AI models

Every AI Agent answers with the OrhanAI default model unless you choose another. OrhanAI picks it for good answers at a low cost, and the model picker names it. On paid plans, open the AI Agent's Settings, General and pick its AI model. The choice is per AI Agent, so one can answer with Claude while another stays on the default.

Only the AI Agent's replies use the chosen model. Background work such as conversation summaries, sentiment, lead scoring, training and knowledge search stays on Gemini. Web search works only on Gemini, so an AI Agent with web search on and no model of its own answers with Gemini.

On the OrhanAI key, a more capable model spends more AI replies for each answer it sends, because it costs more to run. The faster models spend one, like the default model, and the most capable ones spend several. The model picker in the AI Agent's settings shows how many each model spends today, and that number can change.

A model that spends more than one AI reply only answers while enough replies are left for it. Replies your team writes are always free.

• Someone who manages developer settings saves an Anthropic or OpenAI API key under Settings, AI models.
• An AI Agent can then answer with that provider's models on your key. Those answers spend no AI replies; the provider bills your workspace directly.
• Keys are stored encrypted, and only their last four characters are ever shown again.
• If a saved key stops working, for example because it was revoked, the people who manage developer settings are told, and answers come from the default model until the key is fixed.

A visitor always gets an answer. When the chosen model cannot write one, the default model answers instead:

• Your own API key failed or was removed.
• The workspace is no longer on a plan that includes choosing a model.
• There are not enough AI replies left for a model that spends more than one.
• The visitor sent an image and the chosen model does not read images.

An AI Agent set to a Claude model sends the conversation it answers, with the trained knowledge retrieved for it, to Anthropic; one set to a GPT model sends it to OpenAI. Both are listed on the sub-processors page as used only when you choose one of their models.