Published August 23, 2026
Secret Chat AI has a new headline feature, and it changes what "asking an AI" means in the app. The AI Council sends one question to several top AI models at the same time — up to all four of ChatGPT, Claude, Gemini and Grok, one from each company — shows you each answer the moment it arrives, and then has a referee model read all of them and report, claim by claim, where the panel agrees and where it contradicts itself. It is now what the app opens on for new users, and its smallest size is free within the daily quota.
This article is the story behind it: the problem it exists to solve, how a council actually runs, what it costs, and — because this is Secret Chat — an honest account of what it does to your privacy and what it cannot promise.
The Problem: Every Chatbot Grades Its Own Homework
A language model delivers a fabricated answer in the same confident voice as a solid one — nothing in the fluency, the structure or the tone tells you which you just received. And asking the model to double-check itself is not a fix: with no new information to work from, it tends to stand by what it already said, and research on this kind of intrinsic self-correction finds it unreliable — sometimes it makes answers worse.
People who use AI seriously have already worked out the countermeasure: open a second chatbot, paste the same question, and compare. It works — and it is tedious enough that the habit rarely survives contact with a busy day. Four tabs, four logins, four answers to read line against line, with the hardest part left entirely to you: spotting which specific claims the answers quietly disagree on. Two fluent, well-structured replies can contradict each other in one date, one number, one "always" that should be a "sometimes", and a tired reader skims right past it.
The vendors themselves are no help here. ChatGPT will let you regenerate an answer or switch between OpenAI's own models; Claude and Gemini likewise. Putting a rival lab's answer on the same screen as their own is a comparison none of them has any incentive to offer — which makes it exactly the feature a multi-model aggregator should.
What a Council Run Looks Like
You ask once. The panel's models all receive the question in parallel, and each answer appears in its own tab the moment that model finishes — you are usually reading the first reply while the slowest member is still writing, so a council takes about as long as its slowest model, not the sum of the panel.
When every answer is in, the referee goes to work. It pulls the factual claims out of the answers and builds a disagreement table: for each claim, which models assert it, which contradict it, and which simply never mention it. From the claims that at least two models support, it writes a short synthesis; whatever stays contested is flagged separately, under "verify before acting", so the doubt is impossible to miss. Each answer's credit cost is printed right next to the model that wrote it, with the referee's small step on its own line.
The layout is deliberately bottom-up: answers first, then the table, then the summary. A verdict you read before the material behind it is just one more confident paragraph — the whole point is that you can see what the conclusion rests on.
It Gives You Time Back, It Does Not Cost You Time
The obvious objection to a panel is the clock: surely three answers take three times as long to get, and three times as long to read. Neither half holds. The models are all asked at the same moment rather than one after another, so the wait is set by the slowest member of the panel and not by the sum of it — and since each tab fills as its own model finishes, you are reading long before the last one lands. Compare that with doing the check yourself, which is the genuinely slow version: a tab per model, the question pasted again into each, and every wait starting only once the previous one has ended, with the tally of what each answer said held in your head throughout. To be plain about the comparison that matters: a council is not quicker than putting the question to a single model — nothing with four models in it could be — but it is much quicker than the four separate conversations it stands in for.
The reading is where more of the time comes back. Several independent answers to one question are several screens of text that mostly agree, and the part that earns your attention is the sentence or two where they do not — which is precisely the part a tired reader misses, because each model phrases it differently and buries it in a different paragraph. Doing that comparison is the referee's entire job. It extracts the factual claims, lines them up model by model, and returns a finished conclusion: a short synthesis assembled only from what at least two models arrived at independently, with everything still contested set aside under "verify before acting". What you actually have to read at the end of a council is therefore shorter than one chatbot's answer, not longer. The full replies stay above it, unread unless you want them — there for the moments when you would rather check the summary than take it.
Why One Model From Each Company
Every seat on a council belongs to a different provider, and this is a rule rather than a coincidence. Two tiers of the same family agreeing with each other is weak evidence — they were trained by the same lab, on overlapping data, with shared blind spots. Separate companies converging on the same claim is a meaningfully stronger signal — still not proof, as the limit section below spells out — and separate companies contradicting each other is precisely the warning you paid to see.
The three sizes only change how many companies are at the table:
- Duet — two models plus the referee. You pick the pair; paying users default to ChatGPT + Claude, and there is a free pair (more below).
- Trio — three models plus the referee, any three of the four. The size to reach for when a pair splits one against one: a third seat shows you which way the panel leans — a lean, not a verdict, since the two agreeing models can still share an error.
- Quartet — the whole bench: ChatGPT, Claude, Gemini and Grok. Nothing to configure, because leaving one out would mean leaving out the one that might have disagreed.
Every size runs the same referee and returns the same disagreement table. A bigger council buys more companies at the table, never a different class of analysis.
It Searches the Web When the Question Deserves It
Four models answering a current-events question from memory would just be four training cutoffs nodding at each other. So before a branded panel starts, the question is read once: if the answer turns on anything current or checkable — news, prices, laws, versions, who holds which job — every model on the panel answers with its own provider's live web search attached, and the referee ends up comparing what they found rather than what they remember. If the request is creative or reflective — rewrite this, translate this, help me think — nothing is looked up and the panel runs on faster, cheaper models instead, so you are not billed for research that would add nothing.
That decision is made once, for the whole panel, never per model. A council is only readable when its members are comparable; if one had searched and another had not, their disagreement would be about the setup, not about your question.
The Free Council
The Duet runs free within the ordinary daily quota, on a lighter pair: DeepSeek Flash and a fast GPT tier. Both answer from memory — that pair is never web-grounded — but the mechanism is the full one: two independent models, the referee, the claim-by-claim table. You can watch two AIs contradict each other about your own question, and have an answer from anywhere checked, without paying anything. Credits buy the branded panels and the bigger sizes.
Checking an Answer You Already Have
The council's second job may be its best one: it does not need to start from a question. Paste in an answer you were given somewhere else — by ChatGPT, by a colleague, by a confident stranger on a forum — and ask the panel whether it holds up. Each model works through the text independently, and the referee lines their verdicts up so that a claim two models both doubt is flagged in the table rather than buried in two long replies.
It also works mid-conversation. Ask a single model something, then switch that same chat to a council and ask "is that right?" — the panel reads the whole conversation, so the answer under review is the one already on your screen, with nothing to copy across. One cheap model to draft, a council to check the parts that matter, is often the most economical way to use the whole app. And when the referee leaves specific claims unresolved, the app offers to carry exactly those claims — not the whole question — into a deep research run with live sources; the suggestion lands in your message box for you to read and send, never fired on its own.
What It Does to Your Privacy — Said Plainly
Here is the part a marketing page would whisper: a council widens where your question travels. Instead of one provider, the text you type reaches every model on the panel, and the referee model additionally reads the question alongside the panel's answers. Every one of those requests leaves through the same gateway as any other Secret Chat message — anonymously, with no account, name or IP attached, never used for training — and the finished check is stored only in your own browser, like the rest of your chats. The per-message Session Privacy Report lists every model that took part in the turn, not just the one whose answer you liked.
But anonymization protects who you are, not what you typed. Four providers reading your question verbatim is four providers reading it — so the standing advice applies with more force on a council, not less: strip names, case numbers and identifying details before you ask. Secret Chat removes you from your queries; it does not remove the data from your messages.
The Limit We Will Keep Repeating
Agreement between models is evidence, not proof. The major models are trained on much of the same public data and share failure modes, and research on cross-model comparison keeps finding correlated errors even between different architectures and providers — so a unanimous panel can still be unanimously wrong. What a council gives you more reliably is the opposite signal: when the models disagree, you have learned, cheaply and early, that a claim needs checking before you lean on it — and the table tells you which claim, which is the part that used to be guesswork. The referee deserves the same skepticism: it is a language model too, and it can misread a paraphrase as a contradiction or miss a real one — which is why every answer stays on screen above the table, so the comparison is there to be checked rather than taken on faith. Treat the synthesis as "what the panel converged on", and for medical, legal or financial decisions, use the council to decide which claims to bring to a professional, never as the professional.
Try It in Under a Minute
- Open the Secret Chat app — the AI Council is the default selection for new users, so it may already be in front of you.
- Pick a size: the free Duet is preselected if you have no credits, and the size picker offers the Trio and the Quartet with your choice of companies.
- Ask something you actually care about being right — or paste in an answer you have been given and ask the panel to check it.
- Read the answers as they land, then the table: the rows where models split are your to-do list.
The full reference — every size, panel, grounding rule and cost detail — lives on the AI Council feature page.
Frequently Asked Questions
- What is the AI Council on Secret Chat?
It is a mode where one question is sent to several top AI models in parallel — up to ChatGPT, Claude, Gemini and Grok, one from each company — and a referee model then compares their answers claim by claim, showing where they agree, where they contradict each other, and what stays unsettled.
- Is the AI Council free?
The two-model Duet is free within the daily free quota, running on a lighter pair (DeepSeek Flash and a fast GPT tier) with the same referee and the same disagreement table. Credits buy the branded pairs and the larger Trio and Quartet panels.
- Why not just ask two chatbots myself?
You can — the council automates exactly that workflow. It runs the models in parallel instead of one after another, and the referee does the part humans skip: extracting the individual claims and checking them against each other, so a one-word contradiction between two fluent answers does not slip past.
- Does a council take longer than asking one model?
Longer than one model, yes — a panel finishes when its slowest member does. But it is not the sum of the models on it: they are all asked at the same moment, not one after another, and each answer appears as that model finishes, so you start reading before the last one arrives. Against the manual alternative it saves time twice over: you are not pasting the same question into several tabs and waiting through each in turn, and the referee returns a ready conclusion — a short synthesis of what at least two models independently agreed on, with the contested claims listed separately — so the thing you have to read at the end is shorter than a single chatbot's answer rather than several times it.
- Does a council search the web?
On the branded panels, yes, when the question calls for it: if the answer depends on current or checkable facts, every model on the panel runs its own provider's live web search, and the decision is made once for the whole panel so the answers stay comparable. The free pair always answers from memory.
- Can it check an answer I already have?
Yes — paste the answer into a council and ask whether it holds up, or switch an existing chat to a council and ask "is that right?": the panel reads the whole conversation, so the answer being checked is the one already on your screen.
- Is a council private?
Each request goes out like any other Secret Chat message — anonymized, with no account, name or IP attached, and never used for training — and the result is stored only in your browser. Note that a council sends your question to every model on the panel plus the referee, so keep identifying details out of the question; anonymization protects who you are, not what you type.
- If all the models agree, is the answer correct?
Not necessarily. The major models are trained on much of the same public data and share blind spots, so they can be confidently wrong together — agreement is evidence, not proof. The more dependable signal runs the other way: disagreement tells you exactly which claims need verification before you rely on them.