The three bills behind every AI answer
It helps to stop thinking about "the cost of AI" as one thing. When an AI search engine or chatbot answers you, the company running it is usually paying for three very different things.
First, compute — the servers, GPUs, networking, and storage that actually run the model. This bill exists no matter whose model is used, and it's the one most people forget about.
Second, model access — the fee to use the language model itself. This one only applies when a company is renting someone else's model instead of running its own.
Third, content — payments to the publishers and rights-holders whose articles, books, and data the system draws on. This is the newest, messiest, and most contested bill of the three.
Does Perplexity pay OpenAI?
Usually, yes — but through an API, not some special arrangement. When Perplexity routes your question to a model from OpenAI, Anthropic, Google, or xAI, it pays that provider per use, at roughly the rates any developer would pay. Perplexity's own developer pricing is explicit about this: third-party models are billed at direct provider rates, measured in tokens — chunks of text going in and coming out.
The exception is when Perplexity uses its own in-house model family, Sonar. Then there's no external model fee to pay, because Perplexity owns that one. It still pays to run it, and it still pays to search and index the web — but the per-token check to another AI company disappears.
So the honest answer to "does Perplexity pay OpenAI" is: when it uses OpenAI's model, yes, per request — and when it uses its own, no.
Then what does ChatGPT actually pay for?
This is where a lot of people get confused. If OpenAI owns its models, does ChatGPT pay any of these bills?
It pays most of them — just not the external model fee. When ChatGPT answers using OpenAI's own model, there's no second company to cut a check to for the model itself. But the compute bill is still very real: every answer burns server time, GPU capacity, and engineering. When ChatGPT searches the web, there are search and indexing costs. And when it draws on licensed news content, OpenAI pays for that too.
So your instinct is basically right: a chatbot running its own model pays the same categories of cost as a search engine renting one — minus the line item for somebody else's model.
The content bill is where it gets complicated
Compute and model fees are straightforward: you use it, you pay for it. Content is the part nobody has fully figured out, and it shows up in at least three different flavors.
Licensing deals are forward-looking partnerships, where an AI company pays a publisher to use its content. OpenAI has signed many — with the Associated Press, Axel Springer, the Financial Times, News Corp, Condé Nast, and The Washington Post, among others. In December 2025, Meta jumped in with multi-year deals covering CNN, Fox News, USA Today, Le Monde, People Inc., and more, to feed real-time news into Meta AI. Microsoft has taken a different route with a pay-per-use content marketplace, while OpenAI has leaned toward larger upfront sums.
Lawsuits and settlements are backward-looking — payments for content already used without permission. The clearest example: in 2025, Anthropic agreed to a $1.5 billion settlement with authors and publishers over books used to train its Claude models, the largest copyright settlement in U.S. history. But a settlement is not the same as a partnership. It closes the book on past use; it does not license future content. Perplexity, for its part, has been sued by several outlets — including CNN — over how it uses their reporting.
Free scraping is the third flavor, and it's the one driving most of the fights: content pulled straight from the open web at no charge, the way search engines always have. The argument is no longer really about whether crawling is allowed — it's about whether using that content to generate an answer that replaces the click is a different bargain entirely.
So is Google more powerful than OpenAI here?
In the broad sense — yes, easily. Google still handles around 90% of search queries, which means most publishers depend on it for discovery, traffic, and ad revenue. That's structural power OpenAI simply doesn't have. But raw power over the web isn't the same thing as holding the most paid content licenses, and the two keep getting tangled together.
Because OpenAI lacks Google's built-in distribution, it has had to negotiate access — which is part of why its list of public publisher deals looks longer. Google, by contrast, can often reach the same content through its existing search crawl, which is exactly what regulators have started pushing back on.
That tension came to a head this week. On June 3, 2026, the UK's Competition and Markets Authority issued a world-first binding order requiring Google to let publishers opt out of having their content used in AI Overviews — without losing their normal placement in search. Google has nine months to comply. Notably, the order does not force Google to pay for that content; the regulator said it will wait at least a year before deciding whether to require licensing negotiations. The backdrop is rough for publishers: "zero-click" searches, where people get their answer without ever visiting the source, have climbed sharply since AI summaries rolled out.
What it adds up to
Strip away the noise and it's simple. Compute always gets paid. Model access gets paid when a company rents instead of builds. Content is the bill still being argued over — in boardrooms and in courtrooms — sometimes settled with a licensing deal, sometimes with a lawsuit, sometimes not paid at all.
The next time an AI hands you a clean, confident answer in two seconds, it's worth remembering how many invisible meters were running to produce it — and how many of the people whose work made that answer possible are still fighting over their cut.