Frontier Model

Computer programs that learn from huge amounts of text or images are called AI models. There are very many of these models, and they differ greatly in their capabilities. The term frontier model refers to the strongest representatives at any given time: the systems that can do more than all previous ones. They cost the most money, require the largest data centers, and are built by a handful of companies. The English technical term for this is “frontier model,” meaning a model at the frontier. What is meant is the frontier of what is technically possible at all.

Why a single model can move billions

Whoever owns the current frontier model sets the industry’s benchmark for a few months. All other providers measure themselves against it, and customers base their purchasing decisions on it. That’s why every new release triggers a visible echo on the stock markets. Shares of chipmakers and data center operators sometimes react on the very same day.

Building such a model now costs hundreds of millions of dollars, and for some projects even more. Only a few companies can afford this amount. This leads to concentration: OpenAI, Google, Anthropic, Meta, and a handful of Chinese conglomerates divide this field among themselves. Smaller companies instead build cheaper models for narrowly defined tasks.

Politics has also picked up on the term. In laws such as the EU’s AI regulation, particularly capable models are subject to stricter obligations than small ones. The reason is simple: a system that millions of people use, and whose capabilities no one fully understands, carries greater risks. Anyone releasing a frontier model must therefore test it beforehand for potential misuse.

What propels a model to the top

Three factors are especially decisive: the amount of training data, the number of internal adjustment knobs, and the computing time used. These adjustment knobs are called parameters; you can imagine them as millions of small dials that the model adjusts itself while learning. For years, the rule held: more of everything yields a better model. This observation is known as the scaling law.

Now a second approach has been added. Newer frontier models spend more time thinking before answering difficult questions, checking multiple solution paths along the way. So they’re not just trained to be bigger — they also think more thoroughly during operation. This additional computing time after training especially improves performance on math and coding tasks.

Whether a model truly leads the pack is checked using standardized tests, so-called benchmarks. These are collections of tasks from mathematics, law, medicine, or programming, similar to an exam. However, these tests are controversial: if tasks accidentally end up in the training data, the model already knows the answers. That’s why many experts additionally rely on open comparisons, in which humans rate two answers against each other.

Frontier models in everyday life and in the headlines

In everyday life, people usually encounter frontier models through chat programs and search functions. The paid subscriptions of major providers typically include the strongest model available, while the free versions offer a smaller one. That’s exactly why you sometimes get noticeably different answer quality for the same question. In coding tools, the difference is especially noticeable.

In the news, the term almost always appears in connection with money or regulation. Typical headlines: a provider announces a new frontier model, an agency demands safety reviews, or a corporation orders chips for the next training run. It’s important to distinguish this from related terms. A frontier model is not the same as a large language model — the latter describes the type of construction, the former only the rank.

A common misconception is that frontier models are the best choice for everything. They are more expensive per request and often slower than small models. For routine tasks like sorting emails, a small system is perfectly sufficient. Companies therefore usually combine both approaches, sending only the difficult cases to the expensive model.