Deliberation.

What the words mean, which of three paths your question actually takes, and what the pipeline really does — measured across 4,004 discussions rather than described.

The words, in order of appearance

Every other page in this library uses these terms. This is where they are defined, and every other use links back here.

Classification

Before anything is spent, the question is scored for difficulty, depth and task type. This is the same step POST /v1/estimate exposes for free, and it decides everything that follows. Roughly 0.9 seconds.

Seat

One model, answering independently, with no sight of what the others said. A panel is normally three seats, drawn from different provider families so they do not share a failure mode. Independence is the point: three seats from one family would agree more and mean less.

Critique

A seat reading the other seats' answers and arguing with them. This is the step that separates deliberation from an ensemble — an ensemble collects answers, a deliberation makes them contend.

Round

One full pass of seats and critique. A second round happens only when the first produced real disagreement worth resolving.

Judge

A separate model scoring the seats against each other and deciding whether the round made progress. It also writes the directed prompts that tell round two what to actually go and settle.

Convergence

Whether the seats ended up agreeing. Reported on every response as quorum.converged. False is information, not failure — it is the signal that your question is genuinely contested, which is often the most useful thing a panel can tell you.

Synthesis

The single answer written from what the panel produced. Not a vote and not a concatenation — one model writing the resolved answer, with the disagreement resolved rather than averaged away.

Three paths, not one

“Deliberation” is not a single behaviour. The classifier picks one of three, and which one you get is the main thing driving both your bill and your wait.

PathWhat runsWhen
parallel_best Seats answer, best one wins. No critique, no judge. Light question, low difficulty, low hallucination risk — and you did not ask for more.
synthesis_only Seats answer, one synthesis pass. No critique round. Light question but you selected Foundation or Frontier; or a question the panel cannot meaningfully disagree about.
full_critique Seats, critique, judge, and up to four rounds. Anything with real difficulty, and always on Einstein Mode.

The express threshold is difficulty below 0.35, on a light question, where hallucination risk is not high. A Mode can widen that through express_max_difficulty — trading depth for a lower cost floor — but it can never narrow it below what the tier you selected already bought.

Your selection outranks our classifier

Choosing Foundation or Frontier guarantees at least a synthesis pass no matter how easy the question looked. You paid for deliberation; the classifier does not get to decide you did not need it.

What actually happens, measured

Across 4,004 real discussions on 2026-08-20. This is the section most products would leave out.

Rounds actually run

1 round87.9%
2 rounds3.8%
3 rounds1.7%
4 rounds5.9%
5 or more0.7%

Seven to eight questions in eight are settled in a single round. That is not deliberation failing to fire — it is the loop stopping the moment a round stops making progress, which is what keeps the median at 28.5 seconds instead of two minutes. But it does mean that if you are paying for a panel on easy questions, most of the time you are paying for three independent answers and a synthesis, not an argument.

The four-round group is larger than the three-round group because four is the ceiling for Modes that always deliberate, so genuinely hard questions pile up against it rather than stopping naturally.

How often each role fires

Share of discussions in which each role appears at least once.

seat 197.7%
seat 391.1%
seat 287.5%
convergence79.9%
maestro67.3%
seat 3 critique66.9%
seat 2 critique64.1%
judge47.8%
seat 1 critique37.3%

Seat 1 critiques far less often than seats 2 and 3 — a real asymmetry in how the critique step is dispatched, not a rounding artifact. It is on the list to explain properly rather than smooth over.

What it costs you in time

StepTimeSample
Classification only~0.9 s10 prompts
Full deliberation, p5028.5 s3,845 deliberations
Full deliberation, p9088.2 s
Full deliberation, p99229.3 s

Against a single frontier model measured on the same harness — median 9.0 s — that is roughly three times the wait, for roughly two to three and a half times the cost. Those are the honest multiples, and they are the ones the About page quotes. A bare comparison against a notional one-second call would flatter nobody and mislead everyone. That trade is the product. Where deliberation thrives is the page about when it is worth taking, including when it is not.

Does it work

On our own published measurement, the deliberated answer is preferred over a single frontier model 63.9% of the time — n=577, ±3.9 at 95% confidence, across four batches, blind pairwise judging.

Read that honestly: a win rate near 64% also means that about a third of the time, the single model was judged at least as good. Deliberation is a favourable bet on questions that deserve it, not a guarantee on every question. The full protocol and the runs behind it are on the evidence page.