# Models

The Mastra Memory Gateway is built on [OpenRouter](https://openrouter.ai) and supports every model in their catalog, including models from OpenAI, Anthropic, Google, Meta, Mistral, and hundreds more. Browse the full list of available models at [openrouter.ai/models](https://openrouter.ai/models).

Use any OpenRouter model ID in your request and the gateway handles routing, authentication, and memory automatically.

## Default routing

By default, all requests route through OpenRouter. Pass any model ID from the [OpenRouter models page](https://openrouter.ai/models) and it works out of the box.

```bash
curl https://gateway-api.mastra.ai/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "google/gemini-2.5-flash", "messages": [{"role": "user", "content": "Hello"}]}'
```

Your Mastra API key covers access to all OpenRouter models with no additional provider configuration.

## Direct providers

When you add a provider binding through the dashboard with your own API key, the gateway routes directly to that provider instead of through OpenRouter. Direct routing is also used in [BYOK pass-through mode](https://gateway.mastra.ai/docs/features).

| Provider  | Model prefix | Endpoints                   | Example model ID              |
| --------- | ------------ | --------------------------- | ----------------------------- |
| OpenAI    | `openai/`    | Chat Completions, Responses | `openai/gpt-5.4`              |
| Anthropic | `anthropic/` | Messages                    | `anthropic/claude-sonnet-4-6` |
| Google    | `google/`    | Chat Completions            | `google/gemini-2.5-flash`     |
| Codex     | `codex/`     | Responses                   | `codex/codex-mini-latest`     |

### Prefix stripping

Direct providers receive the model ID without the provider prefix. For example, `anthropic/claude-sonnet-4-6` is forwarded to Anthropic as `claude-sonnet-4-6`.

OpenRouter models keep their prefix intact. For example, `google/gemini-2.5-flash` is sent as-is.

## Model inference

BYOK pass-through is available on the **Teams plan** and above. In pass-through mode, when a provider key override is present but the model ID has no explicit prefix, the gateway infers the provider from the model name:

| Model pattern                                       | Inferred provider |
| --------------------------------------------------- | ----------------- |
| `claude-*`                                          | Anthropic         |
| `gpt-*`, `o1`, `o3`, `o4-*`, `chatgpt-*`, `codex-*` | OpenAI            |
| `gemini-*`                                          | Google            |

This inference only applies in BYOK pass-through mode. Without BYOK, unprefixed models route through the default provider (OpenRouter).

If the gateway can't infer a provider in pass-through mode, the request is rejected with a `400` error. Use a prefixed model ID (e.g., `openai/gpt-5.4`) to resolve this.

## Provider resolution order

When the gateway receives a request, it resolves the provider using this fallback chain:

1. **BYOK pass-through**: If a provider key override is present (`X-Memory-Gateway-Authorization` + `Authorization`), route to the provider inferred from the model ID
2. **Model prefix**: If the model ID has a recognized prefix (e.g., `anthropic/`), look for a binding for that provider
3. **Key binding**: If the API key has a provider binding, use it
4. **Project binding**: If the project has a provider binding, use it
5. **Default provider**: Fall back to OpenRouter

## Endpoint compatibility

Not all providers support all endpoint formats. The gateway validates that the resolved provider supports the requested endpoint:

| Endpoint                    | OpenRouter | OpenAI | Anthropic | Google | Codex |
| --------------------------- | ---------- | ------ | --------- | ------ | ----- |
| `POST /v1/chat/completions` | ✅          | ✅      | —         | ✅      | —     |
| `POST /v1/messages`         | ✅          | —      | ✅         | —      | —     |
| `POST /v1/responses`        | ✅          | ✅      | —         | —      | ✅     |

If a provider doesn't support the requested endpoint, the gateway returns a `400` error with details about which endpoints the provider supports.

## Custom providers

For providers not in the built-in registry, add a binding with a custom base URL through the dashboard. The gateway treats unknown providers as OpenAI-compatible and forwards requests with a standard `Authorization: Bearer` header.