Multi-model AI coding: Claude, GPT, Grok, Kimi and Muse in one room

Multi-model AI coding is when a team builds one codebase with AI models from more than one lab: one person working with Anthropic's Claude, another with OpenAI's GPT, a third with xAI's Grok, and all of the work landing in one place. In Poly (usepoly.co), that place is a room. As of September 2026 a room can use 13 models from five labs (Anthropic, OpenAI, xAI, Moonshot AI and Meta), and every model works in the same workspace, the same git history and the same timeline. Each member picks their own model and thinking effort. When the conversation moves from one lab to another, the room briefs the next model on what it missed.

Key takeaways

• A Poly room can use 13 models from five labs, and every lab has at least one free model.
• The model is chosen per person: each prompt runs on the model its author picked, and the turn is billed to that person.
• Labs can't read each other's conversations, so Poly keeps one conversation per lab family and briefs each model on what the others did.
• A shared room runs one turn at a time. Multi-agent mode gives each member their own agent and branch, so several labs can work at once.

Why would a team use more than one AI model?

Because the people on it already do. Developers have preferences, and the model one person trusts is the model another person avoids. There are practical reasons too:

• Cost. A small fix doesn't need a frontier model. Free, fast models such as GPT-6 Luna, Kimi K2.7 Code and Grok Build handle the small turns; Opus 5.5 or GPT-6 Sol take the hard ones.
• A second opinion. A model from a different lab was trained differently, so it tends to catch different mistakes. Cross-model code review covers what the research says, including the direction that didn't help.
• No single point of failure. When one lab is slow, rate-limited or down, members switch to another and keep working in the same room.

Which models can a Poly room use?

Every lab has a free model, and any paid plan or credit top-up unlocks the rest. Google's Gemini and DeepSeek are not in Poly as of September 2026.

How does one room run models from five labs?

Poly runs two real agent harnesses, not a lookalike of either. Claude, Kimi and Muse run in Claude Code, through Anthropic's Agent SDK: Moonshot AI and Meta both serve Anthropic-compatible APIs, so their models drive the same agent. GPT and Grok run in OpenAI's Codex. Every turn runs in its own sandbox on Poly's servers, and nobody brings an API key.

The room works the same way whichever lab is answering. Commands, file edits, the agent's questions and image generation arrive as the same cards, and any member can approve or answer them. A few differences remain:

• Questions. GPT and Grok can ask the room a question only in plan mode, so a GPT or Grok room's first build usually starts as a plan. Claude, Kimi and Muse can ask in any mode.
• Web search. Grok has none in Poly. xAI rejects the web search tool that Codex sends, so Poly removes it.
• Images. Grok turns generate images with xAI's Grok Imagine; every other model uses OpenAI's image model.
• Thinking effort. Grok Build sets its own, so its effort slider is greyed out.

Who picks the model?

Each member does, per room. The pick is saved with their membership and becomes their default for new rooms. Every prompt runs on its author's model, the member list shows who is using what, and the agent's briefing lists each member's model too, so the agent knows who it is working with. Thinking effort is chosen the same way.

Each turn's receipt records the model that actually ran, and the turn's cost goes to the person who sent the prompt. Watching, chatting and approving are free.

What does each model know about the others' work?

Everything that lives in files, and a briefing on the rest.

• The workspace and git history are shared by every model. They are the ground truth.
• The conversation is where labs differ. Claude, Kimi and Muse share one conversation per room, because all three read Anthropic's format. GPT keeps its own, and so does Grok: a Codex conversation carries the lab's encrypted reasoning, which another lab cannot decode.
• The catch-up bridges them. When a prompt runs on a conversation that missed recent turns, Poly adds what it missed, verbatim or summarized. What happens when you switch AI models mid-conversation has the rules.
• The authorship record keeps the shared conversation honest. Replies there carry no author, so a model can mistake another model's work for its own. Poly tells it who wrote what and signs every commit with the agent's name: why AI models claim each other's work.

Can two models work at the same time?

In a shared room, no. The room runs one turn at a time, first in, first out, whatever model each prompt uses. Two people on two labs take turns, and each turn starts from the last one's changes.

In multi-agent mode, yes. The room's owner turns it on, and each member gets their own agent on their own branch, refreshed from the main branch before every turn. Members merge their branches when the work is ready. Each personal agent keeps its own conversation, so there is no catch-up between them; their work meets in git.

What are the limits?

• Switching labs has a cost. The first turn on a different lab starts without a prompt cache and carries a catch-up, so it costs more than the turns after it.
• Context windows differ. Haiku 4.5 holds 200K tokens, Kimi K2.7 Code and Grok Build 256K, Grok 4.7 500K, and the rest about a million.
• The labs are not identical in the room: the question, search, image and effort differences above.
• Gemini and DeepSeek are not available.
• Poly is hosted and desktop-browser only.

Who is it for?

Teams whose members already prefer different models, cofounders who want a second lab to check the first, and anyone who wants to try Grok, Kimi or Muse on a real project without setting up a key for each. If you work alone and want one model in your own terminal, Claude Code or Codex on its own is simpler.