Essays on AI softwareSaturday, October 10, 2026Independent · No sponsors
TalkAI Tools
EssayBy The Talk AI Tools editorsOctober 9, 2026

Unlimited Is a Pricing Word, Not a Product Promise

Unlimited Is a Pricing Word, Not a Product Promise

There is no such thing as an unlimited AI plan, and almost nobody selling one is lying.

That combination is worth sitting with, because the reflex when someone points out that unlimited has limits is to call it deceptive and move on. It is more interesting than that. Unlimited is a real commercial instrument with a specific job, and the job is not describing capacity. Once you see what it is actually for, the word stops being a claim to verify and becomes a signal to read - and what it signals is how the vendor expects you to use the thing, and what happens if you do not behave as expected.

What this note covers, in order.
What this note covers, in order.

What the word is actually doing

Unlimited is not a capacity statement. It is a pricing decision, and specifically a decision to stop metering.

Metering costs something. A vendor that charges per unit has to define the unit, count it, display the count, explain it in documentation, defend it in support tickets, and watch customers modify their behaviour to avoid it. That last cost is the largest and the least visible: a metered product makes people ration themselves, and a rationed product gets used less, demonstrates less value, and is cancelled more often.

Flat pricing buys all of that away. The customer stops counting, support stops explaining, and usage rises - which, for a product whose value grows with use, is exactly what the vendor wants.

So unlimited is a bet. The vendor is betting that the median customer's usage is low enough that averaging the cost across everyone is cheaper than metering it.

Which means the word tells you something real, just not what it appears to say. It does not say capacity is infinite. It says the vendor has looked at the distribution of how people use this and concluded that most of them use little enough that counting is not worth the friction.

Every unlimited plan has a limit, and it is usually a shape rather than a number

The bet only works if a small number of very heavy users cannot sink it. So every flat plan carries a mechanism for handling them, and the mechanism is where the real terms live.

What is notable is that these limits are usually not expressed as a quantity. A number would be a competitor's advertising copy. Instead the constraint takes one of several shapes, each of which does the same job while leaving the headline intact.

Notice that all six are about managing the edge of the distribution, not about restricting the middle. For most customers, none of them ever activate. That is the design, not a loophole.

The honest reading of an unlimited plan

So here is the translation, and it is not cynical: unlimited means the vendor does not expect your usage to be a problem, and reserves the right to act if it is.

For most buyers that is a good deal and should be taken as one. You get a predictable bill, no rationing, and no time spent understanding a credit system. The absence of a counter is itself a feature, and the mental overhead of metered pricing is a genuine cost that flat pricing removes.

The buyers for whom it is not a good deal are the ones who intend to be the edge of the distribution. If your plan for the tool is the thing that makes the vendor's bet lose - continuous automated use, one seat serving a whole team, volume that would be expensive if metered - then unlimited is the riskiest plan on the page, not the safest. You are relying on a promise whose enforcement mechanism exists specifically for you.

The asymmetry matters. A metered plan that runs out tells you so, with a number, in advance. An unlimited plan that runs out tells you by degrading - and degradation is hard to detect, hard to prove, and hard to complain about.

Why this is sharper for AI tools than for storage

Unlimited has been a software pricing word for decades, applied to mailboxes, storage and support. AI tools make it behave differently, for one structural reason: the marginal cost of serving a request is high and it is paid at the moment of use.

Unlimited storage is a bet on space, which is cheap, mostly idle, and bounded by how much you happen to have. Unlimited inference is a bet on compute, which is expensive and consumed per request. A customer who uses ten times the median costs ten times as much to serve, immediately.

That tightens the economics in two ways. The gap between the median and heavy user is wider, so the mechanisms for handling the edge have to be more active. And because capability itself is tiered - a larger model costs more per request than a smaller one - there is a lever that storage never had: keep the volume unlimited and vary the quality.

This is the development worth watching. In a metered world, hitting the limit is an event. In an unlimited world where the relief valve is model substitution, the limit arrives as a vague sense that answers got worse this week. The first is a billing problem. The second is an epistemics problem, because you cannot tell whether the tool changed, your prompts drifted, or your work got harder. The related question of whether you can leave when this happens is argued in the claim that you can always switch.

What to do with the word when you see it

None of this argues for avoiding unlimited plans. It argues for reading the word as information rather than as a guarantee, and for one specific habit: find the mechanism before you depend on the promise.

The mechanism is always documented somewhere, because it has to be enforceable. It will be in the acceptable use policy, the fair use clause, the rate limit page in the developer documentation, or the sharing restriction in the licence terms. It is rarely on the pricing page, which is where the word unlimited lives.

If you are in the middle of the distribution, take the flat plan and stop thinking about it. That is what it is for, and the freedom from counting is worth real money.

The word is a description of the vendor, not the product

The conclusion that holds up is that unlimited describes a relationship rather than a capability.

It says: we have decided not to count, we expect you not to make us regret it, and we have kept the means to respond if you do. Stated that way it is a reasonable arrangement, and most of the time both parties honour it without ever discovering where the edge was.

The failure happens when a customer reads a commercial convenience as a technical guarantee and builds something load-bearing on top of it. That is not usually the vendor's deception. It is a category error about what kind of statement a price is - the same error as reading a tool's feature list as a description of your workflow, which is the subject of the tool is not the workflow.

Prices are not promises about physics. They are statements about what a company has decided to stop charging for, revocable in the manner set out in the terms. Unlimited is the clearest example because it is the most absolute-sounding word on the page, and the one with the most machinery behind it.

Read it as an invitation to use the thing freely, which it genuinely is. Just do not read it as a commitment that there is no edge, because the edge is the reason the word was affordable to print.