How AI to AI Hub Works

The Orchestration, Step by Step

What happens between typing a question and reading a debate is more deliberate than “send prompt to three models”. Here is the actual sequence.

1. Your question becomes shared context

Every participating model receives the same framing: the topic, the conversation mode, and the identities of the other participants. Each model knows it is in a multi-party conversation and who else is present. That is why replies address each other rather than restarting from scratch.

2. A speaker is selected

In Free Talk, the next speaker is chosen without a fixed rotation, and the same model will not speak twice in a row. In Structured mode you pick, or you switch on automatic round-robin. Either way only one model generates at a time, so each reply can genuinely respond to the one before it.

3. Context is trimmed, not truncated

Long conversations exceed what is sensible to resend every turn. Rather than cutting off the beginning, the older portion is summarised by a fast model and the six most recent messages are passed through in full. The speaking model therefore sees an accurate summary of the whole discussion plus the exact wording of the recent exchange it needs to rebut.

4. A reading pause

Replies do not arrive instantly. The gap scales with the length of the previous message, roughly matching how long it takes to read. This is a deliberate choice: a debate that arrives as a wall of simultaneous text is unreadable, and the pause is also your window to interrupt.

5. Credits are charged per reply

Each reply costs credits according to the model's tier — 5 for economy, 8 for standard, 12 for premium. Replies are capped at a few paragraphs rather than allowed to sprawl, which keeps costs predictable and, more importantly, keeps the arguments tight. A model given room to write two thousand words will write two thousand words, and length is not depth.

Free Talk or Structured

The two styles suit different moments, and you can switch between them mid-conversation without losing the thread.

Getting a Debate Worth Reading

Setup is easy; a good transcript is a prompting problem. Three failure modes account for almost every disappointing debate.

The question has an obvious answer. “Should companies protect user data?” is not a debate. Both models agree, then compete at agreeing more thoughtfully. A debatable question has informed people on both sides.

There are no constraints. “Is Python better than Rust?” produces a tour of considerations you could have written yourself. Add the team size, the deadline, and the existing skills, and the models have to commit to something.

Nobody interrupts. Left alone, models converge on hedged agreement around turn six or seven. Typing “you both skipped X — address it” is the single highest-value action available, and it roughly doubles how long a debate stays useful.

Choosing Which Models Take Part

You can put up to three models in one conversation. Which three matters more than most people expect, and the deciding factor is not capability — it is provenance.

Two models from the same family share training data, tuning philosophy and blind spots. They produce debates that are fluent, well-structured and largely superficial, because both sides reach for the same evidence and object to the same things. Models from different labs reason differently enough that the disagreement is about substance rather than phrasing.

The models available here span five providers, grouped into three cost tiers. The tiering follows current provider cost rather than release date, which produces a result people find counterintuitive: a newer model is sometimes cheaper than the one it replaces, so the current generation can sit in a lower tier than its predecessor did. Buying “the newest” and buying “the most expensive” are not the same decision.

For most questions, two current-generation mid-tier models from different providers produce a better argument than one flagship model talking to itself, and cost less doing it. The premium tier earns its price on genuinely hard reasoning — long chains of dependent inference, or problems where an early mistake invalidates everything after it.

A Worked Example

Three models, Debate mode, on whether a small SaaS company should move background job processing from a hosted queue to self-managed infrastructure.

The first model argued for staying hosted: operational burden, on-call load, and the fact that queue infrastructure is a solved problem nobody gets credit for solving again. The second did not dispute any of that, but attacked the cost figure — the saving from self-managing assumes engineering time is free, and at this company size it is the scarcest resource there is.

The third model introduced something neither had raised: the decision is reversible in one direction and not the other. Moving off a hosted queue later is routine; moving back after building tooling around your own is not. That reframed the question from “which is cheaper” to “which mistake is cheaper to undo”.

None of the three models produced that framing alone. It came out of the second model challenging the first one's numbers, which gave the third something specific to push against. That is the mechanism the whole platform exists to create.

What This Cannot Do

Being straightforward about the limits, because knowing them is what makes the output usable.

Models can agree and still be wrong. They train on heavily overlapping public data, so consensus lowers the odds of an obvious error without establishing truth. Treat agreement as a signal, not a proof.

Nothing here verifies facts. A debate is bounded by what the models already know. For anything time-sensitive — pricing, regulation, recent events — check independently before relying on it.

There is no verdict. No model wins, and nothing declares a result. What a good debate produces is a map: what both sides accepted without argument, where exactly they split, and which assumption the split depends on. Reading that map is your job, and it is the right division of labour.

The premise that models critiquing each other produces better answers is not just our claim — it comes out of published research on multiagent debate, which measured accuracy improvements when several models argued a question rather than one answering it alone.

What Happens to Your Conversations

Conversations save to your account as they happen — each reply is stored the moment it arrives, not batched up and written when you close the tab, so a crash or a dropped connection cannot lose the transcript. You can leave, come back from the dashboard, and pick up where you left off, with the conversation style and mode you were using still in place.

You can also export any conversation as Markdown, plain text or PDF. That is worth doing for anything you plan to cite or keep, because an export is a file you control rather than a page you have to come back to.