What is GPT? The AI model family, explained simply

In plain words

GPT is not an app. It is the name of a series of AI models made by OpenAI. Each model is a large language model: a program that learned from huge amounts of text and can write, summarise, translate and answer questions.

The name stands for generative pre-trained transformer. Generative: it creates new text. Pre-trained: it first learned from a vast pile of writing before being shaped for any job. Transformer: the design of the network inside.

OpenAI numbers each new generation. GPT-1 came in 2018, and each one since has been larger and more capable; the box above shows the newest. ChatGPT is one app built on these models, and many other apps use them too.

Everyday example

An engine and the cars that use it

A carmaker builds an engine and keeps improving it: version 1, 2, 3. The same engine goes into a hatchback, a taxi and a delivery van. GPT is the engine. ChatGPT is the carmaker's own car. Other companies buy the engine and put it into their own vehicles, which is what happens when apps use GPT through OpenAI's API.

Where the example breaks: An engine works the same way every time. A GPT model's answers can vary from one try to the next, and each version is trained afresh rather than upgraded part by part.

Latestas of October 2026

Newest generation: GPT-6, from September 2026
Top model: GPT-6 Astra, which OpenAI calls "the world's most intelligent and aligned model"
In ChatGPT: GPT-6 Sol and GPT-6 Luna, from 7 October 2026
Downloadable: gpt-oss-120b and gpt-oss-20b, open-weight models from August 2025

These details change often. The rest of this page explains what it is, which stays true.

How it actually works

Every GPT model is built in the same broad way, at a much larger scale each time.

Pre-training on text

The model reads a huge collection of text split into tokens and learns to guess the next one. Getting better at that guess forces it to pick up grammar, facts and patterns of reasoning. This is pre-training.

Turning it into an assistant

A pre-trained model only continues text. OpenAI then trains it on example conversations and on people's ratings of answers, so it follows instructions and answers helpfully. This uses instruction tuning and RLHF.

Thinking before answering

Since late 2024, OpenAI's models can spend extra time on hidden working steps before replying, which helps with maths, code and planning. This style is called a reasoning model.

Many sizes for many jobs

Each generation comes in versions that trade speed against capability, often with their own names. A larger, slower version handles hard problems, while lighter, faster ones serve everyday chat and high-volume work.

Made by

OpenAI

First described

June 2018, in the paper "Improving Language Understanding by Generative Pre-Training"

Generations

GPT-2 (2019), GPT-3 (2020), GPT-4 (2023), GPT-5 (2025), GPT-6 (from September 2026)

Open-weight versions

gpt-oss-120b and gpt-oss-20b (August 2025), downloadable under the Apache 2.0 licence

People sometimes say "GPT" for any chatbot model. Strictly it means OpenAI's family. Others, such as Claude, Gemini and Llama, use a similar transformer design under their own names.

Questions people ask

Is GPT the same as ChatGPT?

No. GPT is the family of models; ChatGPT is an app that uses them. Think of GPT as the engine and ChatGPT as one car built around it. Many other apps, including some of Microsoft's Copilot tools, have also used GPT models.

What does GPT stand for?

Generative pre-trained transformer. It creates text (generative), learned from a large body of text before being tuned for tasks (pre-trained), and is built on the transformer neural network design introduced by Google researchers in 2017.

What is the latest GPT model?

New versions arrive often, so this page keeps the current names in the box above. OpenAI's own website lists its models and which of them run in ChatGPT and in the API for developers.

Can you download a GPT model?

Most GPT models run only on OpenAI's servers, reached through ChatGPT or the API. The exception is gpt-oss, released in August 2025 with open weights that anyone can download and run.