> ## Documentation Index
> Fetch the complete documentation index at: https://docs.agentbees.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Adding multiple models

> Offer several models per org and let users pick one per task — or race several at once.

You're not limited to one model. An org can allow several models per agent, and
users pick one **per task** — or run one prompt across **several models at once**.

## Allow more than one model

In **Admin → Agent LLM → Org policy**, set a comma-separated **model allowlist**
per agent:

* **Claude model allowlist** — e.g. `claude-opus-4-8, claude-haiku-4-5`
* **Codex model allowlist** — e.g. `gpt-5, gpt-5-mini`
* **Gemini model allowlist** — e.g. `gemini-2.5-pro, gemini-2.5-flash`

<Note>
  An empty allowlist means "any / the configured default." With **two or more** models, the New Task
  form shows a **Model** dropdown so users choose per task — Opus for one task, Haiku for another,
  all on the same configuration.
</Note>

## Per-task model override

When creating a task, the **Model** dropdown offers exactly the models allowed for
the chosen agent. The selection is sent as a per-task override and validated
server-side against your allowlist — an off-list model is rejected with a clear
error. Pipelines can likewise set a **per-stage model** for each agent in the
lineup.

## Race one prompt across models

For **Code** tasks you can enable **Race variants**: run the same prompt across two
or more `(agent, model)` variants **in parallel**, each in its own workspace, then
review and **merge the winner**.

<Steps>
  <Step title="Open Advanced options → Race variants">Available in Code mode.</Step>

  <Step title="Add variants">
    Each row is an agent + model (e.g. Claude/Opus vs Gemini/2.5-pro).
  </Step>

  <Step title="Start the race">
    Each variant becomes its own task, grouped together. Compare the diffs and merge the one you
    like; discard the rest.
  </Step>
</Steps>

<Tip>
  Racing is a fast way to A/B two models (or two agents) on the same real task and keep whichever
  result is best.
</Tip>