Edit-and-regenerate throws away the reply you liked. Branching keeps it. Try a new prompt, swap the model, explore a tangent. Your original answer is still there when you come back.
Claude · GPT · Gemini · Grok · DeepSeek. One thread, many branches.
You can edit and retry. You can't keep the original and try again. That's not branching, that's undo.
Every major chat app lets you edit a past message and retry, but the previous answer is gone. If the new one is worse, you can't go back and compare.
To test a different model on the same question, you paste the whole thread into another app. Context dies, memory resets, and comparing two outputs means manually scrolling between tabs.
You can't ask "which provider answers this sub-question better?" without committing to one and losing the other. Single-vendor tabs physically can't.
Three turns deep into a direction you didn't want, you have to start over. No "what if I'd pushed it harder at turn 4?" The tree only grows one way.
The problem isn't that models disagree. It's that your chat app forces you to pretend one of them didn't speak.
Major chat apps call their replay feature "branches." They overwrite. Cognis forks. The difference shows up the moment you want the earlier answer back.
Retry / edit
ReplaceOne answer at a time. Previous answer destroyed.
Drafts
PickMultiple outputs for a single prompt. No continuation, no follow-up.
Copy-paste to another tab
FragmentNew context, new memory, no link back to the original thread.
True branching
ForkBoth answers live. Conversation continues on either or both. One parent history, many futures.
Branching isn't a button. It's a data model. Your current tool doesn't have it.
A fork changes one variable at a time. Keep everything else equal, test the thing you actually care about.
Fork a message and rerun it on Claude, GPT, Gemini, Grok, or DeepSeek. Both answers stay in the thread. You keep whichever one fits the task.
Claude
GPT-5KEPT
GeminiFork with the same provider but a different follow-up. Explore "what if I'd asked for more technical detail at turn 4?" without losing your current thread.
The mechanics behind the demo above. Walk the motion yourself in the live preview below.
Every message in the thread has a fork affordance. AI turns, your turns, anything in history.
Swap to a different provider, or send a different follow-up. The parent chain stays intact.
Switch between them with one click. Keep whichever wins, or keep both and come back later.
Every chat tool has a word for it. Very few have the data model.
edit + regenerate
retry
drafts
--
Supported Partial (destructive or single-prompt only) Not available
No overwrites. No bleed. No lost threads.
Try anything. Keep everything.
Every branch is saved. Come back in a week, the answer is still there, exactly as it was.
Fork at turn 4, the new branch sees turns 1 to 4 and nothing else. No bleed between branches.
A branch run on GPT-5 stays labeled GPT-5 forever. Follow-ups run on the same provider, not your current default.
Branches live inside the original conversation. Sidebar doesn't explode. Search still works. Exports include every branch.
Always satisfied with the first answer? Cognis probably isn't for you, and that's OK.
This feature works because of one, and enables the other. Both pair with it in real workflows.
Branching works because the chat layer is provider-agnostic. Every branch can run on any model: Claude, GPT, Gemini, Grok, DeepSeek. Single-vendor apps physically can't.
Read the Multi-LLM page →The winning branch lands in Liveboard: editable in place, previewable live, one tap to Gmail, Sheets, Calendar or your CRM. No copy-paste out of chat.
Read the Liveboard page →Every frontier model, one conversation, forkable at any point.
Free plan includes branching across every supported model. No credit card.