
Frontier AI is the policy world’s name for the most powerful AI models in existence. It is used by governments, AI labs and now in law, but it is a moving target: today’s frontier model is next year’s ordinary one. This guide explains what frontier AI means, who builds it, how laws define it and how it relates to super intelligence.
Short answer: Frontier AI means the most capable general-purpose AI models at any given time. The UK government defines it as “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.” Laws in the EU and California now identify frontier models mainly by the computing power used to train them.
What is frontier AI?
Short answer: the cutting edge of general-purpose AI, measured against whatever is most advanced right now.
The term spread in 2023, when the UK government published Frontier AI: capabilities and risks ahead of the Bletchley Park AI Safety Summit. Two features define it:
- General-purpose: the model can do many kinds of tasks (writing, coding, analysis, image understanding), unlike a narrow system built for one job.
- At the frontier: its capabilities match or exceed the best models of the day. The definition is deliberately relative, so the set of frontier models changes every few months.
Because frontier models can do so many things, they can also carry risks nobody tested for, such as helping with cyberattacks or biological weapons. That is the reason governments created a separate category for them.
How do governments and laws define frontier AI?
Short answer: voluntary frameworks describe capability; binding laws use training-compute thresholds.
| Source | Term used | Definition or threshold | Status |
|---|---|---|---|
| UK government discussion paper (Oct 2023) | Frontier AI | Highly capable general-purpose models that match or exceed today’s most advanced models | Working definition for the Bletchley Summit |
| Frontier Model Forum (Jul 2023) | Frontier models | Industry body for developers of the most advanced models | Voluntary, industry-led |
| Frontier AI Safety Commitments, Seoul (May 2024) | Frontier AI | Highly capable general-purpose AI models or systems | Voluntary; 16 original company signatories |
| EU AI Act | General-purpose AI with systemic risk | Presumed above 1025 FLOP of training compute | Obligations apply from 2 Aug 2025 |
| California SB 53 (signed 29 Sep 2025) | Frontier model | Foundation model trained with more than 1026 operations, including fine-tuning | Law; extra duties for “large” developers with over $500 million revenue |
Compiled from GOV.UK, European Commission and Future of Privacy Forum sources listed below.
Compute thresholds are a blunt tool. They are easy to measure, but a smaller, more efficient model can be more capable than a bigger one. The European Commission says its threshold is under review for that reason.
Who builds frontier AI?
Short answer: a small number of very well-funded labs.
The Frontier Model Forum was launched in July 2023 by Anthropic, Google, Microsoft and OpenAI to share safety research and best practices among developers of the most advanced models. At the AI Seoul Summit on 21 May 2024, 16 companies from North America, Asia, Europe and the Middle East, including OpenAI, Google, Anthropic, Meta, Microsoft and Amazon, signed the Frontier AI Safety Commitments, promising to publish safety frameworks and set thresholds at which they would not deploy a model.
In practice, the frontier in 2026 is set by OpenAI, Google DeepMind, Anthropic, Meta (now through Meta Superintelligence Labs) and xAI, with Chinese labs close behind on some benchmarks.
What are examples of frontier AI?
Short answer: the newest flagship versions of ChatGPT, Gemini, Claude, Llama and Grok.
The International AI Safety Report 2026, chaired by Yoshua Bengio, describes what today’s general-purpose systems can do: hold fluent conversations in many languages, write code, generate realistic images and short videos, and solve graduate-level math and science problems. It also notes capabilities keep rising through techniques applied after initial training, such as giving models more time to reason.
Examples age quickly. A model that was frontier when it launched can drop out of the category within a year as newer models pass it.
Is frontier AI the same as AGI or superintelligence?
Short answer: no. Frontier is relative; AGI and superintelligence are fixed targets.
| Frontier AI | AGI | Superintelligence | |
|---|---|---|---|
| Type of term | Relative: “the best right now” | Absolute: matches humans | Absolute: far exceeds humans |
| Exists in 2026? | Yes, by definition | Disputed; most say no | No |
| Used mainly by | Regulators, labs | Labs, researchers | Researchers, labs, critics |
| More detail | This page | What is AGI | Super AI |
The labs building frontier models are explicit that superintelligence is the destination. That is why critics argue frontier rules should cover not just today’s risks but the race itself, the position taken in the Superintelligence Statement. For the jargon around these debates, see our super intelligence glossary.
Why does frontier AI matter?
Short answer: it is where both the biggest benefits and the biggest risks of AI show up first.
New abilities, good and bad, appear at the frontier before they spread to cheaper models. That is why frontier developers are asked to test for dangerous capabilities before release, report serious incidents (a requirement under SB 53) and publish safety frameworks. If superintelligence is ever built, it will first appear as a frontier model.
The bottom line
Frontier AI is not a technology but a position: the leading edge of general-purpose AI. Governments use it to decide which models deserve extra scrutiny, mostly by counting training compute. Watching the frontier is the most direct way to track how close AI is getting to AGI and beyond.
Enjoyed this? Get the SI Briefing
One short email a week: what changed on the road to super intelligence, and why it matters. No spam, unsubscribe anytime.
Frequently asked questions
What is frontier AI?
Frontier AI refers to the most capable general-purpose AI models available at a given moment, the ones at the cutting edge of what AI can do. The UK government defines it as highly capable general-purpose models that match or exceed today’s most advanced models.
What is an example of frontier AI?
The flagship large models from OpenAI, Google DeepMind, Anthropic, Meta and xAI are the standard examples, such as the latest versions of ChatGPT, Gemini and Claude. Which models count changes as the frontier moves.
What is the difference between frontier AI and AGI?
Frontier AI is a relative term: the best AI that exists now. AGI is an absolute target: AI that matches humans across most tasks. Today’s frontier models are not AGI by most definitions.
What is a frontier model under the law?
California’s SB 53 defines a frontier model as a foundation model trained with more than 10^26 operations. The EU AI Act presumes general-purpose models trained with more than 10^25 FLOP carry systemic risk.
Who regulates frontier AI?
As of 2026, the EU regulates it through the AI Act’s general-purpose AI rules, California through SB 53, and the UK and others mainly through voluntary commitments and AI safety institutes that test models before release.
6 Sources
- UK DSIT: Frontier AI, capabilities and risks discussion paper (Oct 2023)
- Computerworld: Major generative AI players join to create the Frontier Model Forum (Jul 2023)
- GOV.UK: Companies agree safety commitments on development of AI, AI Seoul Summit (May 2024)
- European Commission: General-purpose AI obligations under the AI Act
- Future of Privacy Forum: California’s SB 53, the first frontier AI law, explained
- International AI Safety Report 2026