How to Branch AI Conversations to Compare Approaches

If you have ever wished a chat had a git checkout, branching is the feature you want. Knowing how to branch AI conversations means you can fork a thread at any message, try a completely different answer down one path, and still walk back to the original whenever you like. This guide is for power users who already switch models mid-chat and now want to run parallel experiments without ever overwriting a good reply. It works on web and mobile, and the branch action itself is free.

The alternative is re-prompting in place: you tell the model "actually, redo that," and the old answer is gone. Branching keeps both. You get a tree of attempts instead of a single editable line, which is exactly what you want when the answer has to be right and you do not yet know which approach gets you there.

What a branch actually does

A branch copies the conversation up to a chosen point into a brand new thread, then generates an alternative from there. Think of the message you branch from as the commit you check out. Everything before it comes along for context. Everything after it in the original is untouched and still sitting in your sidebar.

Because it is a separate thread, the new branch is fully independent. You can switch models in it, edit prompts in it, even branch again from inside it, and none of that leaks back into the parent. On web the new thread shows a small GitBranch icon in the sidebar with a Branched from: label, so you can always trace a branch back to where it split off.

This is different from switching models mid-conversation, where you keep one thread and hand the same history to a new model. Branching forks the thread; switching stays on it. The two pair up nicely, and we will get to that.

Step-by-step: how to branch a chat

There are two entry points. The first branches from an assistant reply you already have. The second branches from an edit to one of your own messages. Here is the assistant-reply path, which is the one you will reach for most.

Find the assistant message you want to fork from. This is your branch point. Pick the last reply that was still good, before the conversation went somewhere you want to redo.
On web, open the branch dropdown on that message. It carries the GitBranch icon and the tooltip Create new thread with alternative response. On mobile, open the message menu titled Branch and tap Branch here.
To rerun with the same model, use the Try Again option at the top of the dropdown. The current model is tagged (Previous) so you know which one produced the original.
To fork to a different model instead, use the or switch model section and pick the model you want to try. The new thread starts on that model.
MultiChats creates a new thread with everything up to the branch point and generates the alternative. Your original chat stays exactly where it was. Jump between the two from the sidebar.

The second entry point is the edit path. When you edit one of your own messages, web offers a Branch button with the tooltip Keep current reply and branch from this edit. That spins up a new thread using your rewritten prompt while leaving the original prompt and its reply intact. This is the move for testing two phrasings of the same question side by side.

One gating note: branching itself costs nothing. The only paywall is per target model. If you branch to a premium model on a free plan, that model is locked the same way it is anywhere else (it shows a ✦ Pro badge). Branch to a free model and there is no gate at all.

Three ways to use branches

Branching is a general tool, but three patterns cover most of what it is good for.

Same question, different model. Branch from your question and fork it to a second model to see who answers better. For code, branching the same failing function to a reasoning-heavy model is a fast way to get a second opinion. Our take on which models to reach for lives in the best AI app for coding write-up.
Two prompt phrasings. Use the edit-and-branch path to test "explain this for a beginner" against "give me the rigorous version," both from the same starting context. You keep both answers and can compare wording, depth, and tone without losing either.
Preserve a known-good answer. Got a reply that nails it, but you want to push the model further? Branch first, then experiment on the branch. The good answer is safe in the parent thread no matter how far the branch wanders.

Once you have a few branches running, line up their final answers using the compare AI answers across models workflow to pick a winner.

A non-obvious move: branch to escape a bloated thread

Here is the trick experienced users lean on. When a chat has run long, the model starts dropping details from earlier turns, because a bigger context window does not make a model reason cleanly over every token in it. Instead of fighting the drift in place, branch from the last clean answer. The new thread carries only the history up to that point, so you drop the noisy tail that was confusing the model and keep the part that still made sense.

Picture the git tree. Your main thread is a trunk that picked up ten turns of back-and-forth, three of which sent the model down a wrong path. You check out the commit right before the detour (branch from that assistant message), and the alternative line grows from clean ground. The detour is still on the trunk if you need it. Nothing is lost, and the new branch is running on a tighter, more accurate slice of the conversation.

Pair this with a model swap and it gets sharper. Branch from the clean point, fork to a stronger reasoning model for the hard part, then drop back to a lighter, faster model once the thinking is done. You are pruning context and upgrading the engine in the same move, and the original thread never notices.

Branch vs switch vs new chat

Three actions look similar and solve different problems. Use this to pick the right one.

Action	Keeps history?	Best for
Branch	Yes, copied to a new thread	Exploring an alternative while keeping the original
Switch model	Yes, same thread continues	Handing one ongoing chat to a stronger model
New chat	No, starts empty	An unrelated question with no shared context

If you mostly want to hand a single thread to a better model, the dedicated switch AI models mid-chat guide covers that path in full.

Branching checklist

Run down this list before you fork a thread so the branch stays useful.

Pick the right branch point: the last message that was still good, not the one that went wrong.
Decide same-model rerun (Try Again) or a different model (or switch model) before you click.
For two phrasings, use the edit-and-branch path so both prompts share the same starting context.
Rename a busy branch in the sidebar so you do not lose track of which experiment is which.
On a free plan, check the target model is not premium (the ✦ Pro badge) if you do not want to hit the paywall.

Frequently asked questions

What does branching a chat do?

It copies the conversation up to a message you choose into a new thread, then generates an alternative answer from that point. The new thread is independent, so you can take it in a different direction while the original stays exactly as it was.

How is branching different from switching models?

Switching models keeps you on the same thread and hands the existing history to a new model. Branching forks the thread into a separate copy. Use switch when you want one continuous chat to keep going on a better model, and branch when you want to compare two paths at once.

Will I lose my original conversation if I branch?

No. The original thread is never touched. Branching always creates a new thread, and the parent keeps every message it had. On web the branch is even labelled Branched from: its source so you can find your way back.

Is chat branching a paid feature?

The branch action is free on web and mobile. The only paywall is per model: branching to a premium model on a free plan hits the same lock as using that model anywhere else. Branch to one of the free models and there is no gate.

Wrap-up

Branching turns a linear chat into a tree you can explore without fear of losing a good answer. Pick a branch point, decide whether you are rerunning or switching models, and let the parent thread sit safely while you experiment. Once you are running parallel branches across models, MultiChats Pro opens up every model worth forking to, so your branches can pit the best answers against each other in one place.