Glossary › Players and products
GPT
in English: Generative Pre-trained Transformer
GPT is OpenAI's family of language models: AI models pre-trained on vast amounts of text that write new text on request, and it also powers ChatGPT.
In plain terms
GPT is an artificial intelligence model that writes text: it answers, summarizes, translates, and creates code. The three letters stand for three features: generative (creates new text), pre-trained (learned on a huge amount of text before you received it), transformer (this is the name of its internal architecture).
Imagine someone who has read an entire library and has a strong intuition for how a sentence typically continues. They do not retrieve ready-made answers from an encyclopedia but formulate the most likely continuation word by word. This is why they write fluently and why they sometimes err with confidence.
GPT is not the same as ChatGPT. GPT is the engine; ChatGPT is the car into which it is installed. The same engine works in many other products, such as office software, customer service chatbots, and agents that perform tasks autonomously.
The name originally referred to OpenAI's models (GPT-3, GPT-4, GPT-5), but in common parlance many apply it to any similar conversational AI. Models from other companies operating on the same principle bear different names, such as Claude, Gemini, or Llama.
How it works
- Pre-training: the model practices a single task on a huge amount of text, namely guessing the next word fragment. Meanwhile, language patterns and vast factual knowledge become fixed in its internal numerical values, the parameters.
- Tokenization: when you write to it, it breaks your text into small units called tokens and converts them into numbers.
- Attention: the essence of the transformer architecture is that for every new word fragment, the model looks back at all parts of the text so far and weighs how important each is at that point. This lets it track, for example, who a "she" at the end of a sentence refers to.
- Generation: the model calculates the probability of each token following next, selects one based on these probabilities, appends it to the text, and then repeats the process. The answer is thus built piece by piece. Since there is some randomness in the selection, the same question does not always yield the same answer.
- Fine-tuning: developers later adjust the model using human feedback and rules to make it helpful, polite, and safe. Newer versions handle images and audio as well, and use tools such as web search.
An everyday example
You ask ChatGPT to write a polite complaint letter about a late-arriving package, and the text is ready in half a minute. Behind the scenes, a GPT model assembled the letter piece by piece, in word fragments (tokens). The same happens when a fluent Hungarian answer to your question appears instantly in a web shop's chat window.
Who is affected
A common misconception
Limits and risks
- It can be confidently wrong: it may present fabricated data, quotes, or references in a convincing tone; this is called hallucination.
- Its knowledge is current only up to a specific date; it knows about events after that only if the product includes web search.
- What you type goes to a company's server, so share personal, medical, or business secrets with it only after careful consideration.
- It may give a different answer to the same question the second time, and it can also adopt biases and prejudices from its training texts.
Similar, but not the same
- ChatGPT – ChatGPT is the chat application you use, while GPT is the model working inside it.
- Nagy nyelvi modell (LLM) – The large language model is the broader category; GPT is one family within it, linked to OpenAI.
- AI-agent – An agent independently solves tasks in multiple steps using tools; GPT is only the "thinking" model upon which an agent can be built.
Related concepts
GPT-4, GPT-5
A brief history
Google researchers described the transformer architecture in 2017; building on this, OpenAI introduced the first GPT model in 2018. GPT-3 appeared in 2020, GPT-4 in March 2023, and GPT-5 in August 2025. The general public became familiar with the name in November 2022, with the launch of ChatGPT.
Where things stand
By the end of 2026, GPT models were among the most widely used AI systems globally, integrated into office software, search engines, and agents. Meanwhile, they face strong competition from similar models by other companies.
In the news
- State attorneys general target AI providers: West Virginia sues SnapNov 6, 2026 · Jog
- Man sues OpenAI, alleging ChatGPT helped his wife plan to poison himNov 5, 2026 · Jog
- DeepSeek's agent platform reaches 1.54 million daily users as Chinese open models spreadNov 2, 2026 · Agent Age
- OpenAI to pool ChatGPT and Codex usage allowance from DecemberNov 2, 2026 · Code
- OpenAI's GPT-6 Cyber finds 81 Firefox and 948 Linux flawsNov 2, 2026 · Code
- Mistral's open-weight Le Chonk model ranked alongside Codex GPT-6.1 Terra in codingOct 29, 2026 · Code
- DeepSeek unveils coding "super-agent" and open-sources its architectureOct 27, 2026 · Agent Age
- OpenAI releases GPT-6.1 Terra, which it says is strong at computer useOct 26, 2026 · Agent Age
Also known as
GPT · Generative Pre-trained Transformer · generative pre-trained transformer · generative pre-trained transformer · GPT-modell · GPT model · GPT-4 · GPT-5
Sources
- Wikipedia: Generative pre-trained transformer
- Vaswani et al. (2017): Attention Is All You Need (arXiv)
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.