AI 101 — What is a frontier model?
A frontier model is the most capable general-purpose AI model in the world at the moment you ask — the leading edge itself, not a product tier with a fixed spec sheet. Nothing stays a frontier model for long. The phrase describes a position in a race, which is why it moves every few months, and why the definition carries more weight than it looks like it should.
You have read the phrase a dozen times this week without anyone defining it. It shows up in regulation, in lab safety policies, in price arguments, and in the sentence "frontier labs" — a phrase that quietly assumes you already know which labs those are and why they are grouped together.
Why it matters right now
Three things hang on the term, and none of them are academic.
Law. The European Union's AI Act does not use the words "frontier model", but it regulates the same object. A general-purpose model trained with more than 10^25 floating-point operations is presumed to have high-impact capabilities and lands in the systemic-risk bucket, with extra obligations for evaluation, incident reporting and cybersecurity, under Article 51 of the AI Act. Those obligations have applied since 2 August 2025. The threshold is not sacred: the Commission can adjust it by delegated act, and it can also designate a model below the line if the capabilities justify it.
Company policy. Labs write their own thresholds, and those thresholds decide what gets gated. OpenAI rated GPT-6 Cyber "Critical" for cyber capability under its Preparedness Framework this week — a rating above the previous generation's High — and restricted it to vetted defenders, per OpenAI's announcement. That is a private definition of the frontier doing public work: the model exists, and an application form is what stands between it and everyone else.
Money. The frontier is where the price premium lives. Mozilla's September report put the capability gap between the strongest closed frontier models and the best open-weights models at about 4.4 months, and priced the whole premium off that number. If you are paying frontier prices, you are paying for a position on a moving line.
The mental model
There are two working definitions, and both are relative.
The first is capability-based. The Frontier Model Forum — launched in July 2023 by Anthropic, Google, Microsoft and OpenAI — defines frontier models as "large-scale machine-learning models that exceed the capabilities currently present in the most advanced existing models," according to the forum's founding announcement. Read that twice: exceed the capabilities currently present. It is a test you can only pass once, and only until someone else passes it.
The second is a proxy: count the compute. Training a model takes a measurable amount of arithmetic, and more arithmetic has historically bought more capability, so regulators use a number as a stand-in for a judgment. The International AI Safety Report, the 96-expert review chaired by Yoshua Bengio and published in January 2025, takes the capability route instead, focusing on general-purpose AI "at least as capable as today's most advanced" systems.
Put together, the definition is: the model that matches or beats the best thing that exists today. A lawyer, a safety team and a procurement officer can each apply that sentence and come back with a slightly different list of names, which is exactly the confusion worth understanding.
The parade
Picture a parade moving down a street. The front of the parade is a real place — it has the marching band, the police escort and the best view — but it is not a fixed group of people. Marchers fall back, new ones move up, and the front is always occupied by somebody.
That is the frontier. Being at the front is genuinely different from being in the middle: the front gets the attention, the scrutiny and the rules, and it pays the cost of being first to a problem. A model that led the parade in early 2023 is somewhere in the middle of the procession now, still marching, no longer interesting.
The distances matter too. The gap between the front row and the row behind it is small and getting smaller — months, not years. That is the entire commercial argument for open weights: if the second row is four months behind, most work can wait four months, or use the cheaper marcher today.
Common misconceptions
"Frontier means the biggest model." It means the most capable one on the tasks people care about, which is a different measurement. Parameter counts have stopped being a reliable proxy in either direction: a model can be enormous and mid-table, or small and close to the front, because what a bigger model buys you has always been indirect. Capability also depends on the harness wrapped around the model, which is why the same weights can feel excellent in one product and sluggish in another.
"Frontier means closed." Not anymore. Open-weights labs land near the front — a few months back rather than a generation back — which is why the open-versus-closed question is now about distribution and price rather than whether the capability exists outside a handful of buildings.
"Frontier is a badge you keep." It expires, usually within a year, and sometimes within a quarter. Any article calling a model "frontier" is making a claim about a date, not a property.
"Frontier models are the right buy." They are the right buy for the narrow band of work that has to land now and is hard enough to need the extra capability. For routine work you will still be running next quarter, the premium buys you a head start you did not need.
"The compute threshold measures danger." It measures arithmetic. The EU treats it as a presumption, not a verdict, and keeps the right to designate below it — while labs set their own, separate thresholds for capability categories like cyber or biosecurity. Three definitions, three numbers, one word.
Where to learn more
The primary documents are short and readable. The AI Act's Article 51 is four paragraphs. The International AI Safety Report is long, but its opening chapter defines the object cleanly. If you want the industry's own version, the Frontier Model Forum's membership page carries the capability definition in one sentence.
Related reading: AI 101 — What is an AI benchmark? explains how capability gets measured in the first place, and AI 101 — What are model parameters? covers the number people reach for when they want a shortcut.
Does the word "frontier" still mean anything, now that three different organisations define it three different ways? Write to us via our contact page.
