Questions people actually ask.

They solve different failure modes, and the distinction is worth getting right before deciding whether you need this at all.

A reasoning model thinks longer. It spends more compute working through a problem step by step before answering, and it is genuinely good at catching its own slips — a dropped constraint, an arithmetic error, a step that does not follow. What it cannot do is step outside its own training. More thinking applied to the same priors returns a more thorough version of the same point of view, stated with more confidence.

Deliberation adds a second opinion, not more time. Quorum puts a question to several models from different labs, trained on different data with different conventions, and has them read and challenge each other. When they disagree, that disagreement is information a single model could not have produced — it came from outside its distribution.

So the honest split: reasoning is the better tool when the answer is hard to compute — a proof, a tricky refactor, a constraint problem with one right answer. Deliberation is the better tool when the answer is hard to be sure about — judgement calls, trade-offs, ambiguous requirements, anything where being confidently wrong is expensive.

They are not competing. A reasoning model sitting in a Quorum seat is still a reasoning model, and it thinks just as hard — it simply has to defend its conclusion to two others who worked the problem independently. The receipt on every answer shows where they agreed and where they split, so you can see which of the two you actually needed.

They are a per-question choice, not a plan. Your plan decides which of them you can pick; the selector next to the composer decides what happens on this question. It controls two things at once, which is why the difference is bigger than the names suggest.

Which engines answer. Every plan has a ceiling on how capable an engine it can reach. Express spends that ceiling only when a question needs it — easy questions run on cheaper engines, harder ones escalate automatically. Frontier stops the downgrading: every question runs at your plan’s full ceiling whether it needs it or not.

Whether the panel deliberates. On a question our classifier reads as light, Express skips the deliberation round and simply returns the strongest single answer — three models agreeing at three times the price is not worth paying for. On anything it does not read as light, Express deliberates anyway. Foundation and Frontier remove that judgement call: they always run at least a synthesis pass over the panel’s work.

So: Express is adaptive and cheapest, and is the right default for most work. Foundation guarantees deliberation when you want the panel’s reasoning rather than its fastest answer. Frontier guarantees both the deliberation and the best engines you have access to, for the questions where being wrong is expensive.

Dialling down is deliberate and supported — a Pro subscriber can run Express all day and only reach for Frontier when it matters. The receipt on every answer shows which engines actually ran and whether a deliberation happened, so you can check what you got.

Genuine disagreement is a function of model diversity, not just model count. Quorum selects engines with meaningfully different training backgrounds, architectural approaches, and reasoning styles. When they converge quickly, that convergence is itself signal — it means the answer is robust. When they diverge, Quorum surfaces the specific points of disagreement so you can see exactly where the reasoning differs — visible round by round in Panel. The deliberation is transparent, not a black box.

A mode is a saved configuration of how Quorum deliberates: which model families take which seats, how much internal critique runs, how many rounds, and what the whole panel is tuned for. Every answer you get already runs inside one — Standard by default, or Einstein Mode on Pro. Anyone signed in can build their own, keep it private, share it with teammates, or publish it.

Each mode in the gallery shows its real, auto-derived ingredients — which providers it draws from, how many seats, how much critique, and its per-query price — not marketing copy written by whoever made it. Read the ingredients the way you'd read a nutrition label: they can't overstate what the mode actually does, because they're computed from the config, not typed in by the creator.

Standard is Quorum's default — tuned for a fast, cost-efficient deliberation on everyday questions. Einstein Mode runs the highest tier of reasoning: more seats, deeper critique, more rounds when the question calls for it. It's unlocked on Pro.

Per call, it is roughly two to three and a half times a single frontier model — measured on our own runs, not estimated. The counterargument is about total rounds rather than per-call price: token costs scale superlinearly as a conversation grows, so round six with one model is far dearer than round one, and front-loading the disagreement should reach a defensible answer in fewer rounds overall. Be clear about which half is evidenced: the per-call premium is measured and published; the round-count saving is reasoned and not yet measured. Berkeley AI Research has identified compound AI systems as one of the most impactful architectural directions.

No. Quorum does not use your inputs or outputs to train AI models, and does not use your content for cross-user learning of any kind. Quorum accesses AI providers via enterprise API agreements — not consumer chat products — which include contractual no-training commitments. This provides stronger protections than standard consumer AI interfaces. Full details are in our Privacy Policy and Terms of Service.

Quorum is model‑agnostic by design: new models are evaluated and take seats without changing your workflow. Honestly, though — keeping that current is real operational work, and we have been caught with a stale assignment before. It is now something we test for rather than assume. How we test.

A public catalog of modes — some built by Quorum, others built by outside Mode Makers and published for anyone to use. Every listed mode's ingredients are real and derived, never hand-authored, and a mode can't carry a public performance claim until it passes a certification run against a fixed question set. Today there are two listings and both are ours; the builder, the test rig and the payout rails are all live and open to anyone. If you'd rather build a mode than use one, start here.