GPT stands for Generative Pre-trained Transformer. It's a type of neural network that writes text by guessing, over and over, which small piece should come next. The term appeared in 2018 together with OpenAI's first model, and today the same principle powers ChatGPT, GPT-6 Astra and dozens of models from other companies.
The tricky part is that the letters GPT are now stuck on everything: ChatGPT, GPT-5.6, BloombergGPT, GPT Image. One acronym hides a model, an app, an image generator and products that have nothing to do with OpenAI. Until that's sorted out, it's hard to even know what to choose.
What GPT stands for, word by word
It sounds like the name of a robot from an eighties movie, but every word here carries meaning.
Generative. The model doesn't pull an answer out of a ready-made database — it creates the text from scratch, piece by piece. That's why the same question can get two different wordings. Not a bug, just how generation works.
Pre-trained. Before the first user question, the model has already read a huge body of text: books, articles, code, websites. After that it knows a lot about language and the world, but it doesn't yet have the habit of answering politely — that's taught separately.
Transformer. A neural network architecture that Google researchers described in 2017 in the paper Attention Is All You Need. Its key mechanism is attention: while processing each word, the model looks at the whole text at once and decides which words matter most for it. In "The trophy doesn't fit into the suitcase because it is too big", grammar alone doesn't tell you what "it" is, but attention links it to the trophy. More on the internals in the transformer guide.
GPT and ChatGPT are not the same thing
GPT is the model: text goes in, a continuation comes out. ChatGPT is OpenAI's app built around the model: the chat window, conversation history, file uploads, web search, voice mode. If you want an analogy, GPT is the engine and ChatGPT is one of the cars it was installed in. Over time ChatGPT has run on different "engines": GPT-3.5 at launch in November 2022, then GPT-4, GPT-4o, GPT-5 and later versions.
The practical takeaway: GPT models are available without ChatGPT too — through the API and services that connect to them. That's how GPTunneL works: the same OpenAI models, in its own interface and next to models from other companies.
The reverse trap is GPT Image. This one really is OpenAI, but it's an image generator that grew out of the multimodal GPT-4o. So the letters GPT in a name don't guarantee the model writes text.
Where the term GPT came from: a short history
The idea of "first train a model on a mountain of unlabeled text, then fine-tune it for a task" didn't originate at OpenAI — machine learning had used this trick long before GPT. By 2018, two pieces of the future technology came together.
2017 — the transformer. Google publishes an architecture that processes text as a whole and in parallel, which means it scales well to huge amounts of data.
Early 2018 — pre-trained language models. Two notable papers come out: ULMFiT in January and ELMo in February. Both show that a model pre-trained on large amounts of text handles specific tasks better. But they're built on recurrent networks, not on the transformer.
June 11, 2018 — the first GPT. OpenAI publishes Improving Language Understanding by Generative Pre-Training and joins both pieces: generative pre-training plus the transformer. The model has 12 layers and 117 million parameters; today it's called GPT-1. Tiny by current standards: two years later GPT-3 had roughly fifteen hundred times more parameters. Fun detail — the paper's title doesn't contain the acronym at all, only "generative pre-training".
October 2018 — the acronym goes mainstream. Google researchers, in their paper on BERT, compare their model with the "Generative Pre-trained Transformer (OpenAI GPT)". From then on, GPT is the familiar name for the whole family, and every new generation only cemented the term.
Can other neural networks be called GPT — Gemini, DeepSeek, Claude?
Short answer: by design — yes, by name — no.
By design. If you read GPT as a description of the technology — generative, pre-trained, transformer-based — it fits almost any modern language model. Gemini, according to Google's technical report, is built on top of Transformer decoders — the same scheme GPT uses. DeepSeek-V3 is also a transformer, just with a "mixture of experts": out of 671 billion parameters, 37 billion work on each token. Claude, Qwen and other large language models write text the same way — by predicting the next token.
By name. GPT is OpenAI's brand name for its models. Google calls its model Gemini, DeepSeek calls its model DeepSeek. Saying "Gemini is a GPT" is a bit like calling every tissue a Kleenex: people get it, but it's not accurate. The correct umbrella term for all of them is large language model, or LLM.
Legally, it's contested. In April 2023 OpenAI updated its brand guidelines and banned companies building on its API from using GPT in their product names. But OpenAI couldn't register GPT as a trademark in the US: the office rejected the application in 2023, and on February 6, 2024 the refusal became final — the term was deemed a generic name for a family of neural networks used across the industry. In the EU, OpenAI got the registration in 2023, but it has been challenged since 2024, and in China the registration was denied.
That's why other companies have models with GPT in the name too: GPT-J and GPT-NeoX from the research group EleutherAI (2021 and 2022), BloombergGPT from Bloomberg (2023). They have no connection to OpenAI — the letters just point to the type of model.
And in everyday speech, "GPT" has long become a word for any chatbot: "ask GPT" might well mean Gemini. That's fine in casual conversation, but when you're choosing a model or comparing prices, it's better to call models by their names — the differences are in the details.
How GPT differs from other kinds of neural networks
"Neural network" is an umbrella covering very different systems, and language models are just one branch. The differences are easiest to see through two questions: what the system outputs, and how it gets there.
| Type of neural network | What it does | How it produces the result | Examples |
|---|---|---|---|
| GPT and other large language models | Write text and code, answer questions | Predict the next piece of text, one at a time | GPT-6 Astra, Claude, Gemini, DeepSeek |
| Diffusion models | Generate images and video | Start from random noise and step by step "develop" the whole image | Stable Diffusion, Flux, Midjourney |
| Encoders like BERT | Understand text but don't write it | Read the whole phrase in both directions and turn it into a numeric representation | Search, review sorting, spam filters |
| Recognition networks | Tell what's in a photo or audio | Match the input against learned features and output a label | Face unlock, license plate recognition |
| Classic machine learning | Predicts a number or class from a table | Finds patterns in structured data | Credit scoring, product recommendations |
The table reveals three things that make GPT what it is.
It generates rather than classifies. A recognition network will say "there's a cat in the photo". GPT will write a poem about that cat — and might cheerfully get the breed wrong along the way.
It writes strictly in order. Each next piece depends on all the previous ones. BERT sees the whole phrase but can't continue it; a diffusion model edits the entire image at every step.
One model for hundreds of tasks. Translation, code, an email to a client and a contract review all boil down to one thing for GPT — "continue the text". There's no need to train a separate network for each task, which is exactly why language models spread across every service so fast.
How GPT writes an answer: guessing the next piece
First, the text is split into tokens — fragments of words. An English word often fits into a single token, while words in many other languages get split into several, so the same meaning can cost a different number of tokens depending on the language.
Next, the model takes the whole conversation — your question plus the chat history — and calculates a probability for every possible next token. It picks one, appends it, and repeats the whole process. The answer appears on screen in chunks not for show: it's genuinely born one token at a time.
After "The capital of France is", the option "Paris" has a near-100% probability, while "Lyon" gets crumbs. But for "Come up with a name for a coffee shop" there are many fitting continuations with similar odds, and the model, choosing with a bit of randomness, gives something new every time.
An important consequence: GPT picks a plausible continuation rather than checking facts. When knowledge runs out, the most plausible continuation is a confident-sounding invention. That's where AI hallucinations come from.
Modern versions like GPT-5.6 and GPT-6 Astra can "think" first: generate a hidden draft of reasoning and only then the answer. The mechanism is the same — token prediction — some tokens are just spent on reasoning.
How GPT learns: first it reads, then it learns to help
Pre-training. The model is shown a gigantic body of text — for GPT-3 it was about 300 billion tokens — and asked again and again to guess the next one. After each mistake, the parameters get nudged slightly — the numbers inside the network that store learned patterns. GPT-3 had 175 billion of them. The result is a model that knows language and plenty of facts but behaves like autocomplete: asked "How do I make pancakes?", it might well continue with "How do I make waffles? How do I bake a cake?", as if scrolling through a Q&A forum.
Instruction tuning. People write examples of good answers to requests, and the model learns to answer instead of continuing a list.
Learning from human feedback. The model produces several options, people pick the best one, and it adjusts to those preferences. OpenAI described the approach in its InstructGPT paper in 2022: people preferred answers from the fine-tuned 1.3-billion-parameter model over those of the original GPT-3, which is more than a hundred times larger. This stage is what turned GPT into a conversation partner — ChatGPT launched a few months later.
Training has an end date. The model doesn't know anything that happened after it — unless the app connects it to search.
GPT generations: from GPT-1 to GPT-6 Astra
| Year | Model | What changed |
|---|---|---|
| 2018 | GPT-1 | 117M parameters. Proved that pre-training on text plus task-specific fine-tuning works |
| 2019 | GPT-2 | 1.5B parameters, coherent text several paragraphs long. OpenAI initially held back the full version, fearing misuse |
| 2020 | GPT-3 | 175B parameters. Learned to solve tasks from a few examples right in the prompt |
| 2022 | ChatGPT on GPT-3.5 | Fine-tuning on dialogues and human ratings — the model became a conversation partner |
| 2023 | GPT-4 | Understands images as input, noticeably more accurate on complex tasks. OpenAI stopped disclosing model size |
| 2024 | GPT-4o, o1 | "o" stands for omni: text, audio and images in one model. o1 is OpenAI's first reasoning model |
| 2025 | GPT-5, gpt-oss | GPT-5 decides on its own whether to answer fast or think longer. gpt-oss are OpenAI's first open-weight models since GPT-2 |
| 2026 | GPT-5.6, GPT-6 Astra | The Sol, Terra and Luna lineup for different budgets; GPT-6 Astra launched on September 3 — a flagship for long agentic tasks |
The shift in focus is visible right in the table: at first, generations competed on parameter count; then OpenAI stopped disclosing it and started talking about reasoning, tools and the cost of a solved task.
What GPT can't do
Tell truth from plausibility. A confident tone is no sign of being right: GPT will just as cheerfully recite the multiplication table and the contents of a law article that doesn't exist. Numbers, quotes and links from an answer are worth checking.
Know recent events. Everything after the training cutoff doesn't exist for the model until it's given search.
Remember you between chats. The model itself sees only the current context — for recent GPT models that's about a million tokens, several thick books at once. Memory of past conversations is a feature of the app, not the model.
Which GPT model to choose now
You don't always need the flagship. A mid-tier model covers most everyday tasks; keep the most expensive ones for tasks where a cheaper one clearly falls short.
| Model | Best for | Context | Input / output, $ per 1,000 tokens |
|---|---|---|---|
| GPT-6 Astra | Long complex tasks: agents, code, science, huge documents | 1.05M | 0.02 / 0.10 |
| GPT-5.6 Sol | Complex analysis and code when Astra is overkill | 1.05M | 0.01 / 0.06 |
| GPT-5.6 Terra | Everyday work: texts, emails, reviewing materials | 1.05M | 0.004 / 0.024 |
| GPT-5.6 Luna | High-volume simple requests and automation | 1.05M | 0.0004 / 0.0024 |
| GPT OSS 120B | Open model: cheap and can be self-hosted | 131K | 0.0003 / 0.0015 |
Prices are approximate; current ones are on the pricing page.
Astra's price per token is higher than Sol's, but according to OpenAI it spends fewer tokens per answer on complex tasks — so it's fairer to compare models by the cost of a solved task rather than by the price list. A convenient way to choose is to start in the middle: begin with Terra and, if the answer feels shallow, send the same request to Sol or Astra.
FAQ
What does GPT stand for?
Generative Pre-trained Transformer. The model generates text, is trained in advance on large amounts of data, and is built on the transformer architecture.
When did the term GPT appear?
In 2018. On June 11, OpenAI published the paper on the first model in the family, and by autumn of the same year the acronym GPT was already being used in research papers from other companies.
Can Gemini or DeepSeek be called GPT?
By design, they're the same type of model: generative, pre-trained, transformer-based. But GPT is the name of OpenAI's models, so it's more accurate to say "large language model" or call the model by its name.
Are GPT and artificial intelligence the same thing?
No. Artificial intelligence is the whole field, neural networks are one approach within it, and GPT is one type of neural network — OpenAI's language models.
Can I use GPT for free?
ChatGPT has a free plan with limits. GPTunneL has no permanent free tier: you pay for the tokens you use, top-ups start at $5, and a short chat with GPT-5.6 Luna costs a fraction of a cent.
Where to try GPT
All the models from the table — GPT-6 Astra, the GPT-5.6 lineup and GPT OSS — are available in GPTunneL on a single balance, alongside Claude, Gemini and DeepSeek. No subscriptions, pay per use. The easiest way to feel the difference between generations is to try it: ask Luna and Astra the same question and compare the answers.