Brain and models

Brain & Models opens Models. The other tabs are Connections, Fallback order, Activity and Where work runs.

Agent manages your choices by default. Select a provider to see its model for Quick tasks, Everyday work, Complex work and Hardest problems. Each card shows the model name and ID, documented capabilities, a release date when verified, and official details. A CLI alias can change at runtime; its exact release date is not shown.

Owners and admins can select a card, search models and choose one. Select Save changes. Reset removes one choice. Use Agent defaults in Advanced routing rules removes all organization routing settings. Members can inspect the choices.

Automatic adjusts the tier to the task. Fixed by task uses the purpose rules without adaptive changes. Fixed models uses the exact configured chains. Tier choices apply when routing is on. Worker overrides, budgets and provider access still apply.

A custom server needs its own model ID. Agent does not invent a default for an unknown server.

A provider is a model service, such as Anthropic, OpenAI, Gemini or an OpenAI-compatible server (Ollama, vLLM, OpenRouter, Groq). The banner says Workers can think, Text generation is degraded or No text provider can serve. Only owners and admins add or change providers.

New vendor connections start with a suggested model. Open the picker to change it. Connection settings holds its name, ID and optional embedding model. Local and custom servers holds endpoint presets. Enter an exact server model ID when no list is available. New suggested IDs keep existing connections, including disabled ones. Advanced model settings holds optional price and context data.

Workers think at three tiers: Fast, Balanced and Deep. A chain lists the models for each tier, in order. When one model fails, the next one answers. Each worker can have its own chain. See Worker settings.

The Where work runs tab sets the organization's defaults:

• Required executor: Automatic, Native platform loop, Claude Code CLI, Codex CLI or Gemini CLI.
• Host: Automatic, Managed server host, or Requester's runner.
• CLI runtime: Automatic, Host sandbox or Container.
• Cost payer: see Who pays for work.

Each value comes from the platform, then the organization, then the role, then the worker. Permissions, availability and budgets still apply. Codex Cloud is not available here.