A frontier model means one of the most capable AI models in existence at a given moment, the kind only a few very large companies can afford to build. The line moves: today’s frontier model is next year’s ordinary one.
Think of the cars at the front of a Formula 1 grid. They are the fastest machines this season, built by a few teams with huge budgets. Next season, new cars push the front further, and this year’s leaders become the middle of the pack. “Frontier” describes a position at the front, not a fixed kind of machine.
In a sentence: “Only a handful of labs can afford to train a frontier model, so most apps rent one instead.”
What the term means
The plain point: there is no single official definition. The industry and lawmakers each use their own.
The industry version. In 2023, Anthropic, Google, Microsoft and OpenAI set up a group called the Frontier Model Forum. It described frontier models by what they can do: more than the best existing models, across many kinds of task.1
The legal version (California only). A California law signed in September 2025 measures size instead of skill. It counts the computing power used to train the model: more than 1026 operations, which is a 1 followed by 26 zeros.2 An “operation” here is one small calculation, such as multiplying two numbers.
A simple example
The plain point: the flagship models from the biggest AI labs are the frontier. You usually rent them through an app rather than own them.
When a large lab launches its newest top model, that is a frontier model. You reach it through the company’s chatbot or its paid connection for software. Which labs count, and why the list is short, is set out in Who builds AI. The short answer there is cost: chips, electricity and buildings on an industrial scale.
Just behind sit open-weights models, which anyone can download. Epoch AI, a research group that tracks AI progress, measures how far behind they are.3
“Frontier” is a moving label. A model called frontier a year ago may now be matched by one you can download.
Why it matters to you
The plain point: the word decides which rules apply and which models get the most safety checking.
- Laws are written around it. California’s SB 53 places its duties on the companies that develop frontier models. If you read that a law covers “frontier AI”, check which definition it uses and where it applies.
- “Most capable” is not “best for you”. A smaller model can be faster and good enough for everyday tasks. Choosing an AI model walks through the trade-offs.
- Capability claims come from the makers. “Exceeds the most advanced models” is usually measured on benchmarks, tests the industry picks. How AI is tested explains why those scores need care.
Where the word comes from
The phrase was put forward in July 2023. A policy paper by Markus Anderljung and 23 co-authors, first posted on 6 July 2023, said: “we focus on what we term ‘frontier AI’ models: highly capable foundation models that could possess dangerous capabilities”.4 Later the same month, Anthropic, Google, Microsoft and OpenAI named their joint group the Frontier Model Forum, described above.1
Often confused with
| Word | What it means |
|---|---|
| Frontier model | Means: among the most capable models right now. A position, not a type. |
| Foundation model | Means: a general model trained on broad data that can be adapted to many tasks. California’s law treats a frontier model as a very large foundation model.5 |
| AGI | Means: a debated idea of AI as capable as people across most tasks, with several competing definitions. Being at the frontier does not make a model AGI. See AGI meaning. |
| Frontier (everyday) | Means: a border, or the edge of what is known. The AI use borrows the “edge” sense. |
Questions people ask
Who decides what counts as a frontier model?
There is no single referee. The industry’s Frontier Model Forum goes by skill: models that “exceed the capabilities currently present in the most advanced existing models”.1 California’s SB 53 goes by the computing power spent on training, as the next answer shows. Which definition applies depends on who is talking, and where.
What is the difference between a frontier model and a foundation model?
A foundation model is any general model that can be adapted to many jobs; a frontier model is one of the biggest and most capable of them. California’s law says it directly: “‘Frontier model’ means a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations.”2 See foundation model meaning.
Are open-weight models frontier models?
They are close behind rather than at the front, by Epoch AI’s measure. In May 2026 it reported that “the most capable open-weight models have lagged frontier closed models by an average of four months” since January 2026.6 That gap moves, so treat it as a snapshot.
Learn more on this site
- The full picture of the labs: Who builds AI.
- Rent or hold: Open or closed weighs a frontier model you rent against one you download.
- What it takes to build one: How an AI is actually built.
- The basics: the AI glossary and AI explained for complete beginners.
- Related: Foundation model meaning.
- Related: Superintelligence meaning.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What is a frontier model?
- Any model you can download and run at home
- A model used only near country borders
- One of the most capable AI models at a given moment
Show the answer
Answer: One of the most capable AI models at a given moment. A frontier model is one of the most capable models in existence right now, the kind only a few very large companies can afford to build.
2. Why does the page call ‘frontier’ a moving label?
- Because the models run on moving vehicles
- Because today’s frontier model can be next year’s ordinary one
- Because a law renames the models every week
Show the answer
Answer: Because today’s frontier model can be next year’s ordinary one. ‘Frontier’ describes a position at the front, and new models keep pushing the front forward.
3. How do the industry group and California’s law define a frontier model differently?
- The industry goes by what it can do; California goes by the computing power used to train it
- The industry goes by the company’s size; California goes by the model’s name
- They use exactly the same definition
Show the answer
Answer: The industry goes by what it can do; California goes by the computing power used to train it. The Frontier Model Forum judges by skill, while California’s law counts the computing power used in training.
Related terms: Emergent abilities · Fine-tuning · Foundation model · Inference · Latent space · Mixture of Experts (MoE)
- 1Microsoft, Microsoft, Anthropic, Google, and OpenAI launch Frontier Model Forum, 26 Jul 2023, read at source 1 Oct 2026“The Forum defines frontier models as large-scale machine-learning models that exceed the capabilities currently present in the most advanced existing models, and can perform a wide variety of tasks.” ↑a“The Forum defines frontier models as large-scale machine-learning models that exceed the capabilities currently present in the most advanced existing models, and can perform a wide variety of tasks.” ↑b“The Forum defines frontier models as large-scale machine-learning models that exceed the capabilities currently present in the most advanced existing models, and can perform a wide variety of tasks.” ↑c
- 2California Legislature, SB 53, Transparency in Frontier Artificial Intelligence Act, approved by Governor 29 Sep 2025, read at source 1 Oct 2026“‘Frontier model’ means a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations.” ↑a“‘Frontier model’ means a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations.” ↑b
- 3Epoch AI, Open models lag state-of-the-art closed models by 4 months, read at source 17 Sep 2026 (as cited on Who builds AI): “Since January 2026, the most capable open-weight models have lagged frontier closed models by an average of four months in the Epoch Capabilities Index (ECI)”.↑
- 4Anderljung et al., Frontier AI Regulation: Managing Emerging Risks to Public Safety, arXiv:2307.03718, 6 Jul 2023, read at source 12 Oct 2026“we focus on what we term ‘frontier AI’ models: highly capable foundation models that could possess dangerous capabilities”. ↑
- 5California Legislature, SB 53, read at source 1 Oct 2026a foundation model is “Trained on a broad data set”, “Designed for generality of output” and “Adaptable to a wide range of distinctive tasks.” ↑
- 6Epoch AI, Open models lag state-of-the-art closed models by 4 months, 29 May 2026, read at source 12 Oct 2026“the most capable open-weight models have lagged frontier closed models by an average of four months” in the Epoch Capabilities Index (ECI), measured since January 2026. ↑
California Legislature, SB 53, Transparency in Frontier Artificial Intelligence Act, approved by Governor 29 Sep 2025, read at source 1 Oct 2026. Applies in California, USA.
Epoch AI, Open models lag state-of-the-art closed models by 4 months, read at source 17 Sep 2026 as cited on Who builds AI.
The comparison on this page (the Formula 1 grid) is this site’s own way of explaining, not a claim from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK