Perplexity just shipped a small change to a settings menu: in Model Council, you can now pick which model gets the last word. It reads like housekeeping. It's the most important knob in the system.
Model Council runs three frontier models — GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro — in parallel on the same question, then hands their three answers to a fourth model that synthesizes where they agree, where they split, and what each one alone contributed. For months that fourth model was a fixed default you never saw. Now it's a dropdown.
Easy to miss. Also a quiet admission: the seat that decides what to do with disagreement was always the one that mattered.
The thing you're actually paying for
A council's three members can disagree perfectly — surface a real edge case, catch a confident wrong answer, split on a genuine trade-off — and you will never benefit from any of it if the model that writes the final answer flattens the split into "the models broadly agree." The panel produces the raw material. The synthesizer produces the thing you actually read. Pick the wrong one and you've paid for three models to generate a confident average.
That's why making the choice visible matters more than it looks. One widely-shared explainer this month put the accuracy-and-cost sweet spot at three to seven models, with a blind peer-review round and a single synthesizer on top. Notice what's load-bearing in that recipe. It isn't the count. It's the synthesizer.
Three names for one seat
Call it what it is: the chairman. Perplexity calls it the orchestrator. Karpathy's llm-council calls it the chairman model. Same seat under three names — the one that reads everyone else's argument and decides what the room concluded.
Here's the move that gets missed: a council doesn't get its value from the models agreeing. It gets it from a chairman who knows what to do when they don't. One model has an opinion. A council has a position — but only if someone in the room is willing to write down that two members said X and one said Y for a reason worth taking seriously, instead of splitting the difference.
What a good chairman actually does
The job is not to find the average. It's to adjudicate. When three models converge, that's cheap — any synthesizer can report a consensus. The chairman earns its seat on the hard runs, where one model dissents and the whole question is whether the dissent is noise or the most important thing on the table.
This is where the choice of chairman model stops being cosmetic. A sycophantic synthesizer — one tuned to be agreeable, to smooth, to reassure — will treat a lone dissent as an outlier to be quietly dropped. A sharper one surfaces it: "two models recommend X, one argues Y because of this specific risk, and here's why that's worth your attention." Same three inputs. Opposite outputs. The only variable that changed was who held the pen.
The research keeps arriving at this from different directions. One line of work this year found that a model's confidence barely tracks whether it's right — so a chairman that defers to the most confident-sounding member is optimizing for the wrong signal. Another argued that for genuinely contested questions, collapsing the council down to a single agreed answer destroys the exact information you convened the council to get. Both say the same thing to anyone building one of these: the synthesis step is not a formality. It's where the value is created or thrown away.
Watch it happen
Ask a hard question — a real trade-off, not a fact lookup — and a weak chairman returns one tidy paragraph: "The models broadly agree that you should do X." A real one returns the shape of the disagreement: who held back, what they were worried about, and a recommendation that accounts for the dissent instead of erasing it. The second answer is longer and less comforting. It's also the one that saves you from the confident mistake.
And different chairmen suit different jobs. A careful, literal synthesizer is what you want auditing a contract. A bolder one that's willing to take a side is better when you need a decision, not a survey. Making that a choice, rather than burying it, is the part Perplexity got right this week.
It's also the part Shingikai treats as a first-class strategy rather than a hidden default. Chairperson Synthesis is a named option — the synthesis layer's whole job is to decide what to do with disagreement, not to average it away. And because the deliberation streams live, you don't have to trust the chairman blind. You watch the three models argue first, then watch the synthesis happen on top. The chairman's call is legible because the argument it's working from is, too.
The question that turned out to matter
Perplexity making the orchestrator a setting is a small interface change and a large concession at once. The industry spent two years arguing about which model is smartest. The more useful question turns out to be who gets to decide what three smart models concluded when they didn't agree.
Pick that seat carefully. It's the one writing your answer.
Try it free — no signup. shingik.ai