What “free” actually means in an AI app

Running a model costs real money every single time. So when an app says free, the useful question is not whether they are lying. It is who is paying.

Four ways free gets paid for

There is no version of this where the electricity is donated. Every free AI feature is funded by one of four things, and you can usually work out which within a day of using it.

• You are the product. The conversations become training data, or behavioural data, or both. This is the oldest arrangement on the internet and it is not always a scandal — but it should be stated in the first screen, not the fourteenth paragraph of a policy.
• It is a funnel. Free is a demo with a wall in it. The wall is usually a message count, and it always arrives at the least convenient moment, because that is what makes it work.
• Someone else’s money. Investor funding, burned deliberately to buy users. Genuinely free, genuinely good, and it ends. Every product that has ever done this has later either found a business model or shut down.
• It is small. The free tier runs a cheap model, so the cost per message is low enough to absorb. This is the least exciting answer and the most sustainable one.

What ours is

Ours is the fourth. AiHub has one house model that costs nothing to use, and it costs nothing because it is cheap to run, not because you are paying with something else.

Concretely, in the product:

• No balance check. Every other model checks that you can afford the request before it starts. The free one skips that check entirely — it is precisely the model that has to work for an account with zero tokens.
• No charge, ever. Not a discounted charge or a subsidised one. The code that writes a ledger row is not reached for a free model.
• No message count. There is no daily cap to hit, no cooldown, and no “you have 3 messages left today”.
• No card. Nothing in the product asks for one, because there is no card payment in it at all yet.
• No training. We do not train models. We have none of our own, and your conversations are not handed to anyone for that purpose. That is in the privacy policy in the same words.

One name, several models

AiHub AI is a name rather than a model. Behind it sits a short chain: several models that providers publish at no cost, and one cheap paid model at the end that we carry ourselves. Free models are exactly as dependable as free things usually are — they rate-limit, they go down, they get retired with no notice — so when the first one refuses, the request moves along the chain inside the same call and you get an answer instead of an error.

Which link answered is not a secret, but it is also not something the model can see from where it sits. Ask it what it runs on and it will tell you it cannot tell, because that is true, and a confident invented answer would be worse than admitting it.

What it is not

It is not the frontier. The house model is fast and reasonable — it answers ordinary questions well, it holds a conversation, it will summarise, rephrase, explain and draft. It will not out-think Claude Opus on a hard problem, and it does not take attachments: no PDFs, no screenshots. If your question is really about a file you have, the free model is the wrong tool and it will not pretend otherwise.

We would rather describe it that way than call it “powerful”. A free model oversold is a free model that makes you distrust the paid ones.

Why have one at all

Two reasons, and neither of them is generosity.

The first is that an app which stops working when your balance hits zero is a bad app. Balances run out, usually at the worst moment, and until there is a checkout button here they run out until someone reads an email. A product where that state means a locked door is a product people stop opening.

The second is that you should be able to see the thing before you spend anything on it. Not a stripped demo, not a watermarked preview — the actual interface, the actual conversation, the actual history sync, with a real model behind it. If the app is any good, that sells it. If it is not, no amount of trial machinery will.

The trial that is not a trial

The reason to be careful with the word is that “free” in AI usually means “free until you would find it useful”. The pattern is familiar: generous for a week, then a limit appears exactly where the habit formed. It works, which is why it is everywhere.

The house model is not that. It has no counter behind it and no scheduled deterioration. The upgrade path is not a wall you hit — it is a shelf of forty-odd other models sitting next to it, each with its price printed on it, that do things this one cannot. You move because you want the frontier model or you want to make a picture, not because you ran out of something.

How to tell, in any app

Four questions that sort the honest free tiers from the rest, and they work on us too:

• Is there a counter? If there is a number of messages left, it is a trial. That is fine — but read it as one.
• Does it ask for a card? A card on file for a free tier is a subscription with a delay in it.
• What does the privacy policy say about training? Search it for “train”. If free conversations are used and paid ones are not, that is the price, and it is being charged in a unit you cannot see.
• Can you find out what the paid thing costs without signing up? A product that hides its prices behind a registration form has decided the price is a negotiation. Ours is on one page, every model, including what the provider charges us.

Free should be a real thing you can rely on, or it should not be called free. Ours is a small model that costs us little and asks you for nothing. That is the whole arrangement, and there is no second half to it.