After this explainer you canExplain what makes a model “frontier”, check a maker’s claim against independent tests, and say what new laws require after a serious incident.
The short answerFrontier models are the most powerful AI models. Outside experts test them, and new laws in the EU and California make their makers report serious problems.
In simple words
- Frontier models are the strongest AI models.
- Outside testers check what the makers claim.
- Testers look for dangerous skills, such as hacking.
- New laws ask makers to report serious problems.
The short answerFrontier models are the most capable general-purpose models. Independent testers check makers’ claims. Laws in the EU and California already ask their makers to report serious incidents, and New York and Illinois follow in 2027.
In simple words
- Frontier means the most capable models, at the edge of what AI can do.
- Independent tests matter, because makers’ own claims can be optimistic.
- Safety tests look for dangerous skills, such as helping with cyberattacks.
- In the EU and California, frontier developers must already report serious incidents.
The short answerFrontier status rests on compute thresholds, and the EU can also designate models; a revenue test then picks out “large” developers with the heaviest duties. Capability is measured by held-out benchmarks, preference arenas and dangerous-capability evaluations, and incident-reporting rules differ by deadline and recipient.
Key points
- Thresholds: 10²⁵ operations in the EU; 10²⁶ in California, New York and Connecticut.
- “Large” developers add a revenue test: over $500 million a year.
- Benchmarks can leak into training data, so held-out independent tests matter.
- Incident deadlines differ: 15 days in California, 72 hours in New York and Illinois.
1 Frontier-model laws, 2025 to 2028
- EU duties for the most powerful models
- California 15-day incident reports; safety plans from the largest makers
- Connecticut whistleblower protection
- New York 72-hour incident reportsIllinois 72-hour incident reports
- Illinois yearly independent audits
- Today
2 Makers’ claims, checked
- Backed 9 an outside test supports the claim
- Mixed 6 it holds only in part or in some settings
- Disputed 0 an outside test contradicts it
Words to know
- Frontier model
- One of the most powerful AI models of the moment.
- Independent test
- A test run by people who did not make the model.
- Incident report
- A report to the authorities about a serious problem.
The strongest models
A few companies build the most powerful AI models, such as GPT, Gemini and Claude. These are called frontier models, because they are at the edge of what AI can do.
Who checks them?
Companies often say their own model is the best. Independent testers check these claims. Some let people vote between two answers without knowing which model wrote them. Others run their own tests.
Testers also look for dangerous skills, such as helping someone hack a computer.
New rules
In the EU and California, makers of the strongest models must now report serious incidents. California also asks the biggest makers to publish safety plans. New York and Illinois start similar rules in 2027.
Try it yourself
Pick one new model in our release tracker and check whether an outside test backed the maker’s claim.
Open the model release trackerCheck yourself3 quick questions
A company says its new model is “the best in the world”. What should you look for before you believe it?
Why are the strongest AI models called “frontier” models?
A frontier model causes a serious problem in real use. What do new laws in the EU and California ask its maker to do?
Words to know
- Frontier model
- One of the most powerful AI models of the moment.
- Independent test
- A test run by people who did not make the model.
- Incident report
- A report to the authorities about a serious problem.
What makes a model frontier
There is no single scientific line. Laws use training compute instead, as a simple measure: more than 10²⁵ operations in the EU, and more than 10²⁶ in California, New York and Connecticut.
These models come from a few labs, and we track each new release in our model release tracker.
How they are tested
Makers publish their own benchmark scores, but independent testers matter more. Arena, also called LMArena, ranks models by blind votes from people comparing two answers. Artificial Analysis runs its own tests of quality, speed and price.
Makers’ own scores can be too high when test questions leak into the training data. When a maker’s claim and an independent test disagree, our model tracker marks the claim as mixed or disputed.
Safety tests and incidents
Labs and government institutes also test for dangerous abilities, such as helping with cyberattacks. They also check whether a model behaves differently when it thinks it is being tested; a 2025 study found that leading models can often tell.
When something goes wrong in real use, we log it in our AI incident log, with what the company said and what is still unknown.
What the laws now ask
In the EU, makers of models with “systemic risk”, the EU’s term for the most powerful models, must test them, reduce risks and report serious incidents. California asks every frontier developer for incident reports within 15 days, and the largest must also publish a safety framework.
New York and Illinois ask for incident reports within 72 hours from 2027. Illinois also requires yearly independent audits from 2028. Connecticut already protects staff who warn about catastrophic risks.
Try it yourself
Pick one new model in our release tracker and check whether an outside test backed the maker’s claim.
Open the model release trackerCheck yourself3 quick questions
How do laws usually decide which models count as frontier?
A maker says its model beats all rivals on its own benchmark. In Arena’s blind votes it ranks fifth. How would our model tracker likely mark the claim?
A frontier developer has a critical safety incident in California. No lives are at risk. What does California’s law ask?
Defining the frontier
The EU presumes systemic risk above 10²⁵ training operations (Article 51), and the Commission can also designate models. California’s SB 53 defines a frontier model as one trained with more than 10²⁶ operations, counting later fine-tuning and reinforcement learning.
California, New York and Connecticut add a revenue test for “large” frontier developers: more than $500 million a year with affiliates. Large developers carry the heaviest duties.
How capability is measured
Benchmarks test knowledge, coding and agent tasks, but their questions can leak into training data and inflate scores. Researchers wrote GSM1k, a new test built to match the popular GSM8k maths test, and saw accuracy drop by up to 8%; the strongest models showed few signs of this.
Held-out tests run by independent groups are harder to game, but not free of ties: OpenAI funded the FrontierMath benchmark and has access to part of its problems.
Preference arenas such as Arena (LMArena) turn many blind pairwise votes into ratings with confidence ranges. Votes measure preference, not checked correctness, and a 2025 study found that testing many private variants before release can inflate a lab’s place.
Artificial Analysis publishes its own intelligence index alongside speed and price.
Dangerous-capability evaluations
Labs and government bodies test for help with cyberattacks, chemical or biological weapons, and loss of control. Examples include the UK AI Security Institute and the US Center for AI Standards and Innovation.
Evaluators also watch for models that detect testing and behave differently. A 2025 study found that frontier models can often tell evaluation transcripts from real use, though not yet as well as people, which makes results harder to trust.
Reporting regimes compared
EU: models with systemic risk must report serious incidents to the AI Office, and to national authorities where needed, without undue delay (Article 55). California: critical safety incidents go to the Office of Emergency Services within 15 days, or within 24 hours to an appropriate authority if lives are at risk.
New York (from 2027): reports to a new office at the Department of Financial Services within 72 hours, or within 24 hours to an appropriate authority if lives are at risk. Illinois (from 2027): reports to the state emergency management agency and the Attorney General within 72 hours, or 24 hours if lives are at risk. Yearly independent audits start in 2028.
California, Connecticut and Illinois also protect employees who report catastrophic risks.
Try it yourself
Pick one new model in our release tracker and check whether an outside test backed the maker’s claim.
Open the model release trackerCheck yourself3 quick questions
A developer trains a model with 2 × 10²⁶ operations and has $300 million in yearly revenue with affiliates. How does SB 53 classify it?
A model scores 95% on a public benchmark but 60% on a new held-out test of the same skill. What is a likely explanation?
A critical safety incident at a frontier developer creates an imminent risk of death. What does SB 53 require?
Sources
- AI Act, Article 51 (EU AI Act Service Desk)
- AI Act, Article 55 (EU AI Act Service Desk)
- S8828, RAISE Act amendments (New York State Senate)
- SB 53 (California Legislature)
- Text leaderboard (Arena, LMArena)
- Independent model tests (Artificial Analysis)
- A Careful Examination of Large Language Model Performance on Grade School Arithmetic (Zhang and others, 2024, arXiv)
- The Leaderboard Illusion (Singh and others, 2025, arXiv)
- Large Language Models Often Know When They Are Being Evaluated (Needham and others, 2025, arXiv)
- FrontierMath benchmark (Epoch AI)
- AI Security Institute (UK government)
- Center for AI Standards and Innovation (NIST)
This explainer was generated by AI systems under the editorial rules of Silicon AI News and checked against the sources it lists. The live parts come from our checked stories, trackers, model comparison and rules checker.