Frontier AI: what the term means, who draws the line, and why 10^25 and 10^26 disagree

Frontier AI is the label for the most capable models a handful of labs are building, and it is increasingly the trigger for rules, testing and pledges. But there is no single definition. The European Union, California and the White House's own commitments each draw the line in a different place, and the people who wrote them disagree about whether the line should be drawn by compute at all. This explainer sets out what the term means in practice, who decides, and where the definitions break.

The plain-language meaning

In ordinary use, a frontier model is one at or near the state of the art: the systems from OpenAI, Anthropic, Google DeepMind, Meta, xAI and a few Chinese labs. The word signals capability and risk. It is why the labs' own safety frameworks, the government's testing and the 2026 accord all use it. When a government institute talks about the U.S. frontier, it means the best American models, including some that only vetted users can access.

That everyday meaning is fuzzy. The legal meanings are not, which is why they matter.

The EU line: 10^25, with an escape hatch

The EU AI Act's Article 51 sets a presumption of systemic risk for a general-purpose model when "the cumulative amount of computation used for its training measured in floating point operations is greater than 10^25." The line is a presumption, not a verdict. Developers can argue that their model's capabilities are inferior to those of the most advanced systems.

The Act also builds in a way to move the line. Article 51 says the Commission "shall adopt delegated acts in accordance with Article 97 to amend the thresholds listed in paragraphs 1 and 2 of this Article, as well as to supplement benchmarks and indicators in light of evolving technological developments." The EU decided in advance that a number fixed in law would go stale.

The California line: 10^26, plus revenue

California's SB 53 takes a higher, harder line. As the Future of Privacy Forum summarizes it, a frontier model is a foundation model "trained with more than 10^26 computational operations," counting compute from the initial training and any later fine-tuning or modification. A frontier developer is an entity that trained or initiated the training of such a model.

Obligations scale with money. All frontier developers must publish transparency reports before deploying models and report critical safety incidents to the state's Office of Emergency Services within 15 days, or 24 hours if there is imminent danger. Large frontier developers, those with annual gross revenue above $500 million, must also publish a frontier AI framework covering governance, risk mitigation, cybersecurity and catastrophic risk assessments, and must protect whistleblowers. The attorney general can seek penalties of up to $1 million per violation.

The Center for Data Innovation notes that the 10^26 figure also appears in New York's RAISE Act, Illinois SB 315 and a federal bill it names the Great American AI Act. We have not independently read those texts.

Why a compute number at all

Compute is easy to count and hard to fake. A regulator can ask what a training run cost without understanding the model. That is why it became the proxy for capability.

But it is only a proxy, and the people who study it say it is wearing thin. In a piece dated August 28, 2026, the Center for Data Innovation's Michelle Lopes Maldonado argues that fixed thresholds are technically unfounded and increasingly obsolete as efficiency improves and techniques like distillation spread. Her warning: "Set it too broadly, and relatively benign models face unnecessary regulation; set it too narrowly, and genuinely risky models may escape it." And on drift: "A line drawn to capture a handful of systems is on course to capture a couple hundred which was not the original intention."

Her prescription is that "The definition of frontier models should evolve in response to advances in capabilities and efficiencies, keeping the category focused on the relatively small number of models whose capabilities warrant exceptional regulation." Note the source: it is a policy think tank's argument, not a neutral finding, and it favors lighter regulation of smaller models. The underlying problem it identifies, that compute thresholds age fast, is also the reason the EU gave itself the power to change its number.

The third definition: whoever signs

The most recent frontier line comes from a pledge, not a statute. Six companies signed the White House accord on super intelligence, a voluntary commitment with no enforcement mechanism. We logged who signed and what they promised in our accord tracker. In that setting, frontier effectively means the signers.

That is a very different kind of definition. It is set by who chooses to join. A model from a lab that did not sign, including an open-weight model, sits outside it, even if it is capable enough to worry a government evaluator. We do not have a clean example of this problem in the current record, which is exactly why we flag it as a gap, not a finding.

What the definitions miss

Compute counts the inputs, not the capabilities. Three problems follow.

• Efficiency. As algorithms improve, a smaller training run reaches what once took more, so a model can be dangerous below a threshold.
• After training. Fine-tuning, tool use and agent scaffolding can add capability after the training run, which is why SB 53 counts later modification and the EU lets developers be judged on capabilities.
• Access. A model restricted to vetted users and a model anyone can download are treated the same by a compute threshold, though their risks are not the same.

For how capability is actually measured, which is the alternative to counting compute, see our companion guide on how AGI is measured. For the vocabulary around it, see AGI versus ASI.

Who decides

Today the answer is: several parties, none in charge. Legislators set thresholds in statute. The European Commission can amend its own by delegated act. Labs choose whether to sign voluntary pledges. Government institutes decide what to test.

Our take

We would keep using the word frontier but ask each time who is using it. A compute line is a reasonable starting point and a poor endpoint. The sensible design is the EU's: a number that can be rebutted with evidence of capability and amended as technology moves. What we would watch is whether California or Washington add a mechanism to update the number, and whether any definition ever covers a capable open-weight model released by a lab that never signed anything.