Custom models: provider presets and a connection test

You can bring your own model to General Input from any endpoint that speaks the OpenAI Chat Completions API: OpenAI, Google Gemini, xAI, DeepSeek, Mistral, Groq, Together AI, Fireworks, OpenRouter, Cerebras, Moonshot, or your own deployment. Adding one in Settings, Models is now a guided form. Pick a provider and the URL fills in, choose the model from the list the endpoint serves, and test the connection before it reaches anyone's picker. Your provider bills you directly, and these models sit next to the hosted models members already use.

How it works

• Provider presets. Choosing a provider fills in its base URL and links to the page where you create an API key. Any other endpoint works with a URL you type.
• Model list. The form reads the models your endpoint offers, with each model's context window when the endpoint reports it, so you pick an ID rather than guess.
• Connection test. Before saving, we send the model one short message that asks it to call a tool, because agents work through tools. It reports how long that took, or what went wrong in plain words: a rejected key, an unknown model name, a model without tool calling, or a URL that isn't an OpenAI-compatible API. You can still save after a failed test.
• Images and files reach the model. A custom model marked as reading images or PDFs now receives them in chats and workflows. Before, that setting was saved but ignored.
• Public endpoints only. Our servers call your endpoint over the internet, so its URL must be publicly reachable.

Also shipped

• Custom models get their own neutral icon instead of the General Input logo, which belongs to Auto.
• Each custom model row shows its model ID and host.