Free tiers, and what they quietly ration

What free actually gives you

Every major chatbot has a free tier, and for most people most of the time it is enough. Understanding what it rations tells you when it is not, and why the same product sometimes seems to get worse in the afternoon.

Four things are rationed, and the interface is not one of them.

The model. A free tier gives you a limited number of messages on the company’s best model in a window of a few hours, and then moves you to a smaller, cheaper one until the window resets. The move is announced, if at all, in a line of small text. The lesson on the model changing under you described the mechanism; this is where most people meet it. If a chatbot was sharp at ten in the morning and vague at noon, you have probably crossed the line.

The messages. A cap per window on the best model, and a looser cap on the fallback. The numbers change often and are not worth memorising; the existence of the cap is.

The context. Free tiers often carry a shorter context window, which means the long document you pasted may be silently truncated from the top. If a summary of a long file misses the beginning, this is the first thing to suspect.

The extras. File uploads, image generation, voice mode, long-running research modes and code execution are rationed hardest, because each costs the company far more than a text reply.

Why it is free

A free tier is three things to the company. It is a funnel to the paid tier. It is a source of training data, under the defaults described in the previous lesson. And it is a source of preference votes: every thumbs-up, every regenerate, every choice between two answers is a label for the second training. You are paying with data and with judgement rather than money, and the honest correction to the old saying is that the paid tiers collect too. Paying buys capacity, not privacy; privacy is a setting.

As of 2025 the main chatbots carry no advertising. Several companies have said they are considering it. When it arrives, the free tier is where it will arrive first.

The other free doors

Beyond the chatbots’ own free tiers there are three routes worth knowing.

The comparison arena. A research site lets you put the same prompt to two anonymous models side by side, vote for the better answer, and only then see which was which. It exists to build a public leaderboard from those votes, and it gives free access to nearly every leading model. Use it to discover that the model you were loyal to is not the best at your task.

The developer consoles. The labs’ developer sites often have free allowances that are larger than the consumer free tier, at the cost of a plainer interface and, in at least one case, terms that permit training on your inputs in the free allowance. Read the line about data before pasting anything.

Hosted open models. A large community site hosts thousands of open-weights models with a free “try it” box, and free notebooks from Google and Kaggle give you a data-centre GPU for a few hours at a time. This is also the route to the fully free thing, running a model yourself, which the next lesson covers.

In India specifically

The chatbot with the most users in the country is the one inside the messaging app most people already have, where it is free with no cap that most users hit, and it accepts Hindi, Hinglish and several other Indian languages. It is also a smaller model than the flagship products, and the module on being weaker in your language applies to it in full. For a first draft in Hindi it is fine; for anything with a specific fact in it, the four questions from the first lesson of this module still apply.

What free does not change

The free model hallucinates the same way, stores your words under the same defaults, and sits on the same jagged frontier. Nothing about the price changes the mechanism. What the price changes is how often you are quietly handed the smaller model, and that, in practice, is the difference people notice and cannot name.