IT WRITES THE THING
Large refactors, type discipline, builds that pass first time. When the answer has to run, this is the seat that makes it run.

One request reaches every seat in turn. Each takes the part it leads on and leaves the work better than it found it. What comes back is one answer, with the reasoning underneath it.
Four of five signed off. The fifth disagreed, and that disagreement ships with the answer.
Five models answering the same question in parallel and averaging the result produces the blandest of the five. So they do not work in parallel. Each seat receives the work so far and improves it in the one dimension it leads on — and may rewrite anything earlier if it can genuinely make it better.
The hand-off carries the reasoning, not just the conclusion. That is the mechanism by which several models beat one: each seat inherits the thinking of the seat before it.
Occasionally a seat will not put its name on the result. When that happens it is recorded and shown beneath the answer, rather than averaged away. Most runs have no disagreement, and that is the point — one that always disagrees means nothing.
The seats are shown by capability rather than by model name. The roster changes as models are released, and a name we have not verified is a name we will not print.
Large refactors, type discipline, builds that pass first time. When the answer has to run, this is the seat that makes it run.
Retrieval, citation, and then the source behind the source. It comes back with what the others would have asserted from memory.
Long chains, contested premises, trade-offs with no clean answer. It is the seat that refuses the shortcut the other four would take.
Naming, tone, direction. The part a committee averages into nothing, held by the one seat willing to be disagreed with.
Screenshots, whiteboards, scanned pages, the photograph of a napkin. The seat that can look at what you actually have.
One request. Every seat in turn. They argue, hand work between them, and return one answer — with the disagreement on the record.
Three things about your work, each of which the architecture actually enforces — not a policy we ask you to take on trust.
Each one is named on the sub-processor list, with what it receives and which country it processes it in. We make no claim about whether they train on it — their terms differ per model and change without notice, so it is not a promise we could keep.
You see which model wrote which part, what it was given, and where it would not sign off. Nothing is merged silently.
The row and everything attached to it are removed. There is no hidden copy marked deleted.
Plans differ on how many seats work on a run, how many runs you get, and whether UltraZeta depth is available. They do not differ on model quality — a Starter customer gets the same models as everyone else.
Three seats on every run.
Five seats, and the whole trace.
For teams running all day.
Prices in US dollars, billed monthly. Unused runs do not carry over — the refunds page says exactly what happens when you cancel.

Let them work together.
You sign in with a link sent to your email. There is no password of yours for anyone to take from us.
Every plan is paid, so a card is required at signup and the first payment is taken before you can run anything. Nothing here is free.
Cancelling stops the next payment. You keep the period you have paid for.