general-purpose AI model (GPAI)

advancedwidespread

An AI model not designed for a single task but usable for many, for which the EU AI Act prescribes specific rules.

In plain terms

A general-purpose AI model is a legal term: the EU Artificial Intelligence Act (AI Act) uses this name for models that write text, translate, code, or analyze images—meaning they are not good for just one thing but for many tasks. Such models power the most well-known chatbots.

Think of an electric motor. The same motor can end up in a washing machine, a drill, or a lift, and its manufacturer cannot know in advance what it will be used for. So regulation does not ask what the motor is for, but whether you have documented it properly, so that whoever builds it in knows what they are working with.

This is the essential difference compared to other parts of the regulation. Most rules are tied to the use case (such as recruitment or credit assessment), whereas GPAI rules target the model itself and its creator. Stricter requirements apply to the largest models carrying so-called systemic risk.

How it works

Classification: a model is considered general-purpose if it can perform many different tasks well and can be integrated into many different products. The European Commission's guidance also uses the amount of computing power used for training as a benchmark for this.
Basic obligations: the model provider maintains technical documentation, informs companies building products on the model, keeps a copyright policy, and publishes a summary of the training data.
Top tier: if the computational demand of training exceeds a very high threshold (10^25 floating-point operations), the regulation presumes the model poses a systemic risk. The Commission may also classify models with lower computational requirements in this category. In such cases, risk assessment, testing, reporting of serious incidents, and strong cybersecurity are mandatory.
Compliance verification: providers can sign a voluntary code of practice, which offers a tangible way to fulfill obligations.
Oversight: these rules are not enforced by national authorities but by the AI Office operating within the European Commission. Since August 2026, the Commission can also impose fines for violations.

An everyday example

A Hungarian online store is introducing a customer service chatbot built on a large American company's language model. The model's creator is subject to GPAI rules, so it must provide documentation to anyone building a chatbot on the model: what the model can do and what its limitations are. Relying on this, the online store and the chatbot developer can make responsible decisions about its use.

Who is affected

developerIf you build an application or agent on a pre-trained model, you have the right to receive information from the provider about the model's capabilities and limitations. → Code

entrepreneur, executiveYou must know that the model creator's obligations do not exempt yours: you are responsible for the specific use case.

lawyer, citizenThe public summary of training data and the copyright rules provide new footholds for legal enforcement and transparency. → Law

everyoneThe developers of the models behind the chatbots used daily must, for the first time in the EU, account for how those models operate under a law written specifically for them.

A common misconception

MythA general-purpose AI model is the same as human-level, all-capable artificial intelligence.

In realityThis is a legal category, not a peak capability: it means the model can be used for many types of tasks. The large language models currently available fall into this category, even though they are far from infallible.

Limits and risks

The rules bind the model's developer and do not protect you directly: even if a model complies, its answers can still be wrong.
The compute threshold is a rough metric: due to more efficient training methods, a model with lower compute requirements can still be very powerful.
The code of practice is voluntary, and open-source models receive partial exemptions (unless they pose systemic risks), so not all providers comply with the rules in the same way.
The regulation applies only to models offered on the EU market, and models launched before August 2025 have a grace period until August 2027.

Similar, but not the same

artificial general intelligence (AGI) – AGI is a concept denoting future human-level capabilities, while GPAI is a legal classification of models existing today.
High-risk AI system – The high-risk classification depends on the use case (e.g., recruitment, credit scoring), whereas GPAI rules apply to the model itself, regardless of purpose.
foundation model – The term base model is a technical term for roughly the same model type, while GPAI is the legal definition set out in the EU regulation, which carries obligations.

Related concepts

Broader

EU AI-rendelet, AI regulation, AI-modell

Narrower

GPAI model with systemic risk, large language model

On the edge

foundation model, code of practice, AI-hivatal, High-risk AI system, transparency obligation

A brief history

This category did not appear in the European Commission's 2021 draft regulation. The issue of general-purpose AI was raised by the Council of the EU in 2021–2022, but specific rules for models took shape after the launch of ChatGPT at the end of 2022: the European Parliament proposed requirements for "foundation models" in 2023, and the final solution emerged in the political agreement of December 2023. The regulation was adopted in 2024 and entered into force on August 1, 2024; GPAI obligations have been applicable since August 2, 2025.

Where things stand

GPAI obligations have applied since August 2025, and since August 2026 the European Commission, acting on the basis of the AI Office's procedure, can also impose fines. Models placed on the market earlier must comply by August 2027.

Status: 2026 November

In the news

Appears in 5 Agent Age stories; the curve shows weekly occurrence in the news magazine.

AI Act obligations for general-purpose AI models: biological risk and incident reportingOct 19, 2026 · Jog
EU AI Act after the omnibus: high-risk rules slip, transparency obligations do notSep 28, 2026 · Jog
GPAI Code of Practice transparency chapter published as Commission readies guidelinesSep 19, 2026 · Jog
European Commission enforces AI Act: GPAI obligations binding, transparency rules in effectSep 14, 2026 · Jog
EU Commission can now inspect AI models, restrict market access and fine providers under AI ActAug 3, 2026 · Jog

Also known as

GPAI · GPAI-modell · general-purpose AI model · general-purpose artificial intelligence model · general-purpose AI model · general-purpose AI model · general purpose AI · GPAI model

Sources

European Commission: the AI regulatory framework

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.