An essay arguing that cheap models have crossed a practical usefulness threshold drew 499 points and 226 comments on Hacker News, making it one of the week's most-read AI pieces. The concrete claim underneath the argument is a price: Calvin French-Owen, co-founder of the data company Segment, says a personalized daily-news task that cost roughly $1 per run on the previous generation of mid-tier models now runs at about ten cents.
Key facts
- The essay, "Small Models Have Arrived," was published on 26 August 2026 by Calvin French-Owen and reached 499 points with 226 comments on Hacker News.
- The central anchor is a roughly tenfold cost drop on one repeatable task: about $1 per run previously, "the average cost is ~$0.10" now.
- French-Owen reports seeing around 100 tokens per second from the cheap model he tested, across codebase, email and knowledge-base work.
- Primary source: Small Models Have Arrived.
The essay's framing is the reason it travelled. French-Owen starts from a question investors keep asking him -- why are there so few consumer AI companies -- and answers it with unit economics rather than vision. The classic consumer playbook was to build something cheap to run, grow, then monetise. Add a model call to every request and you have a variable cost per user from day one, which changes how much capital you need before the business works at all. At a dollar per session, a consumer app charging thirty dollars a month is dead on arrival. At ten cents, it is a normal business.
The second half is more interesting and less quotable. Comparing notes with his former Segment co-founder Peter Reinhardt, French-Owen splits work into two buckets: the "IQ 180" work, where someone produces a solution nobody had thought of, and the "token spewer" work -- being ultra-responsive, nudging people, pushing a dozen fronts forward. Reinhardt, who runs multiple companies, estimated that about 95% of his own work falls into the second bucket. French-Owen's argument is that most human labour inside companies looks like bucket two, and bucket two is precisely what a fast, cheap, good-enough model can absorb.
He is careful not to overclaim. "I think demand for frontier-level models is going to keep compounding," he writes, "especially for fields that require novel breakthroughs or discovery." The claim is about a second market opening, not the first one closing -- which is a useful corrective to the recurring "small models will eat the frontier" genre.
The receipts for the general thesis are stronger than the essay's own anecdotes. TielCoder, a 22.4 GB 4-bit local build, fixed 12 of 25 problems on a live software-issue benchmark -- the same count as a frontier model at medium effort on that slice. Z.ai's GLM-5.3-Flash, released under an MIT licence as a 328 GB download, was the anonymous model that topped a public router leaderboard for a week. And at the far end of the scale, an ESP32 microcontroller project keeps 28.9 million parameters in flash and reads only a few hundred bytes per token -- though its own repository states it can write short stories and cannot answer questions, follow instructions, or write code.
That last example is the honest boundary of the argument. "Good enough" is a claim about a task, not about a model, and the essay's own evidence is self-reported: the ten-cent figure comes from French-Owen's personal evaluation, not a published benchmark. The Hacker News thread splits accordingly, with supportive comments about local models being sufficient in practice running alongside sceptics invoking the bitter lesson and the durable advantage of scale.
Where the essay is most useful is its list of what is still missing. Making cheap models work for business, he writes, requires "new harnesses, prompt injection safety, roles, and permissions." That is a precise and slightly deflating engineering agenda -- less a story about model quality than about routing between models, inference economics, and building the permission scaffolding that stops a cheap fast agent from doing something expensive and irreversible.
The reason this argument keeps resurfacing is that the two markets are measured with the same yardstick and shouldn't be. Frontier evaluation asks what the best model can do on the hardest problems. The question French-Owen is actually posing is different: for a task you will run ten thousand times, what is the cheapest model that clears the bar, and what does clearing the bar cost? Those are different curves, and a model can be far from the frontier on the first while dominating the second. It is the same reasoning behind distillation -- deliberately trading capability you do not need for cost you do.
Originally published on Ground Truth, where every claim is checked against the primary source.
