Provider keys are the prerequisite for model routing, the mode in which AI Router calls the model for you. AI Router executes routed requests exclusively on your own provider accounts (bring your own key); there is no execution on AI Router's own keys. Model costs therefore remain on your provider bill and your AI Router plan covers the routing fee only. Without at least one connected and enabled provider, model routing is not available for your team.
Model selection (the Decision API and the SDKs' get_best_model) does not use provider keys at all: AI Router returns the recommended model and you execute the request yourself. If you only use model selection, you can skip this page.
Open your team in the dashboard and go to Settings → Provider Keys. Each supported provider is listed with its connection status:
| Provider | Models |
|---|---|
| OpenAI | gpt-5.2, gpt-5.1, gpt-5, gpt-5-mini, gpt-5-nano, gpt-oss-120b |
| Anthropic | claude-4-5-sonnet, claude-4-5-haiku |
| Google Gemini | gemini-3.1-pro-preview, gemini-3.6-flash, gemini-3.5-flash, gemini-3.5-flash-lite, gemini-3.1-flash-lite |
| Mistral AI | mistral-large, mistral-small |
| Meta | llama-3.3-70b, llama-3.3-70b-fast |
| Qwen | qwen-3-235B-A22B, qwen-3-30B-A3B |
| Microsoft | phi-4 |
Click Connect, paste the provider's API key and save. AI Router validates the key against the provider before storing it; a rejected key never replaces a previously stored one. Keys are stored encrypted and are never shown again in the dashboard.
Connecting a key stores it; it does not yet route traffic to it. Each provider row has a switch Execute on your own key, which becomes available once a key is connected:
models when you execute requests yourself.Only enabled models from the team's model configuration are provisioned, so the Models page and the provider switches together define what AI Router can execute for you. When a request falls back to another model, it only falls back to models provisioned on your keys.
On the Models page you can enable Automatically add new models from enabled providers. When enabled, every model we add for one of your enabled providers is added to your allowlist and provisioned on your key. When disabled, new models are ignored until you enable them yourself.
Disconnect removes the stored key and every model provisioned with it. This cannot be undone; connect the key again to resume execution on that provider.