Renting AI by the Word: What It Costs, Who Owns the Output, and What Ends It

Most writing about using AI to build something argues about whether the output is good. That argument is not settled here and this page does not try to settle it. What this page does is answer the three questions you actually have to answer before you can plan around it: what it costs, who owns what comes out, and what happens when access stops.

All three have published answers. Almost nobody puts them in one place.

You rent by the word, and that changes the arithmetic

Neither provider charges a plan fee for API access. Anthropic sells prepaid credits on a pay-as-you-go basis with no subscription, and the cheapest entry rate is Haiku 4.5 at $1 per million input tokens. OpenAI bills per token from the first call with no plan fee listed, and gpt-4o-mini is $0.15 per 1M input tokens.

Two things follow from this, and they pull in opposite directions.

The good one: a real test costs pennies. You do not have to commit to anything to find out whether this works for your case, which is unusual in a branch where most tools want a year up front.

The bad one: there is no bill that stops you. A subscription has a ceiling built into it. Per-token pricing has no ceiling at all, only whatever limit you remember to set. People do not overspend here because the rate is high; they overspend because nothing tells them to stop.

The only cost comparison that means anything

A price per million tokens is an abstraction. It becomes a real number when you measure one actual job.

Anthropic charges no renewal fee, only per-token spend, with Sonnet 5 at $2 per million input tokens and $10 per million output tokens. Take a task you would otherwise pay someone to do, run it once, look at what it consumed, and multiply by how often you would run it in a month. That figure, against what the alternative costs, is the whole decision.

Do it in that order. Choosing a model from a price table before you have measured one job is how people end up on the expensive model for a task the cheap one handles, or on the cheap one for a task it fails at, which costs more than either.

Who owns what comes out

This clause matters more than the price, and both providers answer it more clearly than most platforms in this branch.

Anthropic states that the customer owns the Outputs and keeps all rights to Inputs, and assigns its interest to the customer. OpenAI states that the customer owns all Output, and assigns to the customer all of its right, title and interest in it.

Compare that with the licence clauses on the newsletter and ad platforms in this branch, where "you own your content" is followed by a perpetual worldwide sublicensable grant back to them. Here the assignment runs the other way. Whatever else is uncertain about building on these tools, the thing you make with them is yours.

Nothing comes back on the way out

Anthropic auto-deletes API inputs and outputs within 30 days, so there is nothing to port from their side. OpenAI deletes what it holds when the agreement terminates.

This is not a complaint. For most uses it is exactly what you want, and it is the opposite of the problem you have with platforms that keep your data and hand back a thin export.

But it has one hard consequence: your archive is your responsibility from the first call. If the prompts that make your system work exist only in somebody's API logs, you do not have a system, you have a habit. Store every input and every output on your own disk from day one, before there is anything worth storing.

What ends it

Anthropic may suspend all access over a suspected Usage Policy breach, and disclaims liability for any loss of data or profits that suspension causes. OpenAI may limit or suspend access if required by law or if the customer breaches the agreement.

Read the second half of the Anthropic clause again, because it is the honest part: not only can access stop, but what the stoppage costs you is explicitly not their problem. That is standard in the industry and it is stated plainly here, which is more than many providers manage.

Neither account is sellable on its own. Assigning the Anthropic contract needs their prior written consent, and the OpenAI agreement is not assignable except under its own narrow exception. So if the thing you build is meant to be sold one day, the API account is not part of the sale; the buyer opens their own.

What this means for planning

Treat the model as a supplier, not a partner. That framing gets every decision right:

You would not run a business where one supplier could stop shipping tomorrow with no notice and owe you nothing. So write the parts that matter to be portable, keep your own copy of everything, and know what your system does if access disappears on a Tuesday.

And you would not sign a supplier contract without measuring one order first. Run one real job, look at the bill, and decide from that number rather than from a table of rates you have not tested against anything.