Frontier model

The term started with the companies themselves, not scientists or law. OpenAI, Anthropic and Google DeepMind (Google's big artificial intelligence lab) came up with a label for 'models at the leading edge' together - systems so new that nobody outside has a counter-measure to check their claims. A convenient move: you slap your own sign that says 'we're ahead' before someone else does it for you.

Except the word doesn't measure ability - it measures posture. One model can be strong at language and weak at code. Another can be frontier only because it came out a month ago - and already stop being so, because a competitor shipped a better one on a Thursday afternoon. That's why the label ages faster than the systems themselves.

An example of what the most powerful model looks like in practice today: on October 28, 2026 Google introduced Gemini 4 Carbon, which it calls its strongest. It does not release it to everyone at once, but in steps - first to a limited group of cyber defenders, then to more trusted partners, and for the general public "as soon as possible", with no date. Google also writes that it takes part in the US government's voluntary process for testing models before release.

A second example, from the other side of the fence: on October 25, 2026 MiniMax published MiniMax-M3.1-Flash on Hugging Face, a model with open weights. By the company's own account it is a mixture of experts with about 428 billion parameters in total, of which 23 billion work for each token, and a context of up to 1 million tokens. The million is a ceiling, not a guarantee - usable context depends on GPU memory and server setup. The word 'frontier' is not on the card. In its own table MiniMax leads in two rows and trails in others, and the comparison uses columns the company picked itself. A sign needs no definition when the numbers belong to the maker.

Regulators don't fall for marketing - they grabbed the number behind it. The EU AI Act (the European Union's AI law) sets an automatic presumption of 'systemic risk' on every model trained with over 10 to the 25th power FLOPs - computing operations burned during training. Cross the threshold and the burden of proof falls on the company - it's the one that has to prove the model isn't dangerous.

Crossing the threshold isn't abstract - it leads to concrete obligations. Adversarial tests (deliberate attempts to break the model before someone outside does), a systemic risk report, reporting serious incidents to the European AI office. In other words - the 'frontier' label on paper means more responsibility, not more bragging, even if companies sell it back the other way.

Frontier isn't a category of intelligence. It's a category of the electricity bill.

Wait a minute

I'm allergic to words that sound impressive and mean nothing concrete. 'Frontier' is exactly that - it works as advertising precisely because it has no fixed boundary. Anyone can stand in front of it and say 'we're here', without anyone easily disputing it. It's a trick - the word sells confidence while the actual measure (how much compute, how much data, what the model really does differently) stays hidden behind it.

The distinction I keep for myself: I don't watch who calls themselves frontier, I watch who crosses the threshold the regulator measures in numbers - and then what they do with that crossing. The rest is noise at the leading edge of the wave, gone by next Thursday.

The visual is generated code-art, no third-party images.