Supported AI models

On a Bring-your-own-AI licence, REDA AI runs your agents on your organisation's own AI provider account, and you choose which model each connection and each agent uses. This page lists every model REDA AI supports, what each can do, and how REDA AI behaves when a setting asks for something a model cannot do. Every model listed was tested live with REDA AI before release 1.16003.

Where the model is chosen

• Each AI connection has a default model. You pick it when you create the connection on the Admin Console's AI Connections tab. Only models of that connection's provider are offered.
• Each agent can override it. On the agent's record page, the AI Account card lets you pick a connection and, optionally, a model of the same provider. Blank means the connection's default model; a blank connection means the organisation default connection.
• Admin AI tools and file reading use the organisation default connection — the connection marked Organisation Default. That includes Build Agent with AI, the Prompt Optimizer and the reading of attached files.

Changing a connection or a model takes effect for new conversations. A conversation already in progress cannot continue on a different account or model: the person is asked to start a new conversation, and automation, MCP and API calls start one automatically.

The models

Every model works with REDA AI's tools — reading and updating records, sending email, scheduling actions and the rest. What a model costs is set by your provider's price list; your provider bills you directly.

Reading images and PDF files

When someone attaches a file in the chat, REDA AI reads it with the model of the organisation default connection, not the agent's own model. So:

• If the organisation default connection uses a model that reads images and PDFs (every model above except DeepSeek V4 Pro), attachments work for every agent — including an agent that itself runs on DeepSeek V4 Pro.
• If the organisation default connection uses DeepSeek V4 Pro, attachments cannot be read. REDA AI says so by name rather than guessing at the file's content. Make a connection with another provider the organisation default if your users attach files.

Which file types REDA AI accepts, and the size limits, are described in Document Processing.

Deep Reasoning levels, and “Off” on models that always reason

An agent's Deep Reasoning setting asks the model to think before it answers. Each model offers the levels in the table. Nine models cannot turn reasoning off: Claude Opus 5.5, Claude Sonnet 5.5, GPT-6.1 Sol, GPT-6 Astra and the five Gemini models.

New agents start on Off. On a model that cannot turn reasoning off, an agent on Off runs at the model's lowest level — Low for every model above. REDA AI shows this openly rather than hiding it:

• The agent's AI Account card says: “Off is not available on [model]; this agent runs at Low.”
• The Agent Runtime Preview shows the route with “Deep Reasoning: Low (Off is not available on [model])”.

Choosing a level the model does not offer — for example Xhigh on a Gemini model — is refused when you save the agent, with a message that names the model.

More detail: Deep Reasoning.

Web search

Claude and OpenAI models can search the web for an agent that has Allow Web Search turned on. Gemini models and DeepSeek V4 Pro do not search the web through REDA AI. To limit searches to your own sites, list them in the agent's Web Search Allowed Domains. See Web Search.

When REDA switches a model off

REDA can switch a model off in a later release — for example when a provider retires it. A switched-off model:

• is no longer offered when you create a connection or choose an agent's model;
• keeps working for every connection and agent that already uses it — nothing that runs today stops;
• stays visible on an agent that uses it, so you can see what it runs on and move it to another model when you are ready.

When a model declines a request

Models can decline to answer a request they judge unsafe. When that happens REDA AI shows “The AI model declined to answer this request. Please rephrase it, or contact your administrator if this keeps happening.” instead of a half-written answer, and logs an incident you can review under AI Incident Logs.

Next steps

• Licence & Plan Management — set up AI connections and activate a Bring-your-own-AI licence.
• Configuring Agent Settings — the agent's connection, model and limits.
• Deep Reasoning — choosing a reasoning level.