closed model

in English: closed model / proprietary model

basicwidespread

An AI model whose internals, i.e., its weights, are not released by the developing company: you can typically use it only through its service or partners, under its own terms.

In plain terms

A closed model is like a restaurant's secret recipe. You can order and eat the dish, but you do not get the recipe, cannot enter the kitchen, and cannot cook the same thing at home. The company runs the model on its own or a partner cloud provider's servers. You only send a query and receive an answer.

The most well-known chatbots typically rely on such a model: the GPT models behind ChatGPT, Anthropic's Claude, and Google's Gemini are all examples. The company decides how much it costs, what you may ask it, and when it is replaced with a newer version. These same companies sometimes also release open models, such as Google with Gemma and OpenAI with gpt-oss.

The opposite is the open model. Its "weights" can be downloaded by anyone and run on their own machine. The weights are the numbers formed during training; in a large model, there are billions of them, and they store what the model has learned. The closed model, by contrast, is convenient: it does not require expensive hardware or expertise, and often the most powerful models are of this type.

How it works

The company trains the model at great cost and keeps the final weights as a trade secret. These are stored on its own servers or in the cloud of contracted partners.
You send your query to the provider's servers via an application, website, or programming interface (API).
The model calculates the answer there, and only the finished text, image, or sound comes back to you; the model itself never does.
The company enforces rules in the meantime: it filters prohibited requests, measures usage, and bills based on subscription or consumption.
The company updates or retires the model on its own schedule, so the same service may behave differently over time.

An everyday example

You open ChatGPT on your phone and ask it to draft a complaint letter to your service provider. Your request goes to OpenAI servers, where the response is generated. You cannot download the model, and it does not work without internet.

Who is affected

everyoneMost everyday AI tools are built on closed models, meaning what you type goes to a company's servers.

entrepreneur, executiveIf you build your company's processes on a closed model, you will depend on the provider's prices, terms, and decisions.

developerYou can quickly build an application or agent on top of it, but you cannot see the model from the inside, and an update could change your program's behavior. → Code

lawyer, citizenWith sensitive or personal data, it matters whether they end up on an external company's machines, even in another country. → Law

A common misconception

MythA closed model is safer and more reliable because a large company stands behind it.

In realityA closed model simply means you cannot see inside it. A large company may test extensively, but it is harder to verify from the outside how accurate or safe the model is than with an open one. Openness or closedness alone does not make a model more trustworthy.

Limits and risks

What you type leaves your machine, so check what the data processing notice allows before providing confidential information.
The provider can raise prices at any time, modify the rules, or discontinue the model variant you are used to.
Independent researchers can only study it from the outside, by asking it questions, so its biases and errors are harder to uncover.
It is mostly usable only with internet access and the provider's operational servers. Exceptions include smaller closed models built into some phones or computers.

Similar, but not the same

open model – You can download the weights of an open model and run them on your own machine. A closed model is accessible only through its developer or its partners' service.
paid model – Closedness is not about cost: you can use a closed model for free, and running an open model can also cost money.
feketedoboz-modell – A black box means that the model's decisions are hard to explain even from the inside, and this applies to open models as well.

Related concepts

Broader

MI-modell, tulajdonosi szoftver

Narrower

–

On the edge

open model, open-weight model, model weights, provider dependency, API, foundation model, frontier model

A brief history

The expression spread on the model of closed-source (proprietary) software. With large language models, the turning point came when OpenAI in 2020 made GPT-3 available only through a programmer's interface (API), not releasing the model itself.

Where things stand

As of late 2026, the best-known chatbots and many coding agents are built on closed models. Meanwhile, open models have become a strong alternative, and many companies use a mix of both.

Status: 2026 November

In the news

Appears in 2 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Chinese Qwen search tool removed from US Federal Register site; Alibaba releases new Qwen image modelSep 17, 2026 · Agent Age
Meta releases open-weight Muse Glimmer model alongside Zuckerberg's 'personal superintelligence' essayAug 10, 2026 · Agent Age

Also known as

closed model · closed-source model · closed-weight model · tulajdonosi modell · closed model · closed-source model · closed-weight model · proprietary model

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