Closed model

In short

A closed model is an AI model whose trained weights are kept private by the company that made it. You can use it through an app or a paid connection, but you cannot download it, inspect it or run it yourself.

Also called: Closed-weight model, Proprietary model, Closed-source model, Closed AI model
Topic: The basics
Level: Beginner
First described: There is no single starting point. OpenAI's GPT-4 technical report of March 2023 is a widely cited statement of the closed approach, withholding even the model's size.

Every AI model is, at heart, a large file of learned numbers called weights. With a closed model, that file never leaves the developer's servers. People reach the model through a chatbot app or an API, a paid connection that sends a request in and gets an answer back.

Anthropic's Claude models are an example. Anthropic offers them through its own API and through cloud platforms run by Amazon, Google and Microsoft, priced per million tokens¹. Its documentation explains how to send requests to the models, not how to download them.

Closed developers often keep technical details private too. OpenAI's GPT-4 technical report said it contained no details about the architecture, model size, hardware, training compute or dataset, citing "the competitive landscape and the safety implications"².

The case for closing a model is partly commercial and partly about safety. A July 2024 US government report noted that closed model providers can more easily block harmful uses through limits on prompts and APIs, and that direct access to weights makes safety features easier to remove³. The case against is that outsiders cannot fully check how a closed model works, and users depend on the provider's prices and decisions.

Closed is not all or nothing. OpenAI kept GPT-4's details private², but in 2025 released two open-weight models, gpt-oss-120b and gpt-oss-20b⁴.

Example

A UK retailer that adds an AI chat assistant to its website using a closed model sends each customer question over the internet to the model provider and pays for every reply. It never holds a copy of the model itself.

Why it matters

Many of the best-known AI products run on closed models, so the providers control price, availability and safety rules. Knowing a model is closed tells you that your data travels to someone else's servers.

Questions people ask

What is the difference between a closed model and an open-weight model?

An open-weight model's weights can be downloaded and run by anyone. A closed model's weights stay with the developer, and you can only send it requests through an app or API¹³.

Is ChatGPT a closed model?

GPT-4, one of the models behind ChatGPT, is closed: OpenAI's technical report withheld even its size and architecture². OpenAI did release separate open-weight models, called gpt-oss, in 2025⁴.

Are closed models safer than open ones?

It depends on the risk. A 2024 US government report said closed providers can more easily restrict harmful uses through their APIs, but it did not find enough evidence to justify restricting open-weight models at that time³.

Related terms

Tools that use it

Sources

Back to the glossary