GPT stands for Generative Pre-trained Transformer. It is the family of AI models built by OpenAI — the technology underneath ChatGPT.
The distinction that confuses people: GPT is the engine, ChatGPT is the car. GPT is the model. ChatGPT is the app you type into. Other products use GPT models without being ChatGPT at all.
What each word means
- Generative — it produces new text rather than retrieving stored answers. Nothing it writes was looked up; all of it was generated.
- Pre-trained — it learned from an enormous amount of text before you ever used it. It is not learning from your conversation in real time, and it does not remember you between sessions unless the app stores that separately.
- Transformer — the underlying architecture, introduced in a 2017 research paper, that made all of this work. Essentially every modern AI assistant is built on it, including Claude and Gemini.
How it actually works
GPT predicts the next piece of text, over and over, very well. That is the whole mechanism. Given everything so far, what comes next?
It does not work with whole words. Text is broken into tokens — roughly word-fragments, where common words are one token and unusual ones are several. "Unbelievable" might be three. This is why models are priced per token and why they occasionally miscount letters in a word: they are not seeing letters.
That sounds too simple to produce something that can write an essay or debug code, and the fact that it does anyway is the genuinely surprising result of the last several years.
Why it makes things up
The mechanism also explains the failure mode. A model optimized to produce plausible text will produce plausible text even when it does not know the answer.
That is what a hallucination is — not the model lying, but the model doing exactly what it was built to do with insufficient information. It has no internal flag for "I don't know this"; it has a distribution over what usually comes next, and something always comes next.
Practical consequences:
- Citations and quotes are the highest-risk output. A plausible-looking reference is exactly what the mechanism produces best.
- Numbers deserve suspicion. Especially statistics with a confident source attached.
- Recent events are a weak point unless the tool has web access, because training data has a cutoff date.
Context windows
The context window is how much text the model can consider at once — your conversation, plus anything you pasted in. Measured in tokens.
When a conversation exceeds it, the earliest parts fall out. This is why a long chat sometimes seems to "forget" what you agreed at the start. It is not forgetting; that text is no longer in front of it.
Temperature, and why answers vary
Ask the same question twice and you get two different answers. That is not a bug or a sign of learning — it is deliberate.
At each step the model has a distribution over what could come next, and it samples from it rather than always taking the single most likely token. "Temperature" is the dial controlling how adventurous that sampling is. Low temperature gives repetitive, predictable text; high temperature gives variety and, past a point, nonsense.
Most consumer apps set this for you and do not expose it. It is why the same prompt produces different output, and why "regenerate" is a useful button rather than a pointless one.
What GPT models cannot do
Knowing the mechanism tells you where the limits are:
- Arithmetic on long numbers. They pattern-match rather than calculate.
- Counting characters or words. They see tokens, not letters — which is why "how many r's in strawberry" became a famous failure.
- Knowing today's date or news without web access.
- Remembering you between conversations, unless the app stores that separately and feeds it back in.
- Reporting their own confidence meaningfully. A model saying "I'm certain" is generating text, not consulting an internal gauge.
The version numbers
Higher is newer. Each generation is broadly more capable than the last, and the naming has grown complicated as OpenAI ships variants tuned for speed, cost, or reasoning depth alongside the flagship.
Smaller, faster variants often outperform older flagships while costing a fraction as much, so "newest small model" frequently beats "old big model."
You will also see it written as chat gpt, chatgbt, chat gbt, or GPT chat. Same thing.
GPT is not the only option
Because ChatGPT arrived first and became a household name, a lot of people use "GPT" to mean AI assistants generally — the way "Hoover" came to mean vacuum cleaner.
It is worth knowing it is one family among several. Anthropic's Claude, Google's Gemini, xAI's Grok, and DeepSeek's open models are all direct competitors, and each is ahead on something. They are all transformers underneath; they differ in training, tuning, and temperament.
Common questions
Is GPT the same as ChatGPT? No. GPT is the model; ChatGPT is the application built on top of it.
What does GPT stand for? Generative Pre-trained Transformer.
Does GPT learn from my conversations? Not in real time. Whether your conversations are later used for training depends on the product and its settings — usually there is a toggle, and it is usually on by default.
Why does it get things wrong? It generates plausible text rather than retrieving verified facts. When it lacks the information, plausible is all you get.
Is GPT free? There are free tiers of products built on GPT, with usage caps. API access is priced per token.