What it is
GPT is the name of the family of AI models made by OpenAI, the American company behind ChatGPT. When you type a question into ChatGPT, a GPT model is what reads it and writes the answer. The same models also power thousands of other apps, from writing tools to customer service bots, and they sit inside some Microsoft products.
The letters stand for "generative pre-trained transformer". Each word describes something real:
- Generative means it produces new material, such as text, code, or images, rather than only sorting or labelling things.
- Pre-trained means it first learned from an enormous amount of text before anyone taught it a particular job. That first stage is called pretraining.
- Transformer is the design of the neural network inside it, a way of processing text invented at Google in 2017 and explained on the Transformer page.
The name comes from a June 2018 OpenAI paper by Alec Radford and colleagues, "Improving Language Understanding by Generative Pre-Training". Its idea was simple and turned out to be enormously important: first let a model learn language in general by reading lots of text, then adjust it for specific tasks. Every GPT model since has built on that recipe, and so have most rival language models.
The generations, in brief
You don't need to memorise these, but knowing the rough outline makes the news easier to follow. The dates are from OpenAI's own announcements.
- GPT-1 (June 2018): a research model that proved the idea.
- GPT-2 (February 2019): famous for being held back at first over fears of misuse. The story is on the Language Model page.
- GPT-3 (May 2020): much larger, and the first GPT that businesses could use through OpenAI's API.
- GPT-3.5 and ChatGPT (November 2022): a model trained to hold a conversation, released to the public as ChatGPT. See Artificial Intelligence for that launch.
- GPT-4 (14 March 2023): a large step up that could also accept images as input.
- GPT-4o (13 May 2024): the "o" stands for "omni". It handled text, pictures, and speech in one model, and OpenAI used it to make ChatGPT's voice conversations faster and more natural.
- The "o" reasoning models (from September 2024): o1 and its successors, which work through a problem before answering. The Inference page covers o1 and o3.
- GPT-5 (7 August 2025): OpenAI combined its chat models and reasoning models into one system that decides for itself how hard to think.
- GPT-5.1 to GPT-5.6 (late 2025 to mid 2026): a rapid series of point releases, each replacing the last in ChatGPT within weeks or months.
- GPT-6 Astra (3 September 2026): which OpenAI called "the world's most intelligent and aligned model". On 22 September 2026 OpenAI added two cheaper, faster members, GPT-6 Sol and GPT-6 Luna. The debate over whether Astra counts as AGI is covered on that page.
Most GPT models are closed. OpenAI keeps the model files on its own servers and you use them through its apps or its API. The main exception is gpt-oss, two "open-weight" models released on 5 August 2025, which anyone can download and run on their own computer. OpenAI had not released a language model that way since GPT-2.
ChatGPT, the API, and the models are three different things
People often use "GPT" and "ChatGPT" as if they were the same, but it helps to keep them apart.
The models are the trained AI systems themselves, with names like GPT-5.6 or GPT-6 Astra. On their own they are just very large files of numbers.
ChatGPT is the app: the website, the phone app, and the desktop program. It wraps the models in a chat window and adds extras such as memory, web search, file uploads, image making, and voice. It has a free plan and several paid plans, and which models you get depends on your plan.
The API is the doorway for developers, who pay by usage to build GPT models into their own products. It often offers models that ChatGPT doesn't show, and it keeps older models running for longer. For example, when OpenAI retired GPT-4o from ChatGPT in February 2026, it said the API was unchanged.
Chat models and reasoning models
Inside the GPT family there are two broad kinds of model, and ChatGPT lets you choose between them.
A chat model (OpenAI's current label is "Instant") answers straight away. It is quick and good for everyday questions, drafting, and conversation.
A reasoning model (labelled "Thinking" in ChatGPT) spends time working through the problem before it replies, a bit like doing rough working on scrap paper first. This is sometimes called extended thinking, and it builds on the idea behind chain of thought. It is slower and uses more computing power, but it tends to do better at maths, logic, careful analysis, and multi-step tasks.
At the time of writing, ChatGPT's menu offers Instant for fast replies, several thinking levels for harder questions, and "Pro" options on the most expensive plans. Free users get a lighter model and a "Think" button instead of the full set. These labels have changed several times in the past year alone, so expect them to look different when you read this.
Why it matters
It is most people's first AI. OpenAI said in February 2026 that ChatGPT had passed 900 million weekly users, which is the company's own figure. For many people, "AI" simply means ChatGPT, and ChatGPT means GPT. Understanding what sits behind the chat window helps you use it better and judge it more fairly.
The choice of model changes the answer. The same question can get a quick, confident, and wrong reply from a fast model, and a slower, correct one from a reasoning model. Or the reverse: a reasoning model can over-complicate a simple request. Knowing that there is a choice, and when to make it, is one of the easiest ways to get better results from any chatbot.
Things change constantly. Model names, menus, plans, prices, and usage limits have changed many times since 2022. OpenAI releases, updates, and retires models so often that it keeps a public log of the changes (linked under Additional resources below). A prompt that worked well last spring may behave differently today, because a different model is answering it. If you rely on ChatGPT for work, that is worth knowing.
Who controls it matters. A tool used by hundreds of millions of people is shaped by the people who run the company. OpenAI has an unusual structure, with a nonprofit at the top, and it has gone through a public leadership crisis and a major reorganisation. Those events decide who answers for how the models behave, how they are tested for safety, and who benefits from the money they earn. You don't need to take a side to find that worth understanding.
Real-world examples
The weekend OpenAI fired its chief executive (November 2023)
OpenAI started in 2015 as a nonprofit, a charity-style organisation, with the stated mission of making sure that powerful AI benefits all of humanity. In 2019 it created a business arm to raise money, but the nonprofit's board of directors stayed in charge of it.
On 17 November 2023, that board announced that it had removed Sam Altman as chief executive. Its statement said it had concluded that he "was not consistently candid in his communications with the board", and named Mira Murati, the company's chief technology officer, as interim chief executive. Greg Brockman, the company's president, quit the same day.
What followed played out in public over five days. Microsoft, OpenAI's biggest investor, offered Altman and Brockman jobs. The board appointed a second interim chief executive, Emmett Shear, the former head of Twitch. On 20 November, CNBC and others reported that more than 700 of the company's roughly 770 staff had signed a letter threatening to leave for Microsoft unless the board resigned and Altman returned. Within days an agreement was reached. On 29 November 2023 OpenAI confirmed that Altman was back as chief executive with a new initial board of three: Bret Taylor as chair, Larry Summers, and Adam D'Angelo, the only member who stayed from the old board.
The new board hired the law firm WilmerHale to review what happened. On 8 March 2024 OpenAI announced the review was finished and that it had found Altman's conduct "did not mandate removal". The report itself was not published. Altman rejoined the board, and three new directors were added.
For ChatGPT users, very little changed on screen that week. But the episode showed that a small board could remove the leader of one of the world's most important technology companies overnight, and that staff and investors could reverse it just as fast. Much of what came next in OpenAI's governance can be read as a response to it.
OpenAI becomes a public benefit corporation (October 2025)
The second big change to how GPT is governed came two years later. On 28 October 2025, OpenAI announced it had completed what it called a recapitalisation, a reorganisation of who owns what.
The nonprofit was renamed the OpenAI Foundation. The business that builds the GPT models became OpenAI Group PBC, a public benefit corporation. That is a kind of company that, unlike an ordinary one, is legally required to pursue a stated mission and consider the interests of the public, not only its shareholders. According to OpenAI, the Foundation still controls the company, appoints all of its directors, and owns a 26 percent stake that OpenAI valued at about $130 billion. Microsoft said it now held roughly 27 percent, valued at about $135 billion. The Foundation committed $25 billion to health research and to what OpenAI calls "AI resilience". OpenAI said the change was completed after nearly a year of discussions with the Attorneys General of California and Delaware, who oversee charities registered in those states and did not object to the final deal.
Not everyone was reassured. Critics quoted by the California news site CalMatters, including community groups and Steven Adler, who once led a safety team at OpenAI, argued that the nonprofit would still be dominated by the business it is supposed to oversee. They pointed to the same people sitting on both boards, and questioned whether a safety committee could stay independent of the pull of profits. OpenAI's position is that the new structure keeps the mission in charge while letting it raise the money it needs. Both are claims about the future, and time will test them.
The model menu that kept coming back (2025 to 2026)
In February 2025, Sam Altman posted a plan for simplifying ChatGPT. He wrote that OpenAI hated "the model picker as much as you do", referring to the long menu of oddly named models, such as GPT-4o, o3, o4-mini, and GPT-4.5, that users had to choose between. The plan was for GPT-5 to decide for itself which kind of model to use.
GPT-5 arrived on 7 August 2025, and the menu disappeared. Within a day, many paying users were complaining. Some found the new model colder than GPT-4o, which they had grown attached to. Others found it weaker, and on 8 August Altman explained that the automatic switcher that chose between the fast and thinking modes had broken for part of launch day, which had made GPT-5 seem "way dumber". He announced that Plus subscribers could choose GPT-4o again, and within days a menu with Auto, Fast, and Thinking options was back.
The story didn't end there. On 29 January 2026 OpenAI announced it would finally retire GPT-4o from ChatGPT on 13 February 2026, saying that only 0.1 percent of users still chose it each day. By the time of writing, the menu offered Instant, several thinking levels, and Pro, and OpenAI's help pages said that on the Plus and Pro plans, Instant no longer switched to a thinking level automatically for complex requests. The lesson for users is practical: the menu is not decoration. What you pick affects what you get, and the company itself has changed its mind about how much of that choice to hand over to you.
An everyday case: the bookkeeper and the tax figures
Picture Priya, who runs a small bookkeeping business. She uses ChatGPT's quick Instant mode all day for emails, meeting notes, and rewording letters to clients, and it's ideal for that. One afternoon she asks it to work out a client's quarterly tax figures from a pasted table. The answer comes back in seconds and looks tidy, but when she checks one line by hand, the total is wrong. She asks again with a thinking level switched on. This time the reply takes forty seconds, shows the steps it took, and the numbers match her own. She doesn't conclude that the fast mode is bad. She concludes that it's the wrong tool for arithmetic that matters, and she now checks any figures herself whichever mode she uses.
Try it yourself
This exercise shows you the difference between a fast model and a reasoning model using ChatGPT's own menu. It works on the free plan. Menus and names change often, so if the labels differ from the ones described here, look for the closest match. At the time of writing, free users see a "Think" option in the message box rather than the full list of thinking levels. Check OpenAI's current plan pages for what yours includes.
- Open ChatGPT at chatgpt.com, or in the app, and sign in. A free account is enough.
- Find the model choice. On paid plans it is a menu near the top of the chat or beside the message box, with options such as Instant and Thinking. On the free plan, look for a Think button or option near where you type.
- Start a new chat in the fast mode (Instant, or with Think switched off) and paste this puzzle: > A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost? Then: I have a 3-litre jug and a 5-litre jug and a tap. How do I measure exactly 4 litres? Explain each step briefly.
- Note what you get. How long did it take? Is the answer right? (The ball costs 5 cents, not 10 cents.) Is the jug method correct if you follow it step by step?
- Open a second new chat, switch on a thinking level (or Think), and paste exactly the same prompt. Compare the time taken, the answer, and how clearly it explains itself.
- Now try an everyday task in both modes. Paste this into a fresh chat in each mode: > Write a friendly two-sentence reminder to a client that their invoice is 14 days overdue. Keep it polite but clear. You will probably find the fast mode is just as good here, and much quicker. That's the point: harder thinking helps with problems, not with simple writing.
- Ask a question where being out of date matters. Paste this in either mode: > Which ChatGPT plans exist right now, and what does each include? Search the web and give me the sources you used. Then open at least one source it gives you and check. Plans and limits change so often that any answer, including this article, can be out of date.
- Write down your own rule. Based on what you saw, finish this sentence in your notes: "I'll use the fast mode for ___ and a thinking mode for ___." Revisit it after the next big model update.
Common mistakes and misunderstandings
Thinking GPT and ChatGPT are the same thing. GPT is the family of models; ChatGPT is one app that uses them. Other products use GPT models too, and ChatGPT itself switches between several of them.
Assuming everyone gets the same model. At the time of writing, free users get a lighter model than paid users, and the most powerful options are limited to the most expensive plans, with weekly message limits even there. If a friend gets a better answer to the same question, they may simply be using a different model.
Believing the newest or slowest option is always best. Reasoning modes are better at hard problems but slower, and they can overthink simple jobs. Match the mode to the task.
Treating a company's claims as independent results. When OpenAI calls a model the most intelligent in the world, that is OpenAI's claim, based on tests it chose. Rivals make similar claims about their own models. Look for results from independent groups, and above all test models on your own tasks. The Benchmarks and Evals pages explain how.
Trusting the model to know its own details. A model may be unsure of its exact version, its limits, or recent events, and it can hallucinate about itself just as it can about anything else. Check the label in the app's menu and OpenAI's help pages instead.
Expecting a model to stay the same. OpenAI updates models in place, changes their default style, and retires them from ChatGPT, sometimes with a few weeks' notice. The GPT-4o saga showed that people can grow attached to one model's personality and then lose it. If a workflow matters to you, keep your best prompts written down so you can adapt them quickly to the next model.
Mixing up GPT with "GPTs". OpenAI also used the name GPTs for custom chatbots that users could build, launched in November 2023. In September 2026 OpenAI said it plans to retire them. They were built on GPT models, but they were a feature of ChatGPT, not a model family.
Assuming GPT is open source. Almost all GPT models are closed: you can use them but not download them. The gpt-oss models from August 2025 are the exception. If you need a model you can run yourself, look at open-weight families such as Llama, Qwen, or Mistral.
Related elements
- Claude: Anthropic's model family, a direct alternative to GPT with its own chat app and API.
- Gemini: Google's model family, which competes with GPT in chat, search, and phones.
- Language Model: the kind of AI every GPT model is, with the stories of GPT-2 and GPT-3.
- Transformer: the "T" in GPT, the network design all these models share.
- Pretraining: the "P" in GPT, the first stage where a model learns from vast amounts of text.
- Extended Thinking: what happens when you pick a thinking mode instead of a fast one.
- Inference: why reasoning models cost more to run, with the story of o1 and o3.
- API: how developers build GPT models into their own products.
- Alignment: includes the April 2025 update that made GPT-4o far too agreeable.
- AGI: the goal in OpenAI's mission, and the argument over whether GPT-6 Astra reaches it.
- OpenAI Images: OpenAI's image-making models, which sit alongside GPT in ChatGPT.
- Sora: OpenAI's video model.