Frontier Model

Programs that write texts, generate images, or produce code learn their abilities from huge amounts of examples. Such programs are called AI models. There are now thousands of them, in very different sizes. A frontier model is one of the few that stand right at the edge of what is technically feasible. “Frontier” refers to the boundary in the sense of uncharted territory. The term therefore says nothing about how the model is built, only about its position relative to all others. That is precisely why it is a moving target: what counts as frontier today will be considered standard in two years.

Who coined the term and why

The expression does not come from technology, but from politics. Starting in 2023, it appeared in government papers, for instance at the AI Summit in the United Kingdom. The idea behind it: you don’t need to monitor every small model, but above all the strongest ones. Because only with them is it unclear what they are actually capable of.

This uncertainty is at the heart of the concern. With a very large model, even the maker doesn’t know exactly which capabilities will show up before testing it. Some abilities only emerge beyond a certain size, without anyone having planned for them. Experts speak of emergent capabilities, meaning sudden leaps in ability. A model that becomes too proficient in chemistry knowledge or hacking tools thereby turns into a security issue.

For investors, the term is interesting for a different reason. Only a handful of companies can even afford such models. A single training run often costs hundreds of millions of dollars. Whoever plays in this league decides on market shares across the entire AI business.

What brings a model to the top

Three things matter. First, the amount of computing power used in training, measured in computational operations. Second, the quantity and quality of the training data. Third, the size of the model itself, meaning the number of its adjustable values. These values are called parameters and work like millions of tiny dials that get adjusted during learning.

Regulators have tried to put this into numbers. A well-known threshold lies at 10 to the power of 25 computational operations for training. Anyone above that must meet additional requirements in the EU. Such thresholds are, however, controversial, because they only measure the effort involved, not the outcome. A model trained economically today can be stronger than an expensive one from the day before yesterday.

It is important to distinguish this from related terms. A foundation model is a universally applicable model for many tasks, regardless of how powerful it is. Frontier models are a small subset of these: the strongest representatives. A model that merely sorts photos on a smartphone is neither of the two.

Frontier models in products and headlines

In practice, people often use them without noticing. The paid versions of the major chat assistants usually access the latest top-tier model. Free access tiers frequently get a smaller, cheaper variant. The difference stands out especially with long tasks, such as programming or analyzing an entire document.

In the news, you encounter the term in the context of rules and money. Companies publish safety policies describing at which danger level they will no longer release a model. At the same time, they announce investments in new data centers needed for the next model. Both types of announcements revolve around the same small circle of systems.

A typical misconception is to confuse frontier with good for every purpose. These models are expensive and slow to run. For a search function in an online shop, that would be wasteful. Companies therefore often use a small model for routine tasks and only call on the top-tier model for difficult cases.