md) that blocks
an entry removes it from the chain without an error. `meta.routing.attempts`
lists every attempt with its status and `meta.routing.served` names the model
that answered. Only `/v3/compat/chat/completions` reads `models`.

**Managed on the gateway, for every endpoint.** Build a
[dynamic route](https://docs.opper.ai/capabilities/routes/overview.md), give it the models you want
to try, deploy it, and call the route's name as `model` from any SDK, on
every endpoint. This is the option to reach for when the chain should live in
one place, with ranking, a version history and rollback. The gateway tries the models in the route and the first one
that answers wins; a model that is down or rate-limited is passed over, and
in a Pool so is one that a [compliance rule](https://docs.opper.ai/control-plane/rules/overview.md)
blocks for your organization — neither comes back as an error.

```text theme={null}
production-main  →  Pool: azure/gpt-5.6-luna, melious/deepseek-v4.1-flash
                    ranked per request on price, latency or throughput;
                    the first that answers wins
```

Two nodes answer a request:

* A [**Pool**](https://docs.opper.ai/capabilities/routes/pool.md) holds several models and re-ranks
  them on every request on live numbers — cheapest, fastest, highest
  throughput, or a blend. Use it when any of the models is acceptable and
  you want the gateway to pick.
* A **Model** node calls one model, with fallbacks in the fixed order you
  wrote down. Use it when the order matters — "this model, and only if it
  fails, that one".

Either can sit on the other's fallback edge, so a route can say "try this
pool, then that model". Every attempt is recorded on the trace, so you can
see which model actually answered and why the others were passed over.

Two things this buys you:

* **Resilience.** If the primary provider is down or rate-limited — or, in a
  Pool, a compliance rule takes it out of play — the call continues to the
  next model instead of failing.
* **Swap models without touching code.** Point your app at the route name,
  then reorder or repoint the route in the platform — no deploy.

Routes are part of the [Control Plane](https://docs.opper.ai/control-plane/overview.md). To put your
own provider deployments into a route, register them as
[custom models](https://docs.opper.ai/capabilities/custom-models.md) first.

> [!NOTE]
>
> Model **aliases** (Settings → Models), a stable name mapping to an ordered
> list of models, still resolve on every call for organizations that use them.
> New fallback chains should be built as dynamic routes: they give you ranking,
> per-attempt traces, and the same skip-on-compliance behaviour.

## What's next

> [card] [Dynamic routes](https://docs.opper.ai/capabilities/routes/overview.md)
>
> Pick the model per request: pools, branches, splits, classifiers.

> [card] [Custom models](https://docs.opper.ai/capabilities/custom-models.md)
>
> Register your own model deployments and API keys.

> [card] [Apps](https://docs.opper.ai/integrations/overview.md)
>
> Use Opper as the provider for your editor, agent, or CLI.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.