Frontier model

What it means

A frontier model is one of the most capable general-purpose AI models of its day. It can handle many kinds of work, not just one narrow task, and it does that work as well as or better than the best models already available.¹² The word describes a position, not a product. Think of the lead pack in a marathon: who is in front changes over the course of the race. The Frontier Model Forum, an industry group, says the set of frontier models is “a collection which will shift over time as the field progresses.”²

Executives usually see the word in two places. Companies use it in announcements about their most advanced models. Rules use it too: California’s SB 53, for example, “establishes new requirements for frontier AI developers.”³

Where the word comes from

We could not find a source that says who first used the term, so we don't name anyone. The earliest uses we could open both date from July 2023. On July 6, 2023, Markus Anderljung and 23 co-authors posted a paper titled “Frontier AI Regulation: Managing Emerging Risks to Public Safety.”⁴ On July 26, 2023, Anthropic, Google, Microsoft, and OpenAI announced the Frontier Model Forum, “a new industry body focused on ensuring safe and responsible development of frontier AI models.”¹ The UK government's Department for Science, Innovation & Technology also defined “frontier AI” in a paper it describes as “a discussion paper to support the AI Safety Summit.”⁵

The meaning has changed since then. At its launch, the Forum defined frontier models by how far they went beyond the most advanced existing models.¹ Its own website now describes them as the “state of the art.”² The word has also moved into law: on September 29, 2025, California's governor signed SB 53.³

How others define it

• Frontier Model Forum, at its 2023 launch: “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.”¹
• Frontier Model Forum, today: “‘Frontier AI’ refers broadly to those general purpose AI models that constitute the state of the art, a collection which will shift over time as the field progresses.” For membership, it uses a narrower test: “a general-purpose model that outperforms, based on a range of conventional performance benchmarks or high-risk capability assessments, all other models that have been widely deployed for at least 12 months.”²
• UK government: “highly capable general-purpose AI models that can perform a wide variety of tasks and match or exceed the capabilities present in today’s most advanced models.”⁵
• Anderljung and co-authors: “highly capable foundation models that could possess dangerous capabilities sufficient to pose severe risks to public safety.”⁴

Where they agree and differ

All four definitions describe broad, highly capable models judged against the best models of the moment. They differ on what puts a model over the line. The Forum and the UK paper go by capability: what the model can do compared with others.¹⁵ Anderljung and his co-authors go by danger: whether a model “could possess dangerous capabilities.”⁴ The Forum's membership test adds a time limit, comparing a model only against models “widely deployed for at least 12 months.”²

The strongest objection is about how rules decide which models count. Some rules draw the line by the amount of computing power used to train a model. In a 2024 paper, Sara Hooker argued that “compute thresholds, as currently implemented, are shortsighted and likely to fail to mitigate risk,” and that “The relationship between compute and risk is highly uncertain and rapidly changing.”⁶

This is our perspective, a hypothesis we are still testing rather than a verdict. We think the label is useful for pointing at a moving group of the most capable models, but the label alone doesn't tell a reader whether a model is safe or right for a job. We would change our minds if a public, widely shared test emerged that reliably sorted models by both capability and risk.

In the news

On November 2, 2026, the newsletter SemiAnalysis published GDPx, written by Max Kan, Andrew Megalaa, and Dylan Patel.⁷ A [benchmark](/definitions/benchmark) is a standard test used to compare models. The authors write that “GDPx is a benchmark that covers the most valuable domains for frontier intelligence.” They explain why: “If your goal is to track the progress of frontier models, you should focus on the domains where they will be most impactful.” The domains they list include chip design, cybersecurity, biology, finance, and legal work.⁷ The story shows the term in everyday use: “frontier” marks the top models of the moment, and new tests are built to keep up with them.

What it means for your organization

When a vendor calls a model “frontier,” it is claiming the model is among the most capable right now. The claim has a date on it, because the group of frontier models keeps changing.² It also tells you nothing on its own about whether the model fits your work. Frontier labels and the rules built on them are mostly aimed at the companies that build these models.³ Your organization's job is to decide where that much capability is actually needed.

One question to ask your team: Which of our tasks truly need a frontier model, and how would we check whether a less capable model does them just as well?