← Back

How the room works

One question goes to five frontier models from five different companies. They answer alone, then read each other and argue for three more rounds, and a synthesizer builds the strongest combined answer. Every arc below is one API call, made from this browser tab.

Who is in the room

Each vendor's current flagship. You can unseat any of them before you ask — the room adapts, and the prompts count the models actually seated rather than assuming five.

What the room guarantees

Round one is blind.

No model is told that anyone else is answering, or who. Not merely kept from the others' text — unaware there is a room at all, so nobody postures, defers to whoever they assume is stronger on the topic, or leaves a gap for someone else to fill. The room is introduced in round two, together with the other answers.

Each model keeps its own conversation.

A seat's memory is its own turns plus what it was shown. It never sees its own answer played back as if someone else had said it, and it never sees a round it missed.

An answer cut off at the token ceiling is a failure, not a short answer.

It is marked failed, it does not enter that model's memory, and it is never passed to the others as though it were a position. Half a thought is worse than none, because the room would argue with it.

If a model drops out, the room carries on.

A rate limit or an outage costs you that seat for that round, not the run — the room says who was missing and carries on with the rest. If fewer than two models are still answering there is nothing left to argue with, so it goes straight to the synthesis.

When something breaks, you can hand over the evidence.

A failed seat offers Copy diagnostics: the request shape, the HTTP status, the finish reason, and the provider's own error text — including whether the stream died before the model started or part-way through an answer. It is built here in the page and copied to your clipboard. It contains no API key and none of the model text.

The synthesizer reads the record, not its own side.

It gets fresh context and the whole transcript, and is told not to take a majority vote or average anyone into blandness. It is deliberately not the model that wrote this page.

With all five seated, a four-round run is twenty-one API calls and costs around five dollars, billed to you by OpenRouter — most of a run's cost is prompt tokens, because every model re-reads the others each round. Quick mode is one round plus the synthesis, for about a fifth of that. Unseat anyone you do not want in the room.

Ask the room