For the last few years, building an AI product meant renting the intelligence underneath it from a frontier lab and owning none of it yourself. Renting their models made sense. It was a quick way to start, with good results, and little to no effort. But as the product scales, the cost scales with it. Two labs now collect around 86 cents of every dollar companies spend on model APIs. Sooner or later, a growing business stops renting and builds its own.
Most teams start with a general model
If you look inside most AI-powered products today, you’ll find one general model doing everything. It writes product listings, pulls fields from invoices, and summarizes calls, often thousands of times a day.
There’s a good reason for this. General models are a wonderful place to start. The task can be set up in an afternoon, and it proves a model can do it well. That makes them useful for validating an idea and getting value right away. We are still early in learning how to build AI products, and starting with a capable frontier model is a sensible first move.
But that starting place doesn’t scale cleanly. A general model is designed to be great at writing, coding, knowledge work, biology, and cybersecurity. When you’re extracting insights from emails at high volume, that one skill becomes crucial to your business. Everything else, not so much. So why pay for everything else? The tasks that matter the most also often depend on things a general model was never trained on and is forced to only infer, like data that changes every week or rules only one business knows.
You might expect the next frontier model to fix this. Frontier models continue to improve and are getting cheaper, which is good news for everyone. A better generalist, though, is still a generalist. As companies build AI products in more complex ways, each with its own data and rules, no frontier lab can produce a model for each one of them. They are pursuing frontier intelligence on every front at once.
There is a competitive cost too. When every company in an industry builds on the same few models, none of them gets a clear edge from the model intelligence itself. They all pay the same price per call, too, so none of them wins on margin either. The edge comes from specializing that intelligence on what only you have, like your data, your customers, and the way you work, and there are now more ways to do that than ever.
Always at the frontier
There are now many ways to specialize a model. Fine-tuning, distillation, and reinforcement learning are well established. New methods keep arriving, like weight-space programming, which edits a model’s weights directly, or new ways to adapt a model’s tokenizer to a task. Whatever method you choose, the goal is the same and has been proven to produce results. Specialized models can beat general models on a narrow task.
What holds products back is not the research. It’s the ML resources and expertise required to turn a result into something their business can actually use. A discovery that has been proven to work should be available to every product, not just the ones that can allocate resources for it.
That is the approach we take at Facet. We adopt each new technique for specializing models, each new base model, and each new way of running inference and updating them, and we keep whichever wins on the task’s evals. Most companies in this space are built around a single method, and their customers are tied to it. Facet’s customers get the flexibility of the frontier, and they own their intelligence however they want it produced.
Agents can verify their work
If specialized models work so well, why don’t more teams build them? Only one in ten teams building AI products trains its own model today, and most of those hire outside help. The methods are published, and you could even ask your coding agent to try. What’s hard is the judgment, like designing quality evals, building the training data, picking a method, running the experiments and keeping the model current. Then it has to be served reliably, which is a feat of its own.
All that work to build and maintain a specialized model is a good fit for an agent, for a simple reason. Its outcome can be measured. Coding agents took off first for the same reason, since code can be verified by testing its functionality. In most other work, like reviewing a contract, knowing whether an agent did a good job takes a person reading over its shoulder. Here the model passes its hidden evals or it doesn’t. An agent can try something, measure it and try again, and nobody has to watch every step. Andrej Karpathy’s autoresearch made this loop popular, with an agent running hundreds of training experiments on its own and keeping only the ones that improved its score. The loop continues after launch. The agent watches production for misses and edge cases, and each one it finds becomes a new eval to train against and improve.
Agents are also getting better at this kind of work. In two years, the best agent’s medal rate on Kaggle machine learning competitions went from 17% to 64%, and the length of task agents can finish keeps doubling every few months. Recent projects like Athanor and AutoSFT have teams of agents post-training small open models and improving their training data on their own.
Put all of this together and the work that once needed a scarce ML research team becomes work an agent can do, and verify, autonomously.
Your own frontier lab
No frontier lab will specialize its model because you asked. Facet will.
Let’s say your product pulls fields from freight invoices. You ask Facet for a model from Claude Code, Slack or the web app, with your own data or from scratch. Facet starts working. It asks clarifying questions, writes the evals, builds the training data, trains a small open model and tests it on rows its training never saw.
For most teams, the hard part isn’t seeing the savings. It’s trusting a new model enough to switch, so we obsess over that move. Before you switch, Facet scores the new model against the frontier model you use today, on the same hidden rows. On that freight task, a Facet model matched Claude Haiku’s accuracy at a fifth of the cost per call. The model is served through a managed inference API that drops in where your current frontier endpoint was. There are no sales calls, no ML hires, no strategy meetings and no time pulled away from building the product.
In production, the model keeps getting better. Facet retrains on new data and corrections, and it learns continuously from your product’s own traffic. Every new version is scored on hidden rows before it replaces the last one, so nothing regresses.
Running it should feel as stress free as using a frontier model, where everything works out of the box. A team using Claude Sonnet doesn’t think about the onboarding or maintenance burden, and a team using Facet shouldn’t either.
Yours to keep
Every Facet model is built on an open-weight base under a permissive license and trained on data that never comes from frontier model outputs. It is yours to export and run wherever you choose, even outside Facet.
Every team, its own intelligence
The last era of AI was defined by a few labs building ever larger models. The next belongs to every team that owns the model its product runs on.
