Fate of Reality

September 2, 2026

Intelligence Is Getting Cheap Fast, and the Bill Still Surprises People

A hobbyist beat a million-dollar AI benchmark for 67 cents on one graphics card. Uber burned its whole year's AI budget by April. Both are true, and a new paper explains why the official numbers can't see either one.

Listen to this articleread aloud, 6 min
0:00 / 6:51

The facts:

  • An independent researcher trained a small model from scratch in an hour and a half on a single consumer graphics card, scored 44% on the ARC-AGI-1 reasoning benchmark, and put the cost at 67 cents; the benchmark has been open six years and carried a million-dollar prize (blog post via Hacker News)
  • A new economics paper assembled 21,024 posted prices across 3,208 models and 86 providers and joined them to 4,605 benchmark scores; measured the way government statisticians measure software, AI prices fell about 10% a year; adjusted for quality, they fell about 73% a year, so 87% of the decline is invisible to current methods (arXiv, September 1)
  • The same paper's catch: counted per completed task, the buyer's price stopped falling, because reasoning models burn tokens faster than token prices drop (arXiv)
  • Uber's weekly requests to its AI agents are up more than nine times since February while its spending has stayed flat since April; cost per thousand requests is down 34% from the April peak; it got there by routing easy tasks to cheap models, capping sessions, and showing engineers the live bill (Barchart via Yahoo Finance)
  • Before it tightened up, Uber spent its entire 2026 AI budget in four months; agents now write more than 70% of its code changes, and management says that lets it slow hiring (Barchart via Yahoo Finance)
  • OpenAI has started letting some large customers pay only when the AI finishes a task, such as closing a support ticket; Salesforce is buying a support-bot company that already bills that way for $3.6 billion, and one coding vendor promises up to $10 million in credits if its software doesn't pay for itself (The Information via The Decoder)
  • A developer's test: building a personalized daily news page cost about $1 a run on last year's mid-tier models and about 10 cents on this year's small fast ones; at the old price, charging $30 a month "is untenable for a consumer app" (Calvin French-Owen)
  • Chinese open-weight model GLM-5.3-Flash is priced at 15 cents per million input tokens; Thomson Reuters replaced Claude with its own model built on Alibaba's Qwen for document review; the share of businesses paying for open-model platforms rose from 4.5% in January to 6.1% in July (Fortune, citing Ramp)
  • Anthropic's most capable model accounts for 6% of the tokens businesses buy from it and 11% of the dollars, at about twice the price of OpenAI's flagship; Ramp's economist says buyers "are not willing to pay a premium for the best model on the market when a cheaper one is good enough" (Fortune)

Two headlines from the same week. A guy with one graphics card and 67 cents matched results that a million-dollar prize couldn't buy for years. Uber, with a real budget and real engineers, spent the year's AI money by April. How can both be true?

The paper out of arXiv on Monday answers it better than any press release. The posted price of a token is falling, slowly. The price of a unit of quality is falling fast, about 73% a year, which is faster than anything the official statistics know how to measure. And the price of a finished job is flat, because the newer models think longer before they answer, and thinking is tokens.

So the intelligence got cheaper and the bill didn't. That's not a contradiction. It's what "cheaper" does when demand is infinite.

the Uber fix

Uber's story is the useful one for anyone running a business, because it ends with a solution. Requests to its agents went up nine times. The cost per request came down by a third. The trick was boring: send the easy questions to the cheap model, save the expensive one for the hard jobs, put a cap on each session, and let every engineer see what they're spending in real time. Same discipline as a cell phone plan.

The part investors noticed is the last line: seventy percent of Uber's code changes now come from agents, and the company is hiring slower because of it. The cheap model didn't just cut a bill. It changed who gets hired.

the cheap end is where the new stuff comes from

Calvin French-Owen's post has the clearest explanation of a thing people keep asking: why aren't there more consumer AI apps? Because every request used to cost real money. A personalized news page cost a dollar a run on last year's models. You can't charge a normal person thirty dollars a month for that and survive. This year's small fast models do it for a dime, and a dime is a business.

Then look at the Chinese open models, which are the other end of the same price curve. Fifteen cents per million tokens. Thomson Reuters, not a startup, swapped Claude out for its own model built on Alibaba's Qwen for reading legal documents, and said out loud that you don't need the biggest model to get useful work done. The share of businesses paying for that kind of platform is still small, 6%, but it's grown by a third since January.

what the frontier is worth

Here's the number that should get more attention than it did. The single best model on the market, priced at about twice its main rival, makes up 6% of the tokens businesses buy from its maker and 11% of the money. Ramp's economist reads that as a ceiling: companies want the best thing to exist, and then they buy the thing that's good enough.

Which is why OpenAI is quietly changing what it sells. Not tokens, but finished tasks. You pay when the ticket closes. Salesforce is paying $3.6 billion for a company that already bills that way. The hard part, as Stripe has already written guidelines about, is proving the AI did it and not your marketing team.

The hobbyist ended his post with a question for the field: the benchmark was open for six years, high profile, with a million dollars on it, and the answer was a small transformer with the obvious representation. "Maybe the cost of experimentation was high enough that they couldn't run ablations properly." Meaning nobody tried the cheap thing because everything was expensive.

That's the week in one sentence.

Sources: mvakde blog via Hacker News; "The Price of Intelligence" (arXiv, September 1, 2026); Barchart via Yahoo Finance; The Information via The Decoder; Calvin French-Owen; Fortune's Eye on AI, citing Ramp.