Why we built this
The starting point was a habit that had become annoying. Ask one AI model a question you actually care about, get a confident answer, then open a second tab and ask another — and get a different confident answer. Now you have two well-argued positions and no way to reconcile them, which is worse than where you started, because you have lost the comfort of a single opinion without gaining the clarity of a resolved one.
The obvious fix was to make the models argue with each other instead of at us. Doing that by hand means copying text between browser tabs, re-establishing context every round, and abandoning the whole exercise by turn four. So we built the thing that does the passing.
What we learned building it
Three things surprised us, and all three shaped the product.
Provider diversity matters more than model quality. Two models from the same family produce fluent, well-structured debates that teach you almost nothing, because they share training data and therefore share blind spots. Two mid-tier models from different labs reliably out-argue two flagships from the same lab. That finding is why the model picker groups by provider and why the tiering is explained rather than hidden.
Models converge if you let them. Left alone past six or seven turns, almost any pairing drifts toward hedged agreement, because agreeableness is what assistants are tuned for. This is why the platform paces turns with a deliberate reading gap and makes interrupting easy — the pause exists so you can steer, not just so the transcript is readable.
Length is not depth. Given room to write two thousand words, a model will use it. Replies here are capped at a few paragraphs. That is partly a cost decision and mostly a quality one: a tight argument is more useful to read than a long one, and the cap forces models to lead with their strongest point.
What this is not
It is not an oracle. Nothing here declares a winner, and we would not trust a tool that did. What a good debate produces is a map — what both models accepted without argument, where exactly they diverged, and which assumption the divergence rests on. Reading that map is your job, and we think that is the correct division of labour.
It is also not a fact-checker. Every model here is bounded by its training data, so a debate about anything time-sensitive will be fluent and possibly wrong. Where two independently trained models agree, that is a real signal — but it is a signal, not a proof, because overlapping training data means they can be wrong together.
And it is not the cheapest way to use one model. If you live in a single assistant eight hours a day, that vendor's own subscription is better value and comes with integrations we do not replace. This is for questions where being wrong is expensive and a second competent opinion would change what you do.
The design decisions people ask about
Why cap replies at a few paragraphs? Because an uncapped model fills the space it is given, and a debate made of five two-thousand-word monologues is unreadable. The cap also forces each model to lead with its strongest point rather than burying it in paragraph nine. There is a cost argument too — longer replies cost more — but the readability argument came first.
Why pause between turns? A debate that arrives all at once is a wall of text. The gap scales with the length of the previous message, roughly matching reading speed, and it exists so you have a window to interrupt. Interrupting is the single highest-value action available in a multi-model conversation, and a UI that gives you no opening to do it wastes most of the format.
Why summarise older context instead of truncating it? Rebuttals fall apart when a model is working from a paraphrase of the thing it is supposed to be arguing against. So the recent messages pass through verbatim and only the older portion is compressed. The model always has the exact wording of what it needs to answer.
Why three models maximum? Past three, transcripts become genuinely hard to follow and each round costs proportionally more while adding less. Two produces the cleanest argument. Three is where a participant can reframe the question rather than just pick a side, which is the most valuable single move in a debate. Four has not justified itself in testing.
Why can you switch styles mid-conversation? Because the right structure changes as a discussion develops. Structured mode is better for establishing positions; Free Talk is better once the disagreement is live and you want to see what emerges. Forcing that choice up front, before you know what kind of argument you have, is the wrong time to ask.
Who this is for
The people who get the most out of it share a pattern: they have a decision that is expensive to get wrong, they have already formed a view, and they would genuinely act differently given a strong counter-argument. That last condition is the one that matters. If you would not change course regardless, a debate is entertainment.
Concretely, the recurring uses are stress-testing a decision before committing to it, learning a contested topic faster than a balanced summary allows, and rehearsing an argument against the hardest available opponent before presenting it to a human one.
It suits people who are comfortable being told they are wrong by a machine, and it frustrates people looking for confirmation. We would rather be clear about that than sell it to the second group.
What we got wrong
Two assumptions did not survive contact with real use, and correcting them changed the product more than any feature did.
We assumed more models would mean better answers. It does not, past three, and the curve is steeper than expected. A fourth participant adds words, cost and confusion while rarely adding a position the first three would not have reached. What actually improves an answer is provider diversity among a small number of participants — not participant count.
We assumed people would want a verdict. Early on we considered having a model summarise who won. It was a bad idea. A summary flattens exactly the disagreement you came for, and a declared winner invites you to stop thinking at the moment you should start. The transcript is the product; a verdict would have been a worse product that tested better in a demo.
Where this goes next
The obvious direction is more models, and we will keep adding them as they ship — but the interesting direction is better steering. Almost all the value in a multi-model discussion comes from the moment a human intervenes, and intervening well is a skill most people have not had a reason to develop. Making that easier is worth more than another provider on the list.
The second direction is honesty about uncertainty. Models are uniformly confident regardless of whether they are on solid ground, and a debate inherits that. Surfacing where participants are guessing versus where they are recalling something well established would change how much weight a reader gives each turn.
How we write about it
Every page on this site includes a section on where the approach falls short. That is deliberate. A tool that only publishes its own strengths is asking you to do the skeptical work yourself, and the whole premise here is that skeptical work is worth paying for.
We verify model names, pricing and capabilities against provider documentation before publishing, because that information changes monthly and stale specifics are the fastest way to lose a reader's trust. When something is uncertain, we say so rather than rounding it into confidence.