UNLIMITED.

The best AI deal was never the model. It was the window where you could use the model like compute was free.

That window is starting to close. Not fully. Not overnight. But the shape is obvious if you have seen this pattern before.

For a while, “unlimited” meant you could stop thinking about usage. You could stream, browse, upload, tether, test, waste, learn. Then the carriers found the edge of the economics. Verizon moved away from unlimited smartphone data for new customers in 2011. T-Mobile kept the unlimited language, but made the high-speed part finite.

Tesla had its own version. Free Supercharging helped make early ownership feel simple. Buy the car, drive long distances, charge on the network. Over time, the system moved toward payment methods, invoices, per-kWh or per-minute pricing, congestion fees, credits, and plan-specific benefits.

Apple’s cafeteria philosophy makes the distinction cleaner. The food is subsidised, not free. The point is not profit. Free buffets get loaded up and thrown out. Subsidy gives people access to quality, but keeps enough friction that they still make a choice.

The cheap phase is where taste gets built

When usage feels free, you do not treat it like a rare resource. You burn through it. You test dumb ideas. You ask the same question four ways. You paste huge context into the window. You let the agent run longer than it should. You generate three versions when one would have been enough.

From a cost-control view, that looks wasteful. From a learning view, it is the point. The cheap phase is where people build intuition before the meter forces them to justify every action.

AI is going through the same phase. Flat-rate plans taught people to stop treating the model like a search box. You could leave a conversation open, attach files, ask for rewrites, make the model critique its own work, and run it against a project rather than a prompt. That freedom is messy, but it is also where people learn which tasks deserve a model, where agents fail, where context helps, and where more tokens are just hiding a weak brief.

Capability changes the appetite

The AI version has one extra force inside it: capability keeps rising while tokens get cheaper in many parts of the stack. That does not calm demand down. It creates more demand.

That is the dopamine loop. It is not just “I want more messages.” It is “I want to see what the next model can do with this.” Can it build the app? Can it reason through the meeting history? Can it call the tools? Can it run for an hour and come back with something usable?

AI will not settle into one clean price-per-message world. We are going to have cheap tiny models, strong everyday models, expensive agentic models, and specialist models for images, video, voice, code, research, and internal workflows. The meter comes back, but the menu gets bigger. Capability changes the appetite.

Unlimited means unpriced behaviour

This is the section that needs to feel human, because this is the part people actually feel. “Unlimited” sounds generous. It sounds like trust. It sounds like the company saying, “Go on, use the thing as much as you need.”

But most of the time, unlimited is not a permanent product promise. It is a temporary answer to an adoption problem. The company wants people to form the habit before the unit economics are fully exposed. The user wants freedom from the meter. For a while, both sides get what they need.

Then behaviour becomes normal and the question changes. It is no longer, “How do we get people to try this?” It becomes, “How do we stop the heaviest users from consuming the margin of everyone else?”

That is when the language shifts. Unlimited becomes fair use. Fair use becomes priority access. Priority access becomes credits, caps, resets, queues, add-ons, and team controls.

You can already see this across AI products. Claude Pro gives more usage than free, but Anthropic is clear that there are limits and reset windows. Cursor talks about included agent usage in dollar terms connected to model inference costs. Replit uses credits for Agent and cloud usage.

That is not a moral failure. It is physics. These products are not selling static software. They are selling ongoing compute. Every long context window, image generation, code agent loop, web search, file parse, and tool call has a cost behind it. Eventually the product has to teach the user what the system already knows.

Usage is not free.

The meter forces judgement

When usage was cheap, the skill was exploration. When usage becomes visible, the skill becomes judgement. You need to know which model belongs to which job. You need to know when context is helping and when it is just being dumped into the window because nobody wanted to write a brief. You need to know when an agent should continue and when the human should take back the keyboard.

Cheap does not mean good. Expensive does not mean wasteful.

That is where teams need to land: not free-for-all access, not panic-driven lockdown, but subsidised exploration with visibility and judgement. Let people use the tools enough to build taste, but make the cost visible enough that they learn to choose.

If you still have access to generous plans, use them properly. Compare the expensive model against the cheap one. Track what actually changed in your work.

Do not just consume the subsidy. Convert it into judgement.

Because once the meter is fully visible, the people who only learned how to prompt will feel constrained. The people who learned how to think in usage, cost, value, and systems will know where to spend.

That is the real advantage of the unlimited phase: some people got a chance to learn before every action had a price tag attached.

- JC