A frontier model is an artificial intelligence (AI) model whose capabilities match or exceed those of the most advanced models available at the time. The term describes a position rather than a technology: it marks the leading edge of general-purpose systems, the large models trained on vast data that can write, code, reason and generate images across many tasks. It exists because policymakers and the laboratories themselves wanted a word for the small group of systems whose novelty makes their risks hard to predict, and which therefore deserve closer testing than the rest of the market.

No single definition governs the phrase. Industry bodies, legislators and safety institutes use it in overlapping but distinct ways, and the differences matter for who carries legal duties.

How the label is assigned

Two approaches compete. The first is relative capability. The Frontier Model Forum, the industry body founded in 2023, defines a frontier AI model as "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," according to its website. The Forum says the definition is intentionally broader than "current state of the art". The Frontier AI Safety Commitments, published by the UK government on May 21, 2024, speak of "highly capable general-purpose AI models or systems" that match or exceed the most advanced models.

The second approach is training compute, measured in floating-point operations (FLOP), the arithmetic a computer performs while a model learns. Compute is easy to count and hard to fake, which is why legislators favour it.

  • The European Union's AI Act does not use the word "frontier". It regulates "general-purpose AI models with systemic risk". Article 3(64) defines high-impact capabilities as those that "match or exceed the capabilities recorded in the most advanced" general-purpose models, and Article 51(2) presumes them once training exceeds 10^25 FLOP. Commission guidelines of July 18, 2025 treat models above 10^23 FLOP that generate text, images or video as general-purpose, and require notification within two weeks of crossing the higher line. That regime is explained in depth in PPC Land's pieces on systemic risk and the AI Office.
  • California's Transparency in Frontier Artificial Intelligence Act, known as SB 53, defines a frontier model as a foundation model trained with more than 10^26 integer or floating-point operations, counting later fine-tuning and reinforcement learning. A "large frontier developer" also has annual gross revenue above $500 million, according to the Future of Privacy Forum.
  • New York's RAISE Act, as amended and signed on March 27, 2026, adopts the same 10^26 line and $500 million revenue test, according to law firm Cooley.

The gap between 10^25 and 10^26 is a factor of ten. Epoch AI, a research group, counted 33 publicly disclosed models at or above 10^25 FLOP from 12 developers as of June 2025, starting with GPT-4 in March 2023. Far fewer sit above 10^26.

What happens to a model on the frontier

The label triggers a sequence of work before and after release. Developers run dangerous-capability evaluations, tests designed to detect whether a model offers meaningful help with chemical, biological, radiological or nuclear (CBRN) weapons, cyber attacks, or autonomous action beyond human control. Results are compared against thresholds the company sets in its own safety framework.

Anthropic's Responsible Scaling Policy, published on September 19, 2023, uses AI Safety Levels loosely modelled on biosafety levels, from ASL-1 to ASL-3 and above. OpenAI's Preparedness Framework, updated on April 15, 2025, tracks biological and chemical, cybersecurity and AI self-improvement capabilities against "High" and "Critical" thresholds. Google DeepMind's Frontier Safety Framework, now at version 3.1, defines five Critical Capability Levels and the security tiers attached to them.

Government testers sit alongside. The UK's AI Security Institute and the US Center for AI Standards and Innovation (CAISI), renamed from the AI Safety Institute in June 2025, run pre-deployment evaluations by agreement. When OpenAI launched GPT-5 on August 7, 2025, it classed the reasoning variant as High capability in biology and said red-teaming with partners including CAISI and the UK institute totalled 5,000 hours. Laws then require publication: SB 53 obliges large frontier developers to publish a framework and report critical safety incidents within 15 days, or 24 hours where there is imminent danger.

Origin and evolution

The phrase moved from research into policy in 2023. A paper titled "Frontier AI Regulation: Managing Emerging Risks to Public Safety", submitted on July 6, 2023 by Markus Anderljung and co-authors, defined frontier AI models as "highly capable foundation models that could possess dangerous capabilities sufficient to pose severe risks to public safety." Twenty days later, on July 26, 2023, Anthropic, Google, Microsoft and OpenAI announced the Frontier Model Forum. Amazon and Meta later joined; the Forum lists six members today.

Governments followed within months. The White House executive order of October 30, 2023 set a 10^26 reporting threshold for "dual-use foundation models", and the UK hosted the AI Safety Summit at Bletchley Park in November 2023. That order was revoked on January 20, 2025. The Seoul summit commitments of May 2024 asked signatories, eventually 20 companies, to publish safety frameworks and not to deploy models whose risks could not be kept below their thresholds.

Legislation came next. The EU AI Act's general-purpose model duties applied from August 2, 2025, with a General-Purpose AI Code of Practice published on July 10, 2025, which Meta declined to sign. Anthropic had proposed a transparency framework for frontier developers on July 7, 2025, with coverage tied to revenue of around $100 million or annual spending near $1 billion. California signed SB 53 on September 29, 2025, effective January 1, 2026.

Why the marketing community follows it

Most AI features that advertisers now buy or use run on frontier models or their smaller derivatives. ChatGPT's advertising business, which opened a self-serve Ads Manager to US businesses with cost-per-click bidding in May 2026, sits on top of OpenAI's frontier line. Google's Gemini models power ad products and search features, and every answer they produce passes through inference, billed in tokens.

The frontier also shapes cost and concentration. Training at this scale requires the data centres of a hyperscaler, the cloud companies large enough to build them. UBS estimates suggest OpenAI and Anthropic could account for 48% of Google Cloud revenue in 2027. For agencies and publishers, regulatory status is commercial status: a systemic-risk model can face periodic penalty payments of up to 5% of average daily turnover under the amended AI Act.

Limitations and disputes

The definition moves by design. A relative test means today's frontier model becomes ordinary within a year or two, and the Forum's 12-month reference window acknowledges that. Fixed compute thresholds have the opposite problem: they age. Efficiency gains let developers reach similar capability with less compute, and NVIDIA researchers argued in June 2025 that models under 10 billion parameters can handle 60% to 80% of agentic tasks now given to far larger systems. A 10^26 line may therefore miss capable models while catching bloated ones.

Self-governance draws scepticism. Each laboratory writes its own thresholds, decides when they are crossed and grades its own mitigations. Disclosure laws mostly require publication rather than independent verification, a gap California's governor has asked officials to study.

Openness is contested. A letter of July 24, 2026 signed by 77 organisations, including Nvidia, Google, Meta and OpenAI, defended open-weight models; Nvidia chief executive Jensen Huang wrote that "the world needs both frontier closed models and frontier open models." Anthropic, which did not sign, denied advocating a ban but called for mandatory pre-release testing of all sufficiently capable models. Europe's position adds another tension: Ecosia chief executive Christian Kroll argued that open-weight models let Europe avoid spending billions on frontier training, as the search engine dropped Mistral for Chinese open models.

Not the same as

Foundation model. Any large model trained on broad data and adapted to many tasks. Every frontier model is a foundation model; most foundation models are not frontier.

General-purpose AI model. The EU AI Act's legal category, triggered from 10^23 FLOP. Only the systemic-risk subset corresponds to the frontier.

Large language model. A model built on text. The term describes architecture and modality, not rank, and frontier models are increasingly multimodal.

Superintelligence. A hypothetical AI exceeding human cognitive performance in virtually every domain. Frontier models are existing products; superintelligence is a stated goal of some of their developers.

Recent developments

Anthropic announced Claude Mythos Preview on April 7, 2026, through a restricted partner programme, Project Glasswing. Mozilla attributed 271 Firefox security bugs fixed in Firefox 150 to the model. The cyber capability pushed Washington back towards review: on June 2, 2026, President Trump signed an order creating a voluntary pre-release review for "covered frontier models", with the threshold set through a classified benchmarking process.

On September 18, 2026, California Governor Gavin Newsom signed an order asking for recommendations by November 16, 2026 on whether to require a kill switch for frontier models and onsite verifiers in large laboratories. On October 9, 2026, Anthropic reported that Claude had taken unintended actions on live websites during tests, including submitting a real form, and cut web access in all internal evaluations. New York's RAISE Act takes effect on January 1, 2027.

Timeline

  • March 2023 - GPT-4 becomes the first model trained above 10^25 FLOP, according to Epoch AI.
  • July 6, 2023 - "Frontier AI Regulation" paper by Markus Anderljung and co-authors defines frontier AI models.
  • July 26, 2023 - Anthropic, Google, Microsoft and OpenAI announce the Frontier Model Forum.
  • September 19, 2023 - Anthropic publishes its Responsible Scaling Policy.
  • October 30, 2023 - US executive order sets a 10^26 reporting threshold for dual-use foundation models.
  • November 2023 - UK hosts the AI Safety Summit at Bletchley Park.
  • May 17, 2024 - Google DeepMind publishes Frontier Safety Framework 1.0.
  • May 21, 2024 - Frontier AI Safety Commitments published at the Seoul summit.
  • January 20, 2025 - US executive order on AI safety testing revoked.
  • February 2025 - UK AI Safety Institute renamed AI Security Institute.
  • April 15, 2025 - OpenAI updates its Preparedness Framework.
  • June 2025 - US AI Safety Institute renamed Center for AI Standards and Innovation.
  • July 7, 2025 - Anthropic proposes a frontier transparency framework.
  • July 10, 2025 - EU publishes the General-Purpose AI Code of Practice.
  • July 18, 2025 - Commission guidelines set 10^23 and 10^25 FLOP reference points.
  • August 2, 2025 - EU general-purpose AI model obligations apply.
  • August 7, 2025 - OpenAI launches GPT-5, classed High capability in biology.
  • September 29, 2025 - California signs SB 53.
  • January 1, 2026 - SB 53 takes effect.
  • March 27, 2026 - New York's amended RAISE Act signed.
  • April 7, 2026 - Anthropic announces Claude Mythos Preview and Project Glasswing.
  • April 17, 2026 - Google DeepMind publishes Frontier Safety Framework 3.1.
  • June 2, 2026 - US order creates voluntary review of covered frontier models.
  • July 24, 2026 - 77 organisations sign open-weights letter.
  • September 18, 2026 - Newsom orders kill-switch study, due November 16, 2026.
  • October 9, 2026 - Anthropic reports unintended model actions in evaluations.
  • January 1, 2027 - RAISE Act takes effect.

Summary

Who: Frontier models are built by a small group of developers, including OpenAI, Anthropic, Google DeepMind, Meta, Microsoft, Amazon and xAI. They are overseen by the EU AI Office, California and New York regulators, the UK AI Security Institute and the US Center for AI Standards and Innovation.

What: A frontier model is a general-purpose AI model at or beyond the capabilities of the most advanced models available. Industry defines it relatively; laws define it by training compute, at 10^25 FLOP in the EU and 10^26 operations in California and New York.

When: The term entered policy in July 2023. EU obligations applied from August 2, 2025, California's law from January 1, 2026, and New York's applies from January 1, 2027.

Where: The rules are set in Brussels, Sacramento, Albany, Washington and London, but cover any model offered in those markets.

Why: Capabilities at the leading edge are hard to predict, so governments and developers single out these models for dangerous-capability testing, security controls and public disclosure that ordinary models escape.