GPT is OpenAI's family of language models: transformers first trained on huge amounts of text, then adapted for specific tasks.
In plain words
The first GPT, in 2018, joined two ideas: the transformer design and pre-training on unlabelled text, followed by fine-tuning for specific jobs. Each later model mostly scaled that recipe up. GPT-2 learned by predicting the next word in web text, GPT-3 grew to 175 billion parameters, and GPT-4 added image input. In September 2026 OpenAI's API lists GPT-6 Astra as its most capable model.
A quick example
When you ask ChatGPT to summarise an article, a GPT model reads your request as tokens and writes the answer one token at a time, each chosen based on everything before it.
Where you'll meet it
- ChatGPT.
- OpenAI's developer API.
Related
Sources
- DocsImproving language understanding with unsupervised learning, OpenAI · read 29 Sept 2026
- DocsBetter language models and their implications, OpenAI · read 29 Sept 2026
- DocsLanguage models are few-shot learners, OpenAI · read 29 Sept 2026
- PaperGPT-4 Technical Report, OpenAI (arXiv) · read 29 Sept 2026
- DocsModels, OpenAI API docs · read 29 Sept 2026
Checked by WhatIsAI against the sources above · 29 Sept 2026. Spotted an error? Report it.