Who are the frontier AI labs? The four tests that decide membership

There is no official register of frontier AI labs. In practice four tests get used, and they produce different lists. A company can be counted by a regulator's compute threshold, by membership of the industry body, by a signature on an international safety commitment, or by whether it publishes a capability-threshold safety framework. Those four groups overlap but do not coincide, which is why counts of frontier labs vary from three to sixteen.

Test one: the compute thresholds in law

The only tests with legal force, and they are set at different levels.

The EU AI Act presumes that a general-purpose AI model carries systemic risk when its cumulative training compute exceeds 10 to the 25th floating-point operations. Crossing that presumption brings additional duties around evaluation, serious-incident reporting and cybersecurity.

California's Transparency in Frontier Artificial Intelligence Act sets its bar an order of magnitude higher, at more than 10 to the 26th computing operations, and adds a revenue test: the heavier obligations apply to a frontier developer with annual revenue of $500 million or more.

So a lab can be a systemic-risk provider in Brussels and not a large frontier developer in Sacramento. We have been through how these thresholds were chosen and what they fail to capture separately.

Test two: the industry body

The Frontier Model Forum launched on 26 July 2023 with four companies: Anthropic, Google, Microsoft and OpenAI. Its stated purpose is advancing frontier AI safety research and enabling independent, standardized evaluation of capabilities and safety.

It describes itself as open to other companies that work on frontier models and demonstrate strong commitment to their safety, which makes membership a statement of intent as much as a measure of scale.

The founding four are instructive for what they include. Microsoft is there as a deployer and investor rather than as the trainer of the largest models. That is a reasonable way to draw the line for an industry body, and a poor way to measure capability.

Test three: the Seoul commitments

The widest of the four lists. At the AI Seoul Summit in May 2024, sixteen companies signed the Frontier AI Safety Commitments: Amazon, Anthropic, Cohere, Google and Google DeepMind, G42, IBM, Inflection AI, Meta, Microsoft, Mistral AI, Naver, OpenAI, Samsung Electronics, the Technology Innovation Institute, xAI and Zhipu.ai.

The commitments are voluntary. Signatories agreed to publish safety frameworks describing how they would measure and manage risk, and not to develop or deploy models where identified risks could not be sufficiently mitigated.

The list is geographically broader than any other test, with Chinese, South Korean and Emirati signatories. It is also the least demanding, because signing is free and the undertaking is self-assessed.

Test four: a published safety framework

The most informative test, because it requires the company to have done something specific and to publish it.

Three labs maintain frameworks that set capability thresholds and pre-commit to safeguards when those thresholds are crossed. Anthropic's Responsible Scaling Policy was at version 3.4, effective 8 July 2026. OpenAI's Preparedness Framework reached version 2 on 15 April 2025. Google DeepMind published the third iteration of its Frontier Safety Framework on 22 September 2025, updated to version 3.1 in April 2026. We have gone through DeepMind's capability levels in detail.

Having a framework does not make a lab safe, and a framework can be rewritten by the company that wrote it. But it is the only one of the four tests that produces a document you can read and hold a lab to later.

What does not make a lab a frontier lab?

Three things that get mistaken for it.

Valuation. A company's funding round tells you what investors expect, not what it has trained. Several of the best-funded AI companies in the world build products on top of other people's models.

Benchmark leadership on a narrow task. A specialist model can top a coding or maths leaderboard without being anywhere near the frontier in general capability, which is part of why benchmark scores stop separating models once the leaders cluster at the top.

Owning data centers. Capacity is a prerequisite for frontier training and not evidence of it. Several of the largest compute owners rent most of it out.

The confusion is understandable, because all three are correlated with frontier capability and all three are easier to observe than training compute, which is disclosed at the lab's discretion.

Can anyone verify a lab's compute figure?

Not independently, and this is the quiet weakness under the statutory tests.

A training run's compute is known to the lab, to its cloud provider, and to its chip supplier. It is not published as a matter of course, and model cards vary widely in whether they state it at all. Outside estimates are built from hardware counts, training duration and reported costs, which produces ranges rather than numbers.

So the EU's 10 to the 25th presumption and California's 10 to the 26th threshold are both triggered by a quantity that, in most cases, only the regulated party can measure. Regulators have asked for self-declaration, which works until somebody has a reason for it not to. Our explainer on how compute is counted and costed goes through the arithmetic.

So how many frontier labs are there?

Depends which question you are asking.

If the question is who trains the largest models, the answer has been narrowing. The State of AI Report 2026, published on 8 October 2026 by Nathan Benaich of Air Street Capital, argued the frontier had reduced to three companies.

If the question is who has committed to frontier safety practices, the answer is sixteen and includes companies whose largest models are well behind the leaders.

If the question is who is legally caught, the answer depends on the jurisdiction and on training compute nobody outside the lab can verify independently.

All three answers are defensible. What is not defensible is quoting one and implying the others.

What about the Chinese labs?

They complicate every test except the first.

DeepSeek, Alibaba's Qwen team, Moonshot, Zhipu and others release competitive models, often with open weights, and some of their results sit close to the frontier on public benchmarks. Zhipu signed the Seoul commitments; most of the others have not signed anything comparable.

They are not members of the industry body, and they are not subject to the EU and California statutes in the way a US lab is, though the EU rules apply to models placed on the EU market wherever they were trained.

The practical consequence is that a frontier-lab list built from Western safety institutions will systematically understate how many organizations can train a near-frontier model. A list built from benchmark results will not.

Why does the definition matter at all?

Because obligations attach to it.

The EU's systemic-risk presumption, California's reporting duties and every proposed licensing regime need a definition of who is covered. Compute is the proxy that ended up in statute because it is the only input that is countable in advance, even though it correlates imperfectly with capability, and the correlation has been weakening as efficiency improves.

A model trained with less compute today can match one trained with more two years ago. A threshold fixed in statute therefore catches a shrinking share of capable models over time unless it moves, which is an argument several regulators are now having with themselves.

How to read a claim about frontier labs

Three questions.

• Which test is being applied: compute, membership, signature, or published framework?
• Is the compute figure disclosed by the lab, estimated by an outsider, or asserted by the writer?
• Does the list include deployers and investors alongside trainers, and does the claim depend on that?

A list of frontier labs is always a list of companies that passed one particular test. The test is the interesting part.