A team of AI models builds one scene

One model building a big scene is like one person doing everything: it gets there, but slowly, and its weak spots show. Team mode splits the job. You pick several models, give them a task and a budget, and they build the same scene together — each one as its own agent, visible in the room.

How a run goes

1. Pick the team. In the assistant's model menu choose Team, then tick the models: Claude Haiku, Sonnet or Opus, GPT, Gemini, DeepSeek, Qwen.
2. Set a budget in your account credits: 1,000, 2,000, 5,000 or 10,000. The run stops when it reaches the limit.
3. Give the task, in any language: "a harbor town — docks, houses, boats". The team answers in the language of the task.
4. The lead model writes the plan: who builds what, and where. You see it in the team panel as the work starts.
5. The agents build in parallel, each with its own avatar and name in the scene. They agree on details in the room chat — where the dock ends, how wide the street is.
6. Up to three rounds. After each one the lead reads the agents' reports and hands out what is left. At the end it writes a short summary of what was built.

The run lives on our server, not in your tab. Close the tab, come back later and see where it is; the Stop button ends it at any moment.

What a real run looked like

We gave Sonnet and Haiku the task "a pier with a boat". The lead split the job between them, and in the room chat they agreed on where the pier ends so the boat would sit next to it. Two rounds, about 240 credits.

What it costs

You pay per token at each model's own price plus 5%, charged in your account credits as the run goes. The same credits pay for image, 3D and video generation, so there is no separate subscription for every model. If the balance can't cover the next step, the team stops cleanly instead of going into debt.

When a team pays off, and when it doesn't

• A team pays off on scenes with clear parts: a town, a level, an exhibition with stands. Cheaper models can take the simple parts while a stronger one plans.
• One assistant is better for a single object or a quick edit: a team spends tokens on planning and coordination that a small task doesn't need.

Why we built it

Models are good at different things, and they keep changing. A team lets you combine them in one scene instead of betting on one — and every agent's work stays visible and undoable, because each one is a participant in the room. How agents work in the scene.

See it on the feature page: A team of AI models.