It reads what you meant.
A lead model turns your question into a brief — the real question, the context you left implied, and what would actually help. Vague questions get vague answers everywhere else; this is the step that fixes that.
Ask something that actually matters and four frontier models will argue about it from ten angles at once. Point the council at a decision and it tells you where they agreed, where they genuinely didn't, and what to do. Point it at an open question instead and it does the opposite — ten thinkers pulling in different directions, and none of them reconciled for you.
One question in, four models out. This part is the same whichever panel you pick; what changes here is how hard they work on it. Switch between the two and watch the shape change.
One sitting per model.
Each model is asked once and writes all ten roles in a single answer, so its contrarian has already read what its own analyst said. The ten views hang together as one coherent take — which is exactly why they agree with each other more than ten separate advisors would.
Best for most questions. Fast, cheap, and the coherence is often a feature.
A lead model turns your question into a brief — the real question, the context you left implied, and what would actually help. Vague questions get vague answers everywhere else; this is the step that fixes that.
Four models from four different labs answer that brief in parallel, each one wearing all ten hats. Forty perspectives, each committing to a position and a confidence score, all landing at once.
The lead reads all forty. On Review that becomes one answer — the agreement, the real dissent, what to do. On Explore it stays deliberately open: the distinct directions, kept apart. Either way the tally is counted in code, not guessed at by a model.
Every model answers as all ten. Which ten depends on what you came for — a council can weigh a decision, or it can open a question up. Switch below and watch who turns up.
Ten advisors weigh your question and the council tells you where they agreed, where they genuinely did not, and what to do.
analyst
data and evidence
advocate
argues your corner
optimist
what could go right
contrarian
what you're missing
sage
precedent and patterns
lawyer
liability and ethics
investor
money and returns
engineer
can it be built
strategist
timing and position
psychologist
how you'll actually feel
Every role commits to one of
Counted in code, not guessed at. When models picked for disagreeing agree anyway, that means something.
What comes back
One answer: where they agreed, where they genuinely didn't, and what to do about it.
Where they agreed · Where they split · What to do
A council is only worth more than one model if its members can be wrong in different directions. That is a property you have to engineer and then keep checking, so this is how the panels are picked and how the checking works.
Two models from the same lab are the most correlated pair you can build a panel from — in testing, two Claude models tracked each other at 0.79 where cross-lab pairs sat nearer 0.2. Every seat is a different company.
Training corpus, alignment norms and cultural priors travel with geography, and they are what make two models wrong in the same direction. Both panels deliberately seat labs from more than one country.
Each run scores how much the panel actually split — stance agreement and confidence correlation between every pair of models. You can see the number and the pair-by-pair breakdown on your own sessions.
what the measurement is for
It means no two models that will simply agree with each other ever share a panel. Different lab and different country is where we start, not proof — in our own testing we have found pairs from separate labs on separate continents whose confidence tracked each other perfectly, a 1.00 correlation across every role. A model like that adds another voice to the room without adding a second opinion, so it does not get a seat.
Models change, and they drift toward each other as they train on more of the same world. Measuring every session is how the panels stay genuinely independent rather than just looking it on paper. It keeps us honest in the other direction too: a low score on any one session can simply mean the question had an easy answer, so it is the trend across many that means anything.
The questions where you'd normally text four different friends and get four different answers.
“Should I raise a seed round or bootstrap?”
“What could this product be in three years?”
“Do I take this offer or negotiate?”
“What angles on this story has nobody taken?”
Buy credits, spend them per session. No subscription, nothing to cancel, and credits never expire.
For everyday decisions.
round table · sealed $0.07 – $0.18
Fast, cheap, and honestly good enough for most questions.
For the ones that matter.
round table · sealed $0.48 – $1.50
Frontier models. Slower, pricier, noticeably better at nuance.
per session, charged on completion · 40 perspectives either way
Because one model has one set of blind spots, and it will never tell you where they are. Four models from four labs disagree in useful ways — when three of them flag the same risk you can take it seriously, and when they split, that split is usually the real answer to your question.
You buy credits and spend them per session. Round table costs about $0.03 – $0.07 on Medium and $0.19 – $0.60 on High; Sealed asks every role separately, so it runs about two and a half times that — $0.07 – $0.18 and $0.48 – $1.50. Turning Research on adds roughly $0.06 – $0.11 to a Medium session and $0.12 – $0.25 to a High one. You're charged when a session finishes, based on the models it actually used, so a short question genuinely costs less than a long one. Credits never expire.
What you want out of it. Review puts ten advisors on your question — the analyst, the lawyer, the investor, the one who tells you how you'll really feel in six months — and gives you a verdict: where they agreed, where they split, what to do. Explore swaps in ten different thinkers — a cartographer, an outsider from another field, an inventor told to prefer the idea that might not work — and refuses to reconcile them. Agreement is the signal in one and the failure in the other, so the same disagreement score reads in opposite directions. Review is the default; Explore is a button next to it.
Only if you ask it to. Turn on Research and the lead works out which facts your question actually turns on, sends those to a model with live web access, and folds a short briefing into what every advisor reads. You get the sources, so you can check them. Leave it off and the council answers from what it already knows — which is the right call for questions about your own situation, where there is nothing to look up.
Yes. The run happens on the server, not in your browser. Close the tab, reload, open it on your phone — you'll pick up exactly where it got to, mid-sentence if that's where it is.
Sessions are private by default and only you can see them. If you want to share one you can flip it to a public link, but that's always something you choose.
Treat it as a very well-read friend who has thought about your problem from ten angles — not as a lawyer, a doctor or a financial adviser. It's genuinely useful for thinking. It is not accountable to you, and it can be confidently wrong.
Bring the decision you've been going round in circles on. Four models, ten roles, about a minute, and a couple of pennies.
Ask the council→