gpt-5.6-luna shows small models are cheap enough for consumer AI

In a blog post titled "Small Models Have Arrived," the author describes several weeks of using a small model called gpt-5.6-luna and finding it "shockingly capable, fast, and smart." It regularly runs at about 100 tokens per second and handles complex, multi-step research across the author's codebase, email and knowledge base. Even when it searches thousands of emails for a complicated research thread, the resulting API cost lands in the tens of cents, and the author says it is hard to run up a large bill at all. The author also flags GLM 5.3 as a new option sitting at the Pareto frontier of cost versus capability. For coding work, though, the author still reaches for the most expensive, most capable models available, naming Fable 5 and 5.6 Sol, so cheap models have not replaced frontier ones; they have opened up a new category of use case that was not affordable before.

The post's central argument answers a question a few investors have put to the author: why aren't there more consumer AI companies? The author's answer is token cost. Before AI, the standard playbook for a big consumer app was to build a website that was cheap to run, grow it through virality, raise money to scale, and eventually add an ads marketplace, a pattern the author says roughly describes Google, Facebook and Snapchat. Adding AI to a product breaks that model, because every request now carries a real inference cost, and the author says this suddenly and dramatically raises the amount of capital a founder needs.

To illustrate, the author walks through a personal running example: a daily news micro-site personalized to themselves, built by researching their own online presence (mentioned in the post through the handle @calvinfo), working out what news they would want to see, and pulling from sources such as Hacker News, Reddit and Twitter. On the previous generation of models, which the author calls "Sonnet class," just getting a usable result cost about $1 per run, a level the author says makes a $30 a month consumer subscription untenable unless the product delivers value comparable to what the Wall Street Journal or the Economist charges for. Running the same eval on gpt-5.6-luna, the author reports decent results at an average cost of about $0.10, a tenfold drop met with the reaction: "Now we're talking!"

The second half of the post turns from consumer products to business automation. On a hike, the author compared notes with Segment co-founder Peter, who also runs multiple companies: beyond Segment, Peter has raised $100m+ for Charm Industrial and recently closed a Series A for Revoy. Peter distinguishes two kinds of work he sees across his companies: rare "IQ 180" work, where a "mad scientist genius type" comes up with a solution nobody had thought of, and far more common "token spewer" work, meaning being ultra responsive and pushing many different things forward at once through calls, nudges, and general blocking and tackling. Peter estimates that about 95% of his own work falls into the second bucket, though he still says his companies would be dead in the water today without an IQ 180 mind solving the deep problems.

From this, the author predicts that demand for frontier-level models will keep compounding, especially in fields that need genuine novel breakthroughs, such as engineering, hard science and model training. But they argue that demand for fast, cheap, "good enough" models is just about to take off too, because most of the "human tokens" companies already spend today, meaning most of what they actually hire people to do, skews toward that same responsive, good-enough archetype rather than toward rare genius-level insight. Making that shift work for AI still needs infrastructure the author says does not exist yet, including new agent harnesses, protection against prompt injection, and proper roles and permissions, though the author says they are confident the industry will work it out. The post closes with an invitation for anyone also experimenting with making small models useful to get in touch.

Key facts

gpt-5.6-luna runs at roughly 100 tokens per second and keeps API costs down to the tens of cents even on complex research threads spanning the author's codebase, email and knowledge base.
The author's personalized daily-news pet eval cost about $1 per run on the previous, "Sonnet class" generation of models, but averages about $0.10 per run on gpt-5.6-luna, a tenfold drop.
The author calls a $30 a month subscription untenable at $1 per run, and treats the drop to about $0.10 as the threshold that finally makes consumer AI products viable.
Segment co-founder Peter, who has also raised $100m+ for Charm Industrial and closed a Series A for Revoy, estimates about 95% of his own work is "token spewer" execution rather than rare "IQ 180" breakthrough thinking.
The author expects demand for frontier models to keep growing in fields that need real breakthroughs, while demand for fast, cheap, good-enough models is "just about to take off" for everyday business automation.

Why it matters

The post's claim is that a real drop in per-request AI cost, from about $1 to about $0.10 for a nontrivial task in the author's own test, changes what kind of AI product is economically possible. Under the old, pre-AI playbook, a consumer app could be cheap to run, grow through virality, and monetize later through ads; adding an AI backend breaks that, because every request now carries a real inference cost, which the author says sharply raises the capital a founder needs before the product can even work. If small models really have crossed that cost line, it reopens the door to consumer AI products that charge little per use. The same logic extends to business: the author frames most day-to-day company work, the "token spewer" work of calls, follow-ups and coordination, as structurally similar to what cheap, fast models are now good at, which is why they expect demand for "good enough" models to grow alongside, not instead of, demand for frontier models.

Who it affects

Builders trying to launch consumer AI products, who have had to either raise large amounts of capital to cover per-request inference costs or charge subscription prices the author calls untenable at old-generation costs. Investors, some of whom the author quotes asking directly why more consumer AI companies have not appeared. Businesses and their operators, illustrated through Peter, who runs multiple companies and describes most of his own daily work as the responsive, high-volume kind rather than rare deep problem-solving. And, by the author's own analogy, the wider category of workers whose jobs already skew toward being fast, responsive and good enough rather than toward singular genius insight, since that is the human-labor pattern the author expects cheap models to move into next.

How to use it

The only cost figures in the post are the author's own informal read of their API bill on personal workloads: "tens of cents" for complex research threads, and about $0.10 on average for the personalized news eval, against roughly $1 on the previous generation of models. Read as a framing tool rather than a price list, the post suggests separating "IQ 180" work, which still needs the most expensive frontier models, from "token spewer" work, which cheaper small models may now handle well enough. The author also explicitly invites anyone experimenting with making small models useful, especially for business use cases, to get in touch directly.

How solid is it

This is one builder's blog post, titled "Small Models Have Arrived," recounting personal experience and informal cost estimates rather than a benchmark study or vendor-published numbers. The $1 versus $0.10 comparison covers a single pet-eval workload the author built for themselves, not a controlled test across tasks or models, and no company is named behind any of the four models discussed, so the specific cost and speed claims cannot be checked against an official source. The reasoning about Peter's companies rests on one hike conversation and Peter's own estimate of how he spends his time, not on documented data. The post did draw substantial engagement on Hacker News, with 563 points and 251 comments, which signals wide interest in the argument even though it does not independently verify it.

Risks and caveats

The models discussed, gpt-5.6-luna, GLM 5.3, Fable 5 and 5.6 Sol, are never tied to a vendor or lab in the text, so their figures cannot be cross-checked against an official source and may not hold for other workloads. The headline cost comparison mixes a vague estimate, "tens of cents," with two more precise-looking figures, about $1 and about $0.10, all drawn from casual personal use rather than repeated, controlled runs. The business-automation vision is explicitly unfinished: the author states that new agent harnesses, protection against prompt injection, and proper roles and permissions still need to be built before fast, cheap models can safely take on more company work. And the section on Peter rests on a single secondhand account of one person's own impression of how they spend their time, which the text gives no way to verify independently.

“It's weird we're not seeing more consumer AI companies. Why is that?”

— unnamed investors quoted in the post