Blind spots surface
Different models catch different things. A risk one model glosses over is often the first thing another one names.
// How it works
What happens in the thirty seconds between asking and the verdict, step by step, using one real question.
An LLM council is a setup where several large language models from different AI labs answer the same question independently, and another model weighs their answers into one result. Instead of trusting one model’s view, you see where they agree, where they split, and why.
Think of it as a panel of advisers rather than one. Each model was trained on different data with different methods, so each has different strengths and blind spots. Council turns that difference into something useful: a verdict you can act on, with the disagreement left in view.
Council gives the models fixed jobs instead of letting them all answer the same way: one argues for, one argues against, one works out how to improve the idea. A chairman from a fourth lab then picks the strongest case and commits to an answer, with a confidence level, and every answer shows exactly which model replied.
Different models catch different things. A risk one model glosses over is often the first thing another one names.
When models trained by different labs agree, you can lean on the answer more. When they disagree, you know to check.
Each seat has a job, so you always get the case for, the case against and a way forward, not three versions of the same reply.
Every verdict says high, medium or low. A coin flip is labelled as one, instead of sounding as sure as a fact.
// One question, start to finish
The question
Should I leave my job to work on my startup full time?
01
The gate
About 2 seconds
Before anyone argues, Council asks one thing: does this question have a single correct answer? If it does, you get that answer in one line and a nudge, because a debate would be theatre.
Leaving a job has no single right answer, so the panel is convened.
A fact
What is the capital of Australia?
Canberra. One line, no debate, and a reminder of what Council is for.
A judgment call
Should I leave my job to work on my startup full time?
No single right answer. The council convenes.
02
Three roles
About 15 to 25 seconds
Each seat runs on a model from a different AI lab, so their blind spots differ. The case for is argued by GPT (OpenAI), the case against by Claude (Anthropic), and how to improve it by Gemini (Google). If one is busy, a free backup from another lab steps in.
Each member gets one job and answers alone, in parallel, without seeing the others. That independence is the point: a model that can read the other answers starts agreeing with them.
The case for
GPT (OpenAI)
You already have a working product and a few paying users. Full focus buys speed, and speed is what an early startup lacks most.
The case against
Claude (Anthropic)
Six months of savings is shorter than it sounds. First products usually need more than one change of direction, and each one costs time you would be paying for.
How to improve it
Gemini (Google)
Ask for a four-day week for three months. If revenue covers half your salary by then, leave. If not, you kept the job and learned why.
Illustrative answers, shortened.
03
The chairman
About 5 seconds
A fourth model from another lab, Qwen (Alibaba), reads all three answers. It decides which seat made the strongest case for this person, names where the members disagree, and commits to one answer with a confidence level.
The verdict
Do not resign yet.
Strongest case: the case against. Where they disagree: how much runway is enough. The answer: move to a four-day week, set a revenue target for three months, and let that number make the call. Medium confidence.
04
Verification
Planned for Plus
On Plus, if the verdict leans on something checkable, like a statistic, a date or a rule, that claim will be looked up in an outside source and marked supported, contradicted or not confirmed. Opinions cannot be checked, and Council will say so instead of pretending.
Verification is switched off on the free preview so answers stay fast.
// What you get back
You asked
Should I leave my job to work on my startup full time?
1The case against
The Critic
Answered by Claude (Anthropic), via Anthropic API
Dealbreaker if: you have dependants and no fallback income.
The verdict5Medium confidence
2Why listen to the case against
It is the only answer that deals with the money, and the money decides how many tries you get.
3Where they disagree
How much runway is enough before leaving.
4The answer
Do not resign yet. Go to four days a week, set a three-month revenue target, and let that number decide.
Each seat’s answer in full, with the model that wrote it, so you can read the disagreement yourself.
Which seat the chairman weighed most, and the reason in two sentences.
The split in one sentence. Often the most useful line on the page.
One decision you can act on, including the first step.
High, medium or low. A coin flip is labelled as a coin flip.
The exact model behind each seat and the chairman, shown at the top of its answer as reported by the provider that served it. Nothing on the page is a claim we cannot back up.
// The panel
The case for it. Argues the strongest honest version of yes, so the best case gets made properly rather than assumed.
The case against it. Finds what breaks before you do. This is the seat most single-model answers are missing.
How to make it better. Turns the argument into a next step instead of leaving you with two opposed opinions.
// Behind the verdict
Reads all three, picks the strongest reasoning and issues one verdict with the reasoning attached. It runs on a different model from the three it judges.
A question with one right answer gets a single line and an honest nudge. It stays silent when there is nothing to add.
Planned for Plus: factual claims inside the verdict get checked against an outside source. Not live in the free preview yet.
// One model or a council
| A single model | Council | |
|---|---|---|
| Viewpoints | One | Three, from different labs, plus a chairman |
| The case against | Rarely made unless you ask | Always made: it is a seat |
| Disagreement | Hidden in one confident voice | Shown under the verdict |
| Confidence | Same tone for everything | High, medium or low |
| Which model answered | Usually one you chose | Shown on every answer, with the exact model |
| Speed and cost | Fast and cheap | Slower and dearer: use it when it matters |
// When to use it
Job offers, a drop year, studying abroad, moving cities.
Pricing, launches, hiring, which customers to chase first.
Rent or buy, loans and EMIs, where to put savings. A structured second opinion, not financial advice.
Frameworks, databases, build or buy, where the honest answer is “it depends” and you need to know on what.
You asked one AI and want to know how much to trust it.
// Who sits on the council
GPT argues the case for. Free backup: gpt-oss, OpenAI’s open-weight model.
Claude argues the case against.
Gemini works out how to improve it.
Qwen chairs, from outside the three labs it judges.
Llama steps in as a free backup when a seat is busy.
GPT and Claude are paid models used up to a daily budget; after that, the free backups answer. Every verdict lists exactly which models replied.
// Design decisions
For, against and how to improve it covers the ground a decision needs. More voices add cost and noise, not new angles.
A model asked to check itself mostly agrees with itself. Models trained by different labs have different blind spots, so one catches what another misses.
Members never see each other’s answers. If they could, the later ones would drift toward the first, and the disagreement you came for would disappear.
The judge is chosen from outside the panel whenever possible, so no model ends up grading its own argument.
// What we hold to
Council principle
“It knows when to shut up.”
Ask something with one correct answer and you get one line, not a staged debate. You are not billed for theatre.
Council principle
“It refuses to fake certainty.”
The same confident tone for a settled fact and a coin flip is the problem Council exists to fix.
Council principle
“Every question has three sides.”
For it, against it, and how to make it better. The verdict comes with the split left in view.
// Common questions
Then the verdict says so. Three differently trained models landing in the same place is real evidence, and the confidence goes up with it.
The main panel is GPT (OpenAI) for the case for, Claude (Anthropic) for the case against and Gemini (Google) for how to improve it, with Qwen (Alibaba) as chairman. GPT and Claude are paid models, so each day they answer up to a set budget; after that, or if a model is busy, free open models such as gpt-oss (OpenAI) and Llama (Meta) step in. Every answer is labelled with the exact model its provider says answered it.
Not yet. Choosing your own panel is planned for Plus. The models page shows the panels we plan to offer for each type of question.
About 30 seconds for a full debate, around 10 for a quick take, and a few seconds for a fact.
No. It is a structured second opinion. For legal, medical or financial decisions, use it to get clear on the trade-offs and on the questions to ask a professional.
Not yet. Memory, so Council can draw on decisions you have already made, is the next thing on the roadmap.
// Keep reading
See GPT, Claude, Gemini, DeepSeek and more side by side.
Read →Every model, its price, context and where it sits in Council.
Read →Connect the council to Claude, Cursor or Windsurf.
Read →What is free now, and what each plan will include.
Read →