---
title: "What is a frontier model?"
description: "A frontier model is a general-purpose AI system at the current leading edge. Origin, EU and California compute lines, and what builders should check."
canonical_url: https://www.oguzhan.co/what-is-a-frontier-model/
author: "Oğuzhan Koçaklı"
author_url: https://www.oguzhan.co/about/
date_published: 2026-10-20T09:18:57+00:00
date_modified: 2026-10-20T09:18:58+00:00
language: en
translations:
  tr: https://www.oguzhan.co/tr/frontier-model-ne-demek/
---

# What is a frontier model?

> **Direct answer.** A frontier model is a general-purpose AI model at the leading edge of what current systems can do. The edge moves. Labs, the Frontier Model Forum, and some laws use the label for the most capable models, often tying it to training compute or to capability tests, not to a brand name.

I use frontier model for a place on a moving map of general-purpose systems. Last year’s frontier model is often this year’s ordinary one. Blogs, safety notes, and statutes do not draw the same line, so the first thing I ask is which definition a sentence is using. Longer notes sit on the [AI hub](https://www.oguzhan.co/ai/).

## Where does the term frontier model come from?

The label went mainstream on 26 July 2023, when Anthropic, Google, Microsoft, and OpenAI announced the [Frontier Model Forum](https://www.frontiermodelforum.org/updates/announcing-the-frontier-model-forum/). In that announcement the Forum defined frontier models as large-scale machine-learning models that exceed the capabilities currently present in the most advanced existing models and can perform a wide variety of tasks.

The Forum’s [about page](https://www.frontiermodelforum.org/about-us/) now uses a wider membership test. A frontier AI model, by that test, outperforms every other model widely deployed for at least 12 months, on ordinary benchmarks or on high-risk capability tests. The Forum says it drew the line wide on purpose, so more organizations can join the safety work. It is a member-funded 501(c)(6) and says it does not lobby. On 20 October 2026 its member page listed Amazon, Anthropic, Google, Meta, Microsoft, and OpenAI. A seat is not a ranking, and it says nothing about training compute.

So the original meaning is relative. A frontier model is defined against the best models that already exist, which is why the set changes every few months.

## How does the EU AI Act define a frontier model?

It doesn’t, at least not by that name. The EU AI Act, Regulation (EU) 2024/1689, speaks of a general-purpose AI model and, for the top tier, a general-purpose AI model with systemic risk. According to [Article 51](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-51), a model lands in that tier if it has high-impact capabilities, judged with technical tools and benchmarks, or if the Commission decides it has equivalent capability or impact under Annex XIII, on its own initiative or after a qualified alert from the scientific panel.

Article 51(2) holds the number people quote. When cumulative training computation is greater than 1025 floating-point operations, the model is presumed to have high-impact capabilities. Recital 111 reads high impact as matching or exceeding the most advanced general-purpose models, and the total counts pre-training, synthetic data generation, and fine-tuning. Article 51(3) lets the Commission move the threshold by delegated act. The Commission’s [AI Act FAQ](https://digital-strategy.ec.europa.eu/en/faqs/general-purpose-ai-models-ai-act-questions-answers) still gives 1025, and so does the AI Act Service Desk text, which is based on the EUR-Lex consolidated version as at 27 July 2026, including the Digital Omnibus edits. I found no delegated act that changes the figure.

A second number causes confusion. The same FAQ treats training above 1023 FLOP, with the ability to generate language, images, or video, as a sign that a model is general-purpose at all. That is a hundred times lower and carries the lighter general-purpose duties, not the systemic-risk ones.

## How does California SB 53 define a frontier model?

California put the words in a statute. Senate Bill 53, the Transparency in Frontier Artificial Intelligence Act, was [signed on 29 September 2025](https://www.gov.ca.gov/2025/09/29/governor-newsom-signs-sb-53-advancing-californias-world-leading-artificial-intelligence-industry/) and took effect on 1 January 2026, as Business and Professions Code sections 22757.10 through 22757.16. Under [section 22757.11(i)](https://leginfo.legislature.ca.gov/faces/codes_displayText.xhtml?lawCode=BPC&division=8.&title=&part=&chapter=25.1.&article=), a frontier model is a foundation model trained with more than 1026 integer or floating-point operations. The count includes the original run plus any later fine-tuning, reinforcement learning, or other material modification of a preceding foundation model.

A large frontier developer is one that, with its affiliates, had gross revenue above $500 million in the prior calendar year. It must publish a frontier AI framework. Every frontier developer must publish a transparency report before or when it deploys a new or substantially modified frontier model. The statute defines catastrophic risk narrowly: a foreseeable and material risk of contributing to the death or serious injury of more than 50 people, or more than $1 billion in property damage, from a single incident. The listed incidents are expert help with a chemical, biological, radiological, or nuclear weapon, a cyberattack or listed crime carried out without meaningful human oversight, and a model evading the control of its developer or user.

California’s line is ten times the EU presumption and counts integer operations too. A model can sit above one line and below the other. And a lab can call a model frontier while publishing neither number.

![Timeline from 2023 to 2026 marking the Forum launch, EU AI Act, UK AISI rename and California SB 53](https://oguzhanco.s3.eu-west-3.amazonaws.com/wp-content/uploads/2026/10/20091609/frontier-model-timeline-2023-2026.webp)

*From the Forum launch in 2023 to SB 53 taking effect in January 2026.*

## What does the UK AI Security Institute count as frontier?

The United Kingdom uses the frontier label but, in every source I checked, attaches no FLOP cutoff and no statute. On 14 February 2025 the government renamed the AI Safety Institute the [AI Security Institute](https://www.gov.uk/government/news/tackling-ai-security-risks-to-unleash-growth-and-deliver-plan-for-change). Its work centres on cyber risk and on chemical or biological misuse. On 12 October 2026 the Institute wrote that it had independently evaluated [more than 55 models before release](https://www.aisi.gov.uk/blog/towards-technical-best-practice-for-frontier-ai-development), and it was clear that this testing is voluntary, not a mandate.

The UK view is closest to the Forum’s original one. Frontier means the most capable systems on the tests that matter for security, measured by evaluation rather than by a compute total.

Who should be allowed to move these lines is its own argument. I covered it around the proposed Standards Authority for Frontier AI in [who gets to set the speed limit](https://www.oguzhan.co/ai-standards-body-pact-or-cartel/).

## How is a frontier model different from foundation, open-weight, and small models?

One model can carry more than one of these labels, because each answers a different question.

A foundation model, in the 2021 Stanford CRFM report [On the Opportunities and Risks of Foundation Models](https://arxiv.org/abs/2108.07258), is trained on broad data, generally with self-supervision at scale, and can be adapted to many later tasks. SB 53 uses the same idea in section 22757.11(f): broad data, general output, many distinct tasks. A frontier model in that statute is such a model past the compute line. The EU general-purpose model is a close cousin, built to serve many downstream systems.

Open-weight means the parameters can be downloaded and run. Training data, code, license, and capability are all separate facts. In a [17 July 2026 analysis](https://www.aisi.gov.uk/blog/how-far-behind-the-frontier-are-leading-open-weight-models-on-cyber), UK AISI found that the leading open-weight models performed on its cyber tests about where closed frontier models were 4 to 7 months earlier, down from a 6 to 10 month gap through most of 2025. So the two labels are not opposites, and the distance between them is shrinking. When the file itself becomes the sensitive object, export rules take over, as I wrote in [China, export controls, and model weights](https://www.oguzhan.co/china-ai-export-controls-model-weights/).

Small is a product cut. On 7 October 2026 Anthropic called [Claude Haiku 5.5](https://www.anthropic.com/claude-haiku-5-5) a small model for high-volume summaries, classification, and support. It can still be a foundation model.

|  | Frontier model | Foundation model | Open-weight model | Small model |
| --- | --- | --- | --- | --- |
| Question | At the current leading edge? | Trained broadly, then adapted to many tasks? | Can the weights be downloaded and run? | Cheap and fast, for a narrower job? |
| What sets it | A moving edge. California: more than 1026 operations. EU cousin: systemic risk presumed above 1025 FLOP. | Breadth. Stanford CRFM, 2021. SB 53 section 22757.11(f). | The release, not the score. | Price, latency, and the provider’s own cut. |
| Overlap | Often a foundation model. Sometimes open-weight. | Can be frontier, ordinary, or small. | Near the frontier, or months behind it. | Can still be a foundation model. |
| October 2026 example | Google calls Gemini 4 Argon its latest frontier model. | Claude Opus 5.5, sold for complex work. | A release choice, see the AISI cyber gap above. | Anthropic calls Haiku 5.5 a small model. |

![Four overlapping circles for frontier, foundation, open-weight and small models, with three unnamed dots](https://oguzhanco.s3.eu-west-3.amazonaws.com/wp-content/uploads/2026/10/20091611/frontier-model-foundation-open-weight-small.webp)

*One model can sit in more than one circle.*

## Which models were labs calling frontier in October 2026?

These are examples, not a ranking. None of the posts below states training compute, so none shows that a named model clears 1025 or 1026.

On 7 October 2026 OpenAI said a [new GPT-6 model](https://openai.com/index/gpt-6-for-everyone/) was rolling out to ChatGPT users, a month after the first GPT-6 models reached paid customers. According to that post, GPT-6 in ChatGPT runs on GPT-6 Sol for paid tiers and GPT-6 Luna for the Free and Go tiers.

Anthropic released Claude [Opus 5.5](https://www.anthropic.com/claude-opus-5-5) on 22 September 2026, Claude [Sonnet 5.5](https://www.anthropic.com/claude-sonnet-5-5) on 28 September, and Claude Haiku 5.5 on 7 October. Anthropic says external evaluators, including Frontier Design and METR, tested Opus 5.5 before release.

On 13 October 2026 Google called [Gemini 4 Argon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/) “our latest frontier model” and said it evaluated the model under its Frontier Safety Framework.

Amazon, Meta, and Microsoft are Forum members. I did not verify a primary October 2026 release page for them, so I am not naming one.

## Why does the frontier label matter for builders?

The label decides less than it seems. Three things it does change are price, access, and paperwork, and each has a practical consequence.

**Price falls fast at the edge.** Anthropic lists Opus 5.5 at $4 per million input tokens and $20 per million output tokens, and Sonnet 5.5 at $2 and $10. Google launched Argon at an introductory $2 and $10 per million, rising to $4 and $20 after 30 November 2026. Opus 5.5 is 20% cheaper per token than Opus 5 on Anthropic’s own pricing table. In practice, a frontier price is a snapshot. I budget on the model id I actually call, and I re-check the bill every release, because last quarter’s frontier tier is often this quarter’s mid tier.

**Access is not the API key.** Google points cyber defenders to its Fairwind Program to use Argon without the ordinary cyber guardrails. Anthropic runs a Life Sciences Verification Program and is widening a Cyber Verification Program for higher-risk defensive work. If your product depends on those capabilities, plan for an application process, not a signup form. If it doesn’t, the default safeguards are part of the product you are buying, and a request can be routed or refused.

**The paperwork sits upstream, mostly.** Under Article 55, the provider of an EU general-purpose model with systemic risk must evaluate the model, assess and mitigate systemic risk, report serious incidents, and protect the model and its infrastructure. The [Commission FAQ](https://digital-strategy.ec.europa.eu/en/faqs/general-purpose-ai-models-ai-act-questions-answers) is the short account. Providers also notify the Commission when training compute meets or will meet Article 51(2). Those duties belong to the lab. Your tools, data, and what your users can reach stay with your product, and that is where your own obligations start. In California, I read a developer’s published framework and transparency report before I treat a safety adjective as a spec.

My working rule is simple. I ask which definition a claim uses, what an evaluation actually covered, and what changes if the weights can be downloaded. “Evaluated before release” needs a scope: biology, cyber, agents, or something narrower.

## Frequently asked questions

### Is a frontier model the same as a foundation model?

No. Foundation means broadly trained and adaptable. Frontier, in the Forum’s sense, means the current leading edge. California adds a compute line of more than 1026 operations, with later fine-tuning and reinforcement learning counted. In that statute every frontier model is a foundation model, but most foundation models are not frontier models.

### Does the EU AI Act use the term frontier model?

Not as a legal category. It says general-purpose AI model with systemic risk, presumes high-impact capabilities above 1025 floating-point operations, and lets the Commission designate a model or change the threshold later. If a vendor treats “not frontier” as “outside the Act,” ask which article they mean.

### Can an open-weight model be a frontier model?

Yes. Open-weight describes the release. Frontier describes capability, or in some laws a compute total. UK AISI measured the leading open-weight models at 4 to 7 months behind the closed cyber frontier in July 2026. Commission guidance says systemic-risk duties get harder once a model is released openly, so testing has to happen before the file ships.

### Do I need a frontier model to ship a product?

Usually not. I ship most work on a smaller or mid-tier model and call a more capable one only for the steps that fail a check. Sending every call to a frontier system is a cost choice, not proof of safety.