GPT stands for Generative Pre-trained Transformer. It is the name of OpenAI’s family of large language models, which power ChatGPT and many third-party applications. Each word in the name describes the approach: the models generate text, are pre-trained on vast amounts of data, and are built on the transformer architecture.
Key takeaways
- GPT is a family of LLMs developed by OpenAI, starting in 2018.
- GPT models are decoder-only transformers that predict the next token.
- ChatGPT is an application; GPT models are the engines inside it.
- “GPT” is also used loosely to mean any chatbot or LLM, though other companies’ models are not GPTs.
- Later GPT models became multimodal, working with images and audio as well as text.
What each letter means
| Letter | Word | Meaning |
|---|---|---|
| G | Generative | Creates new text, one token at a time |
| P | Pre-trained | First trained on a huge general dataset before being adapted to tasks (pretraining) |
| T | Transformer | Uses the transformer neural network architecture introduced in 2017 |
GPT explained in simple terms
Think of GPT as an engine and ChatGPT as a car built around it. The engine is a very large model that has learned to continue text by studying enormous amounts of writing. The car adds a steering wheel and dashboard: a chat interface, safety rules, memory and tools. Other companies build their own cars on GPT engines through OpenAI’s API, while other labs build entirely different engines.
A short history
| Model | Year | Milestone |
|---|---|---|
| GPT-1 | 2018 | Showed that generative pre-training plus fine-tuning worked well across language tasks |
| GPT-2 | 2019 | 1.5 billion parameters; initially released in stages over misuse concerns |
| GPT-3 | 2020 | 175 billion parameters; showed strong few-shot learning from prompts |
| InstructGPT / GPT-3.5 | 2022 | Trained with RLHF to follow instructions; powered ChatGPT’s launch in November 2022 |
| GPT-4 | 2023 | Major gains in reasoning and exams; accepted image inputs |
| GPT-4o | 2024 | “Omni” model handling text, images and audio in one model |
OpenAI has continued releasing newer GPT models and separate reasoning models. Check OpenAI’s official site for the current line-up, as it changes frequently.
How GPT generates text
- The prompt is split into tokens.
- The model processes all tokens with stacked transformer layers using attention.
- It predicts the probability of every possible next token.
- A token is selected, influenced by temperature, and appended.
- The process repeats until the response is complete.
GPT vs ChatGPT vs other LLMs
| Term | What it is |
|---|---|
| GPT | OpenAI’s family of language models |
| ChatGPT | OpenAI’s consumer chat application built on GPT and related models |
| Other LLMs | Models from other labs, such as Claude (Anthropic), Gemini (Google), Llama (Meta), Mistral, Qwen and DeepSeek |
| “Custom GPTs” | Configured versions of ChatGPT with instructions and files, not separately trained models |
How businesses use GPT
Companies access GPT models through OpenAI’s API or through Microsoft’s Azure OpenAI service to build chatbots, writing tools, coding assistants, data extraction pipelines and agents. They often combine GPT with their own documents using retrieval, and use system prompts and structured output to control behaviour.
Limitations
- Can hallucinate facts.
- Has a knowledge cut-off unless connected to search or documents.
- Closed weights: you cannot download or fully inspect the models.
- API costs scale with usage.
- Behaviour can change between model versions.