Why AI agents might finally be the perfect micro-payers

Hi, I’m media innovation journalist Ulrike Langer and you’re reading my weekly AI in media newsletter. Thank you for being here! If someone has forwarded it to you please subscribe and never miss any future post.

My next European trip:
I’ll be in Europe again this fall and wanted to share a few dates in case it’s useful for you. I’m speaking at the Reporter:innen Workshop in Hamburg on September 25–26, at the Media Tech Hub Conference #MTHCON in Potsdam on September 29, and at the b°future Festival in Bonn on October 1–3. If you are at any of these events, please say hi! If you’re based in Europe and would like to set up a talk, workshop, or strategy session while I’m around, I still have some availability on October 5, 6, 7, and 8. If you’re interested please hit reply or send me a LinkedIn DM.

Two lines on a media company's ledger are heading toward the same pricing logic this summer: what you pay for AI, and what AI pays you for journalism. On July 20, Anthropic starts charging subscribers for its best model by the token. At the same time, publishers are being pitched consumption-based licensing as the way AI will finally pay fairly and transparently for content. Media executives who understand the connection between the two will have an edge in negotiations.  

The $20 AI subscription was never a calculated price intended to make a profit. "I don't think there was really any strategy involved in charging $20 a month" for ChatGPT Plus, TechCrunch's Sean O'Kane recalled on the Equity podcast. "It was just sort of like, 'Let's spit out a number.'" This number worked because investor money covered the difference between what subscribers paid and what they consumed but the difference has become too big to subsidize. Reasoning models burn tokens by the million, agents run for hours, and the heaviest users cost AI companies many times what they pay.

Anthropic is the first to act on this gap. From July 20, its top model Claude Fable 5 is no longer included in the subscription bundle: Subscribers pay for it by the token, at the same rates developers pay. This appears to be the first documented time a frontier lab has put a consumer model behind usage-based billing. Anthropic says the move is temporary, until capacity catches up. Maybe. The direction matters more: AI intelligence is moving from the fixed cost bucket into the variable cost bucket. Every media budget that treats it as a fixed line item was built for an era that ended this month.

Will pay-per-usage AI pricing work? Past experiments with micropayments have shown that human users hate paying this way. Anthropic learned that lesson within days. The new pricing model was set for July 7, slipped to July 12 after subscriber backlash, then slipped again to July 19. People who happily pay $20 or even $200 a month revolt the moment a price tag becomes visible per use.

German publishers especially know that reflex from past experience. In 2016, Der Spiegel began selling single web articles for 39 cents. Revenue stayed negligible, and by 2018 Der Spiegel had retreated to the flat subscription it sells today. Blendle, which built a Dutch platform on per-article payments, ended the same way. The pattern proved itself time and again: Micropayments turn every act of consumption into a small purchasing decision, and humans will pay extra to be spared that unwelcome feeling.

For publishers, this settles an old question. Reader revenue will remain by subscription. The pay-per-usage future, if it comes, will not be humans paying per article. It will be machines.

Cosmin Ene has sold pay-per-usage news twice – once to humans, and now to machines. His Munich-based company LaterPay powered Spiegel's 39-cent per article experiment. In 2022, the firm renamed itself Supertab, and today Ene pitches its payment rails at AI companies and their agents. LaterPay was built to soften the sting its rivals never solved. Blendle made readers pay per article up front; LaterPay let them read first with one click and settle the tab only once it reached a certain threshold – the "later" in the name was the entire pitch. 

Der Spiegel's version worked exactly that way: Paywalled Plus articles at 39 cents accumulated quietly until a five-euro threshold, and only then did the register open. The model failed anyway. Part of the reason: Tech-savvy users quickly found ways to reset the meter. For less tech-savvy users the friction, even if deferred, was still too much. 

If a softened obligation didn't work, no obligation at all worked even less. The same internet era produced voluntary buttons – Flattr and Kachingle, which let readers reward free content with a few cents if they felt like it. I ran both on my own media industry blog Medialdigital (RIP); they never collected more than pocket change. Strict, soft or voluntary: Every version of small payments for news died of the same human adversity to having to think about whether it’s worth paying for something.

Machines don’t care about any of that. An AI agent fetching an article to answer a question doesn't hesitate at 10 cents. It reads the price, checks it against its budget, and pays or moves on – millions of times a day, aggregated into real money.

Machines not minding the price is only half the case for pay-per-usage. The other half: A flat rate for machines cannot work. The news subscription survives heavy readers because a subscriber who reads 300 articles per month costs the publisher barely more than one who reads 50, and pays the publisher back in attention and loyalty besides. Agents break that math twice over: They consume at machine scale, and they return nothing a publisher can monetize – no ad impressions, no habit, no subscription. "A new category of eaters that never stop," Ene calls them on Brian Morrissey's The Rebooting Show at Cannes. 

So the solution is to charge them per bite. Ene frames it as a choice about where publishers sit in an AI company's books; as a cost that every company tries to shrink or route around, or as a supplier that gets paid whenever the AI earns money. A per-use fee puts publishers on the earning side: The more revenue the AI's answers collect, the more its sources collect. Both sides of the AI economy oblige by the same rule – whoever causes usage pays for usage.

Ene’s own caveat: He expects no meaningful publisher revenue from this for 12 to 24 months. The rails are still being laid – standards that let machines read a publisher's rights, prices and usage terms. The Associated Press joined one such coalition, SPUR, just last week. The bet is unproven. But it is the first version of micropayments whose target customer has none of the traits that killed the previous ones.

There is a working alternative, and Le Monde built it. Chief executive Louis Dreyfus signed licensing deals with OpenAI, then Perplexity, then Meta. None is a revenue share. Each deal pays a minimum guarantee with a small variable on top: fixed income for several years, 25 percent of it passed to journalists under a union agreement. Dreyfus also rejects the idea that chatbots eat his subscription business. People don't pay Le Monde for the news a chatbot can summarize, he argues: "They subscribe to read long-form journalism from our reporters in Ukraine, in Sudan, in Gaza” (The Media Stack).

The problem with this approach: A guaranteed check buys certainty, but it surrenders knowledge. The AI company can measure exactly which articles its systems fetch, how often, and for what purpose. The publisher sees none of that intelligence – and learns nothing about what its content is worth. The income is fixed on day one, which also means it is capped on day one. Pay-per-usage reverses every one of those terms. The publisher sees each access and each payment, and the upside grows with consumption. In exchange, there is no floor under the revenue – and for now, hardly any revenue at all.

And the choice between them is not equally available. More than two years after Le Monde signed its deal in March 2024, Dreyfus says it remains the only French publisher with an OpenAI deal – the guaranteed check is an offer most newsrooms will never receive. The courtroom doesn't rescue that math either. Ene's arithmetic: If you settled every AI lawsuit for $2 billion and spread it across 10,000 newsrooms, each would receive $200,000 – once.

The flat rate, on both sides of the AI economy, was never a realistic price tag. It was a promise that consumption wouldn't matter. AI companies have stopped making that promise to their customers. Publishers should stop making that promise to AI companies.