What Does GPT Stand For? Meaning, Origin and Timeline

In brief

GPT stands for Generative Pre-trained Transformer. OpenAI described the approach in a paper dated 11 June 2018, and ChatGPT adds Chat because OpenAI tuned the model for conversation when it launched on 30 November 2022. GPT-3 had 175 billion parameters, and OpenAI launched GPT-6 Astra on 3 September 2026.

What does GPT stand for in ChatGPT?

GPT stands for Generative Pre-trained Transformer. Chat isn't part of the acronym. It names the product. Each word of the acronym answers one question about the model.

Three glass blocks stacked on a black plinth beside an open notebook and a brushed silver OpenAI knot, in a charcoal studio under white rim light.
WordPlain meaningWhere it comes from
GenerativeThe model writes new text instead of picking from a listIt predicts the next word, then the next
Pre-trainedIt learned from a huge pile of text before you ever used itOpenAI's June 2018 paper on generative pre-training
TransformerThe network design that lets every word look at every other wordGoogle researchers, June 2017

The rest of this page takes each word in turn, dates the versions, and covers the other things GPT can mean.

What does each word in GPT mean?

Generative means the model produces text, and it does so one piece at a time. It reads your prompt, predicts the most likely next token, adds it, and repeats. That's why the same question can get two different answers. What are tokens in AI shows what a token is.

Pre-trained means the heavy learning happens before you type anything. OpenAI's page on its first model says the pre-training ran for about a month on 8 GPUs, over a few thousand books that came to roughly 5GB of text. A small tuning step on labeled examples came after. The page puts it this way: "This provides some insight into why generative pre-training can improve performance on downstream tasks."

Transformer is the design inside the model. A transformer lets each word weigh its relevance to every other word in the same step, which is what let models grow from millions of parameters to billions. The transformer explainer covers attention in depth. Here the one fact that matters is the order of events. The transformer came first, in 2017. OpenAI built GPT on top of it.

Which OpenAI paper named GPT?

OpenAI published the first GPT on 11 June 2018, under the title "Improving language understanding with unsupervised learning". The paper behind it is titled Improving Language Understanding by Generative Pre-Training, and Wikipedia lists the same title and date for it. Alec Radford is the author named on the OpenAI page.

The first model was small and uneven. OpenAI reports a 12 block transformer that read sequences of up to 512 tokens. It set new results on tests of commonsense reasoning and reading comprehension, such as COPA at 78.6 against a prior 71.2. It also lost on others: 56.0 against 61.7 on RTE and 91.3 against 93.2 on SST-2.

So the first GPT wasn't the best model at everything. Its point was a method. Pre-train on plain text, then tune. It worked. By February 2019 OpenAI called its next model "a successor to GPT", which is the earliest use of the short name I found on OpenAI's own pages.

What does ChatGPT stand for?

ChatGPT stands for Chat plus Generative Pre-trained Transformer, and the Chat marks that the model was tuned for conversation. OpenAI's launch post of 30 November 2022 says: "We've trained a model called ChatGPT which interacts in a conversational way." It adds that the model was adapted from the GPT-3.5 family and trained with reinforcement learning from human feedback.

That's the difference between GPT and ChatGPT. GPT is the family of models. ChatGPT is the app that wraps a chosen model in a chat window, with its own rules, memory and plans. Chat was a product label from the launch post onward.

Newer ChatGPT versions keep the pattern. Model names now carry a number and sometimes a tier, such as Astra, Sol or Luna in the GPT-6 family. ChatGPT model history lists every one with its dates, and which ChatGPT model should I use helps you pick.

Which GPT came when? A dated timeline

Eight years from the transformer to GPT-6
The ChatGPT era201820192020202120222023202420252026GoogleTransformer paperOpenAIGPT-1GPT-2GPT-3ChatGPTGPT-4GPT-4oGPT-5GPT-6 Astra

The table gives the date and one fact per release, taken from OpenAI's own pages where I could open them.

ReleaseDateOne factSource
GPT-111 June 201812 block transformer, 512 token sequences, about 5GB of booksOpenAI
GPT-214 February 20191.5 billion parameters, about 8 million web pages (40GB)OpenAI
GPT-328 May 2020175 billion parametersarXiv paper
ChatGPT30 November 2022Adapted from the GPT-3.5 family with human feedbackOpenAI
GPT-414 March 2023Accepts image and text inputs, outputs textOpenAI
GPT-4o13 May 2024Takes text, audio, image and video as input; text and image input launched firstOpenAI
GPT-57 August 2025New default model in ChatGPT for signed in usersOpenAI
GPT-6 Astra3 September 2026First of the GPT-6 family, limited to a set of organizations at launchOpenAI release notes

OpenAI does not state a parameter count for GPT-4 on the page I read, so any figure you see for it is a third party report. Model counts and release dates also move fast: as of November 2026 the GPT-6 page holds the current lineup.

Which GPT is ChatGPT running now?

As of November 2026 the newest family is GPT-6, and OpenAI's release notes date Astra to September 3, 2026, with access limited to a set of organizations at first. Which model answers you depends on your plan and on the model picker, so the app's name no longer tells you the version.

Plans, prices and message limits have their own pages:ChatGPT pricing holds the plan table, and ChatGPT vs Claude vs Gemini compares the three assistants. Neither is repeated here.

Three real cases of the name at work

GPT-1 in 2018: a language model trained on books that then beat the best published results on reading comprehension tests, while losing on others. The lesson: pre-training helped most where a task needed broad knowledge of language, and it didn't help everywhere.

GPT-2 in 2019: the same recipe with 1.5 billion parameters, trained to predict the next word in 40GB of internet text. OpenAI released it in stages. A 345 million parameter version came out in May 2019, and the full model followed in November, as Wikipedia reports.

ChatGPT in 2022: the recipe plus human feedback. OpenAI says ChatGPT is a sibling of InstructGPT, a model built to follow instructions. Same transformer family, different final training step. That final step is what Chat stands for in practice.

What does GPT stand for in computing?

In computing GPT also means GUID Partition Table, a standard for how a disk records its partitions. It doesn't involve AI at all. Wikipedia describes it as part of the UEFI standard, and says Intel developed it in the late 1990s. GUID is short for Globally Unique Identifier.

The disk meaning shows up when a Windows or Linux guide tells you to convert a drive from MBR to GPT. With 512 byte sectors, MBR caps a disk at 2 TiB, while GPT allows up to 8 ZiB, per the same Wikipedia page.

MeaningFieldWhat it is
Generative Pre-trained TransformerAIOpenAI's family of language models
GUID Partition TableComputingThe disk partition layout in UEFI
General-purpose technologyEconomicsListed on Wikipedia's GPT page
GPT GroupBusinessA company, also listed there
Alanine transaminaseMedicineA liver enzyme test, also listed there

Wikipedia's disambiguation page lists more meanings, from a song to an airport code. When you see GPT in a headline about chatbots, it means the AI one.

Can anyone use the name GPT?

Anyone can say it. OpenAI has tried to own it, though. According to Wikipedia, OpenAI said in 2023 that GPT should be treated as its brand and changed its API terms in April 2023 so customers could no longer use GPT in product names.

The US Patent and Trademark Office rejected OpenAI's application in May 2023, finding GPT both descriptive and generic, on that same Wikipedia account. The EU registration is under challenge, and China denied OpenAI's applications. Treat these as Wikipedia's report and check the registries before relying on them.

What can you ask an AI about GPT?

These five prompts work in Claude, ChatGPT, Gemini or Whizi. Fill the brackets.

  1. Explain the name: "Explain generative, pre-trained and transformer to a [age] year old, with one everyday example for each word. Then ask me one question to check I understood."
  1. Split training from design: "In the name GPT, which word describes how the model is trained and which describes how it is built? Answer in one sentence each."
  1. Separate model from product: "What is the difference between GPT and ChatGPT? Give one line each for the model, the app and the plan, as of [month and year]."
  1. Source a date: "Search the web for OpenAI's own page announcing [GPT version]. Give the release date, the link, and say which facts on that page you relied on."
  1. Test a rumor: "Does OpenAI itself state the parameter count of [version]? If only a third party does, label the number as a rumor and name the third party."
  1. Map other meanings: "List what GPT can abbreviate in [field, such as medicine or computing], with one example sentence each."

Where does Whizi fit?

Whizi is one app for asking Claude, GPT and Gemini the same question instead of opening a pile of single purpose tools. Web search is on by default in each chat, so a date question can check OpenAI's own pages.

The free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card. Pro carries Claude, GPT and Gemini in one app, its first week is $1.99, and Compare shows two models' answers side by side. See Whizi pricing for what each plan holds.

What does an acronym explanation leave out?

The name says how a model is built. It doesn't say whether its answers are true. A generative model predicts likely text, and a likely sentence is not always a correct one. Does ChatGPT make up sources shows how that fails in practice, and how to fact check AI answers gives a two minute routine.

Pages that explain GPT also trip over the same errors. One top three result calls GPT "the application of ChatGPT", which is backwards: GPT is the model and ChatGPT is the product. A version number or a parameter count from a blog is a report, not a fact, until OpenAI states it. Don't quote it as one. What is an LLM places GPT among the other families.

Workflow checklist
  • Read GPT as three words: generative, pre-trained, transformer
  • Remember Chat is the product, not part of the acronym
  • Take release dates from OpenAI's own pages
  • Treat parameter counts that OpenAI has not stated as rumors
  • Check the field before reading GPT as AI: it can also mean a disk partition table
Common questions

Frequently asked questions

What does GPT stand for in ChatGPT?

Generative Pre-trained Transformer. Generative means the model writes new text, pre-trained means it learned from a large body of text first, and transformer is the network design. OpenAI described the approach in a paper dated 11 June 2018.

What does ChatGPT stand for?

Chat Generative Pre-trained Transformer, with Chat marking the conversational tuning. OpenAI's launch post of 30 November 2022 describes ChatGPT as a model that interacts in a conversational way, adapted from the GPT-3.5 family.

Is GPT the same as ChatGPT?

No. GPT is the family of models, and ChatGPT is the app that runs a chosen model in a chat window. ChatGPT model history lists which model sat behind the app on each date.

What does GPT stand for in computers?

GUID Partition Table, the layout a disk uses to record its partitions. Wikipedia says it is part of the UEFI standard and allows much larger disks than MBR, which caps at 2 TiB with 512 byte sectors.