How a debate runs
- Each model answers on its own.
- Each reads the others' answers and challenges the weak points.
- Each revises.
- A referee model that didn't debate writes the final answer, says how confident it is, and lists what they still disagree on, with the stronger case for each side.
A real run: 34 seconds, $0.06
Should a 20-person accounting firm move its files to SharePoint or Google Drive? Both agreed on SharePoint (Excel integration, finer permissions) and named misconfigured permissions as the biggest risk. Claude adopted ChatGPT's point about wrong-client exposure; ChatGPT adopted Claude's on Excel macros and timing the move away from tax season.
When to use it
Decisions where being wrong is expensive: a policy, a vendor choice, a legal reading, a strategy call. For routine document work use a job instead; it's far cheaper.
Questions people ask
Which is better, GPT or Claude?
It depends on the question, which is the point of Swarm: both answer, both critique, and you see where they agree and where they don't.
What is LLM-as-a-judge?
Using one model to grade or decide between other models' answers. Swarm's referee is a judge that took no part in the debate.
How much does a debate cost?
Usually 5 to 60 cents, shown before it runs.