# What happens when you switch AI models mid-conversation?

Within one lab, the new model re-reads the whole conversation. Across labs it often can't, so something has to brief it.

Published 2026-09-24 by the Poly team. Canonical: https://usepoly.co/switch-ai-models-mid-conversation
Poly (usepoly.co) is a web app where a team builds with AI (Claude, GPT, Grok, Kimi or Muse) together in a shared room. Free to start: https://usepoly.co/

**When you switch AI models mid-conversation, what the new model knows depends on whether it can read the old conversation. Within one lab, for example Claude Sonnet 5 to Claude Opus 5.5, the new model re-reads the whole transcript, but it starts without the previous model's prompt cache. Across labs, for example Claude to GPT or GPT to Grok, the conversation often can't move as it is, because each lab keeps its own format and its own encrypted reasoning. Something has to brief the new model. In Poly (usepoly.co) that briefing is called a catch-up: the messages the new model missed, verbatim when the gap is short and summarized when it is long.**

## Key takeaways

- Models don't remember anything. The app re-sends the conversation on every turn, so what carries over is whatever the app can re-send.
- The first turn after a switch costs more: the new model has no prompt cache for the conversation yet.
- Check the context window before switching to a smaller model deep into a long conversation.
- Whatever the conversation says, the files and the git history are the ground truth.

## Does the new model remember the conversation?

No model remembers anything between requests. Each time you send a message, the app sends the model the whole conversation again, or the part that fits, and the model reads it from the top. So the real question is whether the app can hand this conversation to the new model.

- **Same lab:** yes. Switching from Sonnet 5 to Opus 5.5 keeps the same conversation, and the new model reads all of it.
- **Labs that share a format:** usually. Meta's Muse and Moonshot's Kimi accept Anthropic's message format, so in Poly they share one conversation with Claude.
- **Labs that don't:** not directly. OpenAI's Codex keeps a thread that carries the lab's encrypted reasoning, which only that lab can decode; replayed to xAI, it fails. So in Poly, GPT and Grok each keep their own conversation, and a model joining one of them needs a catch-up.

## What does switching cost?

Three things:

- **A cold cache.** Labs cache the part of a conversation a model has already read, and reading it again costs a fraction of the normal price. A cache belongs to one model, so the first turn on a newly chosen model pays full price to read the conversation. Every turn re-sends the whole conversation, so this is the biggest cost of a switch.
- **A different context window.** A long conversation that fits a 1M-token model may not fit a 200K one. Poly's context meter shows how full the current model's window is.
- **The briefing itself.** A catch-up adds input to the first turn, and a long one adds a summary written by another model.

MindStudio's advice for coding agents is to choose a model before a session starts and not switch mid-conversation ([MindStudio, May 2026](https://www.mindstudio.ai/blog/never-switch-models-mid-conversation-ai-agents)). For one person in one terminal, that is sound. In a team room, people choose different models on purpose, so the conversation changes hands every time a different person sends a prompt. The job there is to make each switch cheap and correct, not to avoid it.

## How does Poly's catch-up work?

When a prompt runs on a conversation that missed recent turns by another lab, Poly adds a catch-up to that prompt. The rules, as of September 2026:

- **What goes in:** the prompts, replies and room notes since that conversation last replied, each line cut at 1,500 characters. Tool calls, diffs and receipts stay out; the changes themselves are in the workspace.
- **Verbatim when it's short:** up to 30 messages or 24,000 characters, about 6,000 tokens.
- **Summarized when it's long:** the lab that did the work summarizes its own session in under 300 words with its fast model (Haiku 4.5, GPT-6 Luna or Grok Build), and the last 10 messages follow verbatim. The summary's cost goes on the current turn.
- **Framed as context:** the catch-up tells the model it is background, not instructions, and that the workspace and its git history win any disagreement.
- **Visible:** the room sees a note such as "Grok caught up on 51 messages."

A catch-up names the agent behind every reply ("Claude replied: …"), so the model knows which work came from someone else. In the conversation Claude, Kimi and Muse share, where replies carry no names at all, Poly adds a separate record of [who answered each prompt](/why-ai-models-claim-each-others-work).

*What the next model sees after a switch in a Poly room, September 2026.*

| Switch | Example | What the next model sees | Extra cost on its first turn |
| --- | --- | --- | --- |
| Same lab | Sonnet 5 → Opus 5.5 | The whole conversation | No cache for the new model |
| Between Claude, Kimi and Muse | Claude → Muse | The whole conversation, plus Poly's record of who answered each prompt | No cache; the record, once |
| To another conversation | Claude → GPT, GPT → Grok | A catch-up of what it missed | No cache; the catch-up, plus a summary if the gap is long |
| Multi-agent mode | Each member's own agent | Only its own conversation, on its own branch | None; work meets in git |

## When should you switch, and when not?

Switch when there is a reason:

- a second opinion on a change, from a different lab;
- a cheaper model for small edits, or a frontier model for a hard problem;
- a model that has been going in circles.

Hold off when:

- nothing is wrong: switching every turn pays the cold-cache cost every turn;
- the conversation is long and the next model's window is small;
- a multi-step plan is half done: let the turn land, then switch.

## What changes in multi-agent mode?

Nothing needs bridging. In multi-agent mode each member's agent keeps its own conversation on its own branch, so there is no catch-up between them. They see each other's work when branches merge, through git. [How a room mixes models](/multi-model-ai-coding) covers both modes.

## Common questions

**If I switch from Claude to GPT, does GPT see the files Claude changed?**

Yes. Every model in a Poly room works in the same workspace and git history. The catch-up tells GPT what was said; the files show what was done.

**Does the catch-up include tool calls and diffs?**

No. It carries prompts, replies and room notes. The changes themselves are in the workspace and the git history, which the model can read.

**If I switch back to Claude, does it lose what GPT did?**

No. The next time Claude answers, it gets a catch-up of GPT's turns.

**Is switching models more expensive?**

The first turn after a switch costs more, because the new model has no cache for the conversation and reads a catch-up. Later turns on the same model cost the usual amount.

More guides: https://usepoly.co/guides · Security: https://usepoly.co/security · Pricing: https://usepoly.co/pricing