Models

A bot answers with a model. The model decides how well the bot reasons, how much text it reads at once, and whether it understands images. Owners and Admins pick the model of a bot. Members do not pick models.

Choose the model of a bot

In the bot editor, open Behaviour, Model and instructions. The Model list shows each model with its source. You can pick a model, or "Org default". Next to it, Thinking sets how long the model thinks before it answers. The levels are Low, High, and Max. "Default" uses the default thinking of the organization, or the level of the model when there is none. The list names it, for example "Default (High)". The list shows only the levels the model supports. A model that does not reason has no Thinking setting.

"Org default" follows the default model of the organization. When the default changes, bots on "Org default" use the new model from their next message. A model change on a bot applies to the next message in every chat, also in chats that already exist.

Images in a chat work only when the model of the bot reads images. In Chat, the + button accepts images for such bots.

The Personal bot has one model for the organization. Members do not choose.

The Models page

Open Models in the Console to see the models your organization can use. The table shows the name, the model ID, the type (Text, Text + image, or Embedding), and the context length. People who may see costs also see the market price for input, cached input, and output per million tokens. The market price is what you pay for the tokens of a model. It is the purchase price plus a fee. The purchase price is what the tokens cost GEKI at the providers right now. The market price changes with the purchase price and is lower with a package. You can search and filter the table. Test opens a conversation with a model. Every table in the Console has a Refresh button above it. The tab Use the models shows the base URL, the endpoints, and a code sample, with a button to create an API key. The Model API is OpenAI compatible. See API.

A GEKI model can have aliases: shorter, more general names such as flash. The API accepts an alias as the model, and GET /v1/models lists the aliases of each model. In the Console, the model pickers list the aliases first, under "Updates automatically", and the fixed versions second. Choose an alias such as flash so the bot gets model updates. Choose a fixed version only when the model must not change. A bot on an alias follows it when GEKI moves the alias to a newer model. A bot on the model ID stays on that model. Disabling a model also disables its aliases.

Default models

The Models page starts with the Default models card. Everyone who opens Models sees it. Only an Owner or Admin changes it.

The card has six settings:

• Default model. Used by bots on "Org default" and by API calls without a model.
• AUX model. Used for internal tasks such as chat titles and name detection. "Same as the default model" is the default.
• Assistant model. Used by the Console assistant. Until you choose, it shows "GEKI default (DeepSeek V4.1 Flash, High)". Reset removes your choice.
• Assistant thinking. How long the assistant thinks. It shows the same GEKI default until you choose a level. If you only change the model, the level stays when the new model supports it.
• Default thinking. How long models think when a bot, an API call, or a chat does not set it. Low, High, or Max. Without a choice, each model uses its own level. A model runs at the nearest level it supports. The assistant uses its own thinking setting.
• Fall back to the default model. When a model is not available, the call uses the default model.

A child organization uses the model settings of its partner. It can set its own values.