The specialist models

In September, Jev introduced a new kind of model, one that decides instead of writing. You give it some context and a question, and it hands back a decision, yes or no, one option out of a few, or a score, with a probability attached. It was the first model built purely for this, and it’s far faster and cheaper than the general models most of us use. Two weeks later, OpenAI announced its own Decisions API.

Decision models feel like a missing piece of the puzzle, and I expect more models like them, each built for one part of the problem. Some models write, some judge what was written, others handle voice, video, the physical world, or the back-and-forth of a live conversation. I’ve come to think that going after real-life workflows efficiently will take a family of specialized models working together, more than one type of model doing everything.

Behind every agent workflow, whether it ships a feature or runs a campaign, sits a long tail of tiny calls: is this result good enough, should this step be retried. Having a large model write a paragraph about each one is slow, expensive, and inefficient. A software factory is a good example. These decisions would make up a large share of what it needs, combined with frontier code generation, deterministic checks like tests and rules, knowledge graphs, and a few other pieces.

Another place where decision models shine is adtech. Adtech is all about scale, trillions of decisions a day. We’ve run on narrow machine learning models for years, each trained for a single decision: one for pricing, one for bid shading, one for predicting a click. It isn’t the same thing, but the parallel is there. Most programmatic ads go through a chain of them before the page finishes loading, and each one has to be fast, cheap, and good at exactly one thing.

A general decision model, one that reads language and can take on many different calls, opens new ways to optimize. Probably not at the bid level yet, since the bid has to fit in a fraction of a second at the lowest possible cost to protect margin, though who knows how fast that changes. But easily at the page level, deciding what a page is about or whether a creative fits, and asynchronously on the heavier workflows around campaigns and deals.

Agentic advertising is where that gets most concrete. Buyer and seller agents are starting to negotiate the direct deals that still run on email, spreadsheets, or order management systems today, and most of that conversation comes down to quick yes or no questions. If those get cheap, some of the early promises of agentic advertising can turn into reality, and more complex workflows can finally see the light.

Outside our industry, most of this will stay out of sight, and people will never know which part of the family they used. Assistants already route between tiers of the same model, choosing how much thinking a question deserves. The same kind of router could pick between types of models, and from the outside the assistant would just get better.

As these workflows get built, with new types of models added and combined, the part that seems unclear to me is accountability. A decision model returns a number, and once it sits inside a chain of other models, it gets hard to follow the full reasoning behind what comes out the other end. How do we keep that chain understandable, and who answers for it?