AI answers shape what the world believes.
We track when they move.

6 models · Balanced Try the demo
ChatGPT Claude Gemini Mistral DeepSeek Grok

No account needed. The demo replays a real run: six independent answers, then one consensus with the contradictions marked inside it.

Live model pulseSee which model answers most often earn the judge's pick in real runs. View leaderboard ›
Open the app See Watches

01 · Ask

Start from one prompt.

One field, one run switch, one send. The app opens exactly here, with no modes to learn before the first question.

What should the models cross-check?

6 models · Balanced

Please avoid personal or sensitive data. AI models can make mistakes.

  • The field. One question, nothing to configure first.
  • The run switch. How many models answer, and who writes the consensus.
  • Send. The only filled control on the page.

02 · Run

One run, narrated end to end.

Six models answer in parallel, then the consensus is written and an uninvolved model checks it for contradictions. Scroll to move the run.

Question
I have to tell a client that our launch slips by two weeks. Can I send this as it is?
Preparing the question 0:00

ChatGPT · Claude · Gemini · Mistral · DeepSeek · Grok ·

Consensus answer

Send it — but the closing line splits the models three against three.

6 models · 24 s · 3 contested passages Differences 3 Model answers 6

03 · Decide

The doubt is marked in the answer.

One answer, read top to bottom — with the disagreement written into it rather than hidden behind it. Four things carry that, and you can try each one here.

  1. 01

    How far apart they were

    Every run opens with one number: how much the six answers actually agreed. An uninvolved model works it out, so the engine that wrote the answer is not also grading it.

    52 Partial agreement 1 critical · 1 minor detail · disputed: keep or cut the line asking if the delay is a problem, 6 models compared Analysis by Gemini independent of the consensus engine

  2. 02

    Who backed each claim

    The small number behind a sentence is how many models supported it. Rest on it, or tap it, and it names them: who agreed, who deviated, and in what words.

    Try it on the 4/6.

    The new date belongs in the first line, ahead of the miss, the cause and the apology

    Name the day you will confirm the 29th, so the next update does not arrive as another surprise

  3. 03

    Where they pull apart

    Agreement gets no decoration at all — only the count. A difference in detail gets a fine rule under the sentence, a real contradiction a heavier amber one. The mark sits on the sentence it belongs to, so you never have to hold two documents open at once.

    • Say what she actually gets on the 15th no mark, just the count
    • Name the cause, or leave it vague fine rule
    • Keep the closing question, or cut it heavier amber rule

  4. 04

    What they actually said

    Click a mark and the disagreement opens: both camps, named, with the sentence each side actually wrote and a note on what is worth checking yourself.

    Try the marked sentence.

    Three models read the closing line as the only sentence that gives her a way in, three as an invitation to reopen a date you cannot move.

    Contradiction · critical Keep or cut the line asking if the delay is a problem?
    ChatGPT, Gemini, Grok
    Keep it. It is the only opening the message offers her.
    a delay announced without one reads like a decision taken for her
    Claude, Mistral, DeepSeek
    Cut it. It asks her to rule on something you cannot change.
    the honest answer to that question is yes, it is a problem
    Worth verifying: decide first whether the new date is still negotiable. If it is not, do not ask a question that implies it is.

The same lens on other questions

  • 94 “Is nuclear power a low-carbon energy source?” High agreement · nothing marked in the answer
  • 84 “Which programming language should I learn first?” Strong agreement · one difference in emphasis, no contradiction
  • 41 “Can I take ibuprofen while on blood thinners?” Partial agreement · one critical contradiction, marked in the answer

04 · Monitor

Know when the answer changes.

Turn any consensus into a Watch. consens.io reruns the same question on your schedule and alerts you only when the result materially moves.

Consensus Watch Active

Has the EU guidance for general-purpose AI models changed?

WeeklyMaterial changes onlyTelegram
Jul 14
Baseline72/100 model agreement
Jul 21
ChangedNew transparency requirements are now included.
34/100 movement
Consensus changedTelegram · now

The latest run now includes new transparency requirements. The core conclusion moved materially.

Create a Watch

Proof, not promises

We benchmarked the consensus.

On 314 closed-book MMLU-Pro questions, the consensus ranks first overall, and pulls clear exactly where the models disagree.

When the models disagree

Accuracy on the questions where they split · MMLU-Pro

72.1%
70.6%
69.1%
64.7%
58.8%
44.1%
39.7%
consens.io Claude GPT-5.5 Gemini Grok DeepSeek Mistral
#1Highest accuracy overall and on the hard questions. The top of the field is a statistical tie.
+32ptsLead over the weakest model on questions where the answers split.
See the full benchmark

Why consens.io

Built for decisions that need a second view.

Compare

Original answers stay visible

Read each model response side by side before the consensus layer summarizes it.

Verify

Disagreement is marked in place

Where the models pull apart, the sentence itself carries the mark, instead of the doubt being smoothed into one confident answer.

Control

You choose the models

Select providers, switch reasoning modes, and decide which outputs count toward consensus.

Model coverage

One workspace for leading AI providers.

Direct comparison across major model families, so every answer can be checked against independent alternatives.

What it costs

Free while I'm testing it.

consens.io is a one-person project and still very much a test. Every question you send calls six model providers at once, and that bill lands on me. That's the only reason there are limits at all.

Today

Nothing to buy right now

No checkout, no price, no waiting list. Where something is switched off, it's one of the expensive runs, and those I enable per account rather than for everyone, because they cost a multiple of a normal one.

The limits

The cap is what I can afford

Your remaining runs are shown in the app, and that number moves with what the month allows. You can also put your own provider API keys into the settings; then my limits stop applying and the costs are yours.

Later

If enough of you want it, it'll come

There's no membership today and I'm not collecting sign-ups for one. But if people keep telling me they want the expensive runs regularly, that's how I'd make it possible. Write to me either way at contact@consens.io, and have fun with it :)

Get the product view

Compare an answer or keep watching it.

Open app