⏱️ Lectura: 11 min
Researching a complex topic with an AI assistant used to cost as much as a premium subscription per session: nearly a dollar each time. Today, with the latest small AI models, that same work (cross-referencing thousands of emails, tracking a person’s activity online, and putting together a personalized summary) costs just ten cents.
📑 En este artículo
Developer Calvin French-Owen, former co-founder of Segment, documented the shift in an essay published on August 26, 2026, titled “Small Models Have Arrived.” His central argument: fast, cheap models have now reached a level of capability that makes viable consumer products that were previously impossible to sustain economically.
TL;DR
- gpt-5.6-luna handles complex research, like searching through thousands of emails, for just a few cents, according to Calvin French-Owen.
- With Sonnet-generation models, that same task cost around $1 per session: unsustainable for a $30-a-month app.
- GLM 5.3 joined the cost-performance Pareto frontier alongside gpt-5.6-luna, according to the chart from artificialanalysis.ai.
- Peter Reinhardt, co-founder of Segment, says 95% of his daily work is “token spewer” work: fast responses, not isolated genius.
- Reinhardt raised over $100 million for Charm Industrial and closed a Series A for Revoy.
- French-Owen uses the more expensive models (Fable 5, 5.6 Sol) only for coding; for research he uses the cheap ones.
- Demand for frontier models will keep growing in science; demand for fast, cheap models is just taking off for business.
Introduction
For years, the public conversation about generative AI centered on the biggest, most expensive models: the ones that solve research problems, generate complex code, or reason through mathematical proofs. But behind that race for the frontier, another category of small AI models advanced quietly: fast, cheap, and surprisingly competent at day-to-day work.
Calvin French-Owen knows both worlds. He co-founded Segment, the customer data platform, and now writes about the state of AI development. In his August 26, 2026 essay, he argues that the quality leap in models like gpt-5.6-luna and GLM 5.3 isn’t just a technical improvement: it’s a shift in what kind of consumer product becomes viable to build.
What happened: small AI models drove down the cost per query
French-Owen has spent weeks testing gpt-5.6-luna on real tasks: reviewing his code, reading his email, and querying his knowledge base. He reports speeds of around 100 tokens per second and results he describes as surprisingly capable.
What caught his attention most wasn’t the speed but the cost. Even having the model search through thousands of emails, the total spend on API calls ended up in the tens of cents, not tens of dollars. To put it to the test with a concrete case, French-Owen uses a personal “eval”: asking a model to research his public activity, identify what news he’d be interested in, and put together a personalized daily micro-site by searching Hacker News, Reddit, and Twitter.
With the previous generation of models, comparable to Claude Sonnet, that task cost around a dollar per run. With gpt-5.6-luna, the result is “pretty decent” and the average cost drops to about $0.10. That’s the difference between a product that needs to charge like a premium newspaper (the Wall Street Journal or The Economist) and one that can be given away free or charged in cents per use.
The other data point French-Owen highlights is that GLM 5.3 joined the cost-versus-performance Pareto frontier alongside gpt-5.6-luna, according to the chart he cites from artificialanalysis.ai. For coding tasks, however, he still prefers the more expensive, more capable models, Fable 5 and 5.6 Sol: the frontier and small models aren’t competing for the same work, they complement each other.
Context and history: why there are barely any consumer AI apps
Before generative AI, the playbook for building a mass-market consumer app was well known: put together a site that’s cheap to maintain, get users through some virality, raise capital to scale, and eventually build an advertising marketplace. Google, Facebook, and Snapchat more or less followed that script. Amazon and Netflix are the notable exceptions, French-Owen points out.
That playbook breaks the moment a language model gets added to the product. Every generated response carries a real inference cost per request, and that cost doesn’t drop to zero with scale the way the marginal cost of serving a web page did. Suddenly, the capital needed to sustain millions of active users skyrockets.
That, according to French-Owen, is the answer to a question several investors separately asked him: why there aren’t more consumer AI startups. The short answer is cost per token. If every user session costs a dollar in API calls and the product can’t charge $30 a month without delivering the value of an outlet like the Wall Street Journal, the business model simply doesn’t work.
Technical details and performance: frontier models versus fast models
French-Owen connects this cost drop to a distinction he borrowed from Peter Reinhardt, his co-founder at Segment and now head of Charm Industrial (which has raised over $100 million) and Revoy, which just closed a Series A round. Reinhardt divides work within a company into two categories:
- “IQ 180” work: the kind solved by an unconventional genius, with a solution no one else had thought of. It requires the most capable model available.
- “Token spewer” work: being ultra-responsive, moving forward on dozens of fronts at once. Calls, following up on pending tasks, operational coordination.
According to Reinhardt, about 95% of his daily work falls into the second category, though he notes his companies wouldn’t survive without an “IQ 180” mind solving the underlying technical problems. Most human work at any company, French-Owen argues, looks more like “token spewer” than isolated genius: that’s why most hiring favors the fast, cheap, good-enough profile, not the exceptional genius.
This distinction explains why demand for small AI models and demand for frontier models aren’t competing with each other, but growing in parallel, each for a different kind of task:
| Category | When to use it | Advantage | Limitation |
|---|---|---|---|
| Frontier model (Fable 5, GPT-5.6 Sol) | Complex engineering, hard science, model training | Maximum reasoning and discovery capability | High cost per query: doesn’t scale to millions of free sessions |
| Small, fast model (gpt-5.6-luna, GLM 5.3) | Customer support, lightweight research, consumer apps, high-volume agents | Cost per query in the cents, around 100 tokens/s | Lower reasoning ceiling on truly novel problems |
A pattern several teams already use to get the best of both is routing each request by complexity: a small model classifies and handles the routine cases, and only escalates to the frontier model when the task demands it.
flowchart TD
A["User request"] --> B["Complexity router"]
B -->|"routine task"| C["Small model: gpt-5.6-luna"]
B -->|"novel or critical task"| D["Frontier model: Fable 5"]
C --> E["Response to user"]
D --> E💭 Key point: the key isn’t picking a single model, but routing each task to the level of capability, and cost, it actually needs.
How to start testing it
Most small AI models, including the ones French-Owen mentions, are served behind an OpenAI-compatible API, so testing them doesn’t require a new SDK. Just install the official client:
# macOS / Linux
npm install openai
# Windows (PowerShell)
npm install openai
# Alternative with Python, any operating system
pip install openaiThe minimal example is a single chat call. By just changing the model name and the endpoint, you can point to different small-model providers:
import OpenAI from "openai";
const client = new OpenAI({
apiKey: process.env.SMALL_MODEL_API_KEY,
baseURL: "https://api.proveedor-modelo-chico.com/v1",
});
const respuesta = await client.chat.completions.create({
model: "gpt-5.6-luna",
messages: [{ role: "user", content: "Summarize today's tech news in 3 lines." }],
});
console.log(respuesta.choices[0].message.content);
console.log(respuesta.usage);The usage object in the response includes the input and output tokens consumed. With that data you can calculate the real cost per query and decide whether to route that task to a small model or escalate it to a frontier one:
function costoUSD(usage, precioEntrada, precioSalida) {
const entrada = (usage.prompt_tokens / 1000000) * precioEntrada;
const salida = (usage.completion_tokens / 1000000) * precioSalida;
return entrada + salida;
}
// Example prices in dollars per million tokens
console.log(costoUSD(respuesta.usage, 0.15, 0.60));Running this calculation on every call lets you verify in production, with real data, that a model switch actually lowered the spend per query, instead of just assuming it.
💡 Tip: save the usage field from every response in your logs; without it there’s no way to verify the real savings of using a small model.
Impact and analysis
For development teams in Latin America, this price drop matters more than the debate over who has the biggest model. A cost of $0.10 per complex query opens the door to products with real freemium plans, customer support agents that respond at volume, and micro-SaaS products that previously didn’t pencil out with Sonnet-generation models.
French-Owen warns that there’s still infrastructure work needed before this is safe for businesses: new harnesses (the systems that orchestrate the model with tools and memory), protections against prompt injection, and a clear scheme of roles and permissions, so a cheap model doesn’t end up with more authority than it should have over sensitive data.
The flip side of the argument is that demand for frontier models isn’t disappearing: French-Owen still uses Fable 5 and 5.6 Sol for coding, because there the capability ceiling matters more than the price. What changes is that, for the first time, there’s a second viable product category: one that needs a good-enough result at nearly zero cost, not genius at any price.
What’s next
French-Owen expects demand for frontier models to keep growing in fields that require genuinely new discoveries: advanced engineering, hard science, and model training. At the same time, he anticipates that demand for fast, cheap, good-enough models is just taking off for everyday business work: most of the interactions a person today handles with a colleague, a supplier, or a customer.
He closes his essay with an open invitation to anyone experimenting with making small AI models useful: share your findings, because the harness (the layer of tools, permissions, and security around the model) is still a work in progress.
📖 Summary on Telegram: View summary
Try it yourself: run npm install openai, point baseURL to a small-model provider, and measure the usage field of your first response to see the real cost in cents.
Frequently Asked Questions
What is a small AI model?
It’s a language model optimized for speed and low cost per query, in contrast to frontier models, which prioritize maximum reasoning capability regardless of price. gpt-5.6-luna and GLM 5.3 are examples of this category according to Calvin French-Owen’s essay.
How much does it cost to use gpt-5.6-luna, according to this analysis?
French-Owen reports that complex research tasks, like searching through thousands of emails, end up costing tens of cents in API calls, compared to around a dollar with Sonnet-generation models for an equivalent task.
Why aren’t there more consumer AI apps, according to the author?
Because adding a language model to a product introduces a real inference cost on every request, something that didn’t exist in the classic consumer app playbook: cheap site, virality, advertising. That cost drives up the capital needed to scale.
What is “token spewer” work?
It’s the term Peter Reinhardt, co-founder of Segment, uses to describe fast-response and coordination tasks, like calls and following up on pending items, which take up about 95% of daily work at a company, in contrast to “IQ 180” work solving novel problems.
Do small AI models replace frontier models?
No, according to French-Owen: he still uses frontier models like Fable 5 and 5.6 Sol for coding. Small models open up a different product category, one for high-volume tasks with a low margin per query.
What’s still needed for businesses to safely adopt small models?
New orchestration harnesses, protections against prompt injection, and role and permission schemes, according to French-Owen’s essay.
References
- Small Models Have Arrived: original essay by Calvin French-Owen, published on August 26, 2026.
- Artificial Analysis: independent benchmark of language model cost and performance, cited as the source of the Pareto chart.
- OpenAI API Reference: documentation for the compatible API format used by most small-model providers.
📱 Enjoy this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Imagen destacada: Foto de Brecht Corbeel en Unsplash
0 Comments