When the models disagree

A split is a tell about your question. Two models given identical input produced different answers. Something in the question decided the difference: a missing constraint, a contested fact, or a trade-off you had not named. Agreement is confirmation, and confidence in the answer. A split is information: it tells you where to dig. Big-AGI puts every answer in front of you, each labelled with the model that wrote it. It computes no agreement score, no vote and no consensus indicator. Reading the split is yours.

Three kinds of split

Editorial guidance, July 2026:

What you are looking at	What it means	What to do with it
Different answers to the same question	A factual split - at least one of them is wrong	Find the claim they differ on and settle that one claim
Answers to different questions	Your prompt was underspecified, and each model resolved the ambiguity its own way	Rewrite the question around the ambiguity and run again
Same substance, different emphasis	Style and priority, not disagreement	Take whichever reads best; there is nothing to decide here

The second is the most useful: the run has found the assumption your question rested on.

What a merge can hide

The right merge operation depends on what you are after: keep the shortest answer or the longest, fuse all of them, drop the ones that miss a requirement, take the strongest paragraph from each, or score them against criteria of your own. Custom merges do any of those. The built-ins cover the common cases; the custom prompt covers yours (your own merge prompt).

The merging model receives the answers as ordinary messages, with no model labels attached. Neither Fuse nor a plain custom prompt asks it to preserve a position only one answer took. A merged paragraph can therefore read like consensus over inputs that disagreed sharply. Three ways to keep the split visible:

• Read the cards before adding any merge card.
• Use Guided, whose first stage is asked for points of difference and lists them as tick boxes. Or Compare, which keeps one row per answer.
• Overrule the merge by sending a single answer forward instead (combining answers, run your first Beam).

Compare is a judgement, not a measurement. One model invents the criteria and scores answers it did not write. It is also instructed to spread those scores over the full range, so the spread is a formatting rule, not a finding. The winner it declares is that model's opinion.

Related

• Run several models at once
• Combining answers
• Model settings