What Does GPT Stand For? Meaning, How It Works, Every Version

GPT stands for Generative Pre-trained Transformer. The term is used in three ways:

An architecture: a decoder-only transformer that is pre-trained to predict the next token and then adapted to specific tasks.
A model family: OpenAI’s proprietary large language models, from GPT-1 in 2018 to the GPT-6 series released in September 2026.
“GPTs” in ChatGPT: custom versions of ChatGPT that users configure with their own instructions, extra knowledge and tools.

OpenAI introduced custom GPTs on 6 November 2023. When the GPT Store opened on 10 January 2024, the company said users had already created more than 3 million of them.

What each word in GPT means

Generative

“Generative” means the model creates new content instead of retrieving existing documents. A search engine returns pages that already exist; a generative model writes a new answer token by token. Early GPT models produced only text. Newer models also work with images and audio. This is what separates GPT-style systems from earlier natural language processing tools built for classification or extraction. It also places them inside the broader field of generative AI.

Pre-trained

“Pre-trained” describes a two-stage process. First, the model is pre-trained on a very large corpus of unlabeled text. Nobody tells it the right answers; it learns grammar, facts and patterns of reasoning by predicting the next token over and over. Then the general model is adapted. It may be fine-tuned on smaller labeled datasets or trained with reinforcement learning from human feedback so that its answers are more useful.

Transformer

“Transformer” is the neural-network architecture introduced in the 2017 paper “Attention Is All You Need”. It replaced recurrent neural networks with self-attention, a mechanism that lets every token weigh its relevance to every other token, however far apart they are. Because the whole sequence can be processed in parallel, transformers train far faster than recurrent networks. That is what made models with billions of parameters practical. Our visual guide to the attention mechanism walks through the maths.

How GPT works, step by step

1. Text becomes tokens

A GPT model never sees letters or words directly. Text is first split into tokens with byte-pair encoding (BPE). GPT-1 already used BPE with 40,000 merges, and OpenAI’s tiktoken library is described as “a fast BPE tokeniser for use with OpenAI’s models”. OpenAI’s rule of thumb for English is that “1 token is approximately 4 characters”, or about three-quarters of a word. So 100 tokens are roughly 75 words. Token counts matter because API usage is billed per token and the context window is measured in tokens. See tokenization for details.

2. The model predicts the next token

GPT models are decoder-only transformers with masked self-attention. Each position can attend only to tokens that come before it, never to later ones. For every step, the model looks at the whole sequence so far and calculates a probability for each token in its vocabulary. One token is chosen, appended, and the process repeats. This is called autoregressive generation:

The prompt is tokenized.
The model computes probabilities for the next token from the full sequence.
One token is selected, usually with some controlled randomness.
The token is appended to the sequence.
Steps 2–4 repeat until the model emits an end-of-sequence token or reaches a length limit.

3. Training happens in three stages

Pre-training on large text corpora builds general language ability.
Fine-tuning adapts the model to tasks such as following instructions.
RLHF aligns outputs with human preferences.

The third stage matters more than its size suggests. In OpenAI’s InstructGPT paper (January 2022), “labelers prefer outputs from our 1.3B InstructGPT model over outputs from a 175B GPT-3 model”. A model more than 100 times smaller won because it had been aligned.

GPT vs ChatGPT vs LLM vs AI

These four terms are nested layers, not synonyms.

Term	What it is	Examples
AI	The broad field of machines performing tasks that normally need human intelligence	Spam filters, chess engines, self-driving systems
LLM	A type of AI model trained on large amounts of text to understand and generate language	GPT models, Claude, Gemini, Llama, DeepSeek, Qwen
GPT	OpenAI’s family of LLMs, and the name of the architecture they use	GPT-3, GPT-4o, GPT-6 Sol, the open-weight gpt-oss models
ChatGPT	OpenAI’s chat application built on top of GPT models	The ChatGPT web, desktop and mobile apps

Artificial intelligence is the umbrella. LLMs are one kind of AI model, and GPT is one family of LLMs. ChatGPT is the product most people use to reach that family. Every GPT model is an LLM, but not every LLM is a GPT. Claude, Gemini and Llama use similar decoder-only transformers, yet only OpenAI’s models carry the GPT name. In the same way, ChatGPT is the application and GPT is the engine. Since 7 October 2026, paid ChatGPT plans run on GPT-6 Sol and the Free and Go plans run on GPT-6 Luna.

GPT versions at a glance

OpenAI has not published parameter counts for GPT-4 or any later model. Parameter figures below exist only for the first three generations.

Version	Released	What changed
GPT-1	June 2018	Generative pre-training plus fine-tuning; 12-layer decoder, 117M parameters, trained on BooksCorpus
GPT-2	February 2019	1.5B parameters, about 40 GB of web text; full model withheld until 5 November 2019
GPT-3	May 2020	175B parameters, trained on 300B tokens; few-shot learning; API opened June 2020
ChatGPT	30 November 2022	Chat interface on a model fine-tuned from the GPT-3.5 series
GPT-4	14 March 2023	Large multimodal model: image and text in, text out; size not disclosed
GPT-4o	13 May 2024	“Omni”: text, audio, image and video in; text, audio and image out
o1-preview	12 September 2024	First of the o-series reasoning models that “think” before answering
GPT-4.5	27 February 2025	Research preview
GPT-5	7 August 2025	One system: fast model, deeper reasoning model and a real-time router
GPT-5.1 / 5.2	12 Nov / 11 Dec 2025	Point releases of the GPT-5 family
GPT-5.4 / 5.5	5 Mar / 23 Apr 2026	Point releases of the GPT-5 family
GPT-5.6 Sol, Terra, Luna	9 July 2026	Three tiers, generally available
GPT-6 Astra	3 September 2026	OpenAI’s most capable model; API $10 / $50 per million input / output tokens
GPT-6 Sol, Luna	22 September 2026	GPT-6 Sol at $2 / $10, half the price of GPT-5.6 Sol
GPT-6.1 Sol	29 September 2026	Nearly matches Astra at one-fifth of its price

Current per-token prices for every GPT model are tracked on our API pricing page.

The GPT story, version by version

GPT-1 (2018): pre-training plus fine-tuning

The series began in June 2018 with the paper “Improving Language Understanding by Generative Pre-Training”. It proposed “generative pre-training of a language model on a diverse corpus of unlabeled text, followed by discriminative fine-tuning on each specific task”. That two-stage recipe is still how GPT-style models are built. The model was a 12-layer decoder-only transformer with 117 million parameters. It was trained on BooksCorpus, a collection of “over 7,000 unique unpublished books”.

GPT-2 (2019): scale and a staged release

GPT-2 arrived in February 2019 with 1.5 billion parameters, trained on about 40 GB of text from 8 million web pages. Its release is remembered as much as the model. OpenAI wrote: “Due to our concerns about malicious applications of the technology, we are not releasing the trained model”. Smaller versions came out first, and the full 1.5B model followed on 5 November 2019.

GPT-3 (2020): few-shot learning

The May 2020 paper “Language Models are Few-Shot Learners” introduced GPT-3: 175 billion parameters, with all models “trained for a total of 300 billion tokens”. Its central finding was that “scaling up language models greatly improves task-agnostic, few-shot performance”. GPT-3 could pick up a new task from a few examples in the prompt, with no further training. In June 2020 OpenAI opened its API, so developers could build on GPT-3 without running it themselves.

From InstructGPT to ChatGPT (2022)

Raw GPT-3 was powerful but often ignored what users actually asked for. InstructGPT, published in January 2022, applied RLHF to fix that, and human raters preferred a 1.3B InstructGPT model over the 175B GPT-3. On 30 November 2022 OpenAI launched ChatGPT, “fine-tuned from a model in the GPT-3.5 series”. It brought the same instruction-following approach to a free chat interface.

GPT-4 and GPT-4o (2023–2024)

GPT-4 was released on 14 March 2023 as “a large multimodal model (accepting image and text inputs, emitting text outputs)”. It was also the point where OpenAI stopped describing its models in detail. The technical report gave “no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar”. GPT-4o followed on 13 May 2024. The “o” stands for “omni”: it accepts text, audio, image and video and can output text, audio and image.

Reasoning models and GPT-5 (2024–2026)

On 12 September 2024 OpenAI released o1-preview, the first of its reasoning models, “designed to spend more time thinking before they respond”. They still generate one token at a time. The difference is that they produce a long internal chain of reasoning before the final answer. GPT-4.5 followed as a research preview on 27 February 2025. GPT-5 launched on 7 August 2025 as “one unified system”. It combines a fast model, a deeper reasoning model and a “real-time router” that decides which one handles each request. Point releases followed:

GPT-5.1 on 12 November 2025
GPT-5.2 on 11 December 2025
GPT-5.4 on 5 March 2026
GPT-5.5 on 23 April 2026
GPT-5.6 Sol, Terra and Luna, generally available on 9 July 2026

GPT-6 (2026)

GPT-6 Astra came first, on 3 September 2026, as OpenAI’s most capable model, priced at $10 per million input tokens and $50 per million output tokens (full Astra pricing breakdown). On 22 September OpenAI added GPT-6 Sol and GPT-6 Luna. Sol costs $2 / $10, half the $4 / $20 of GPT-5.6 Sol. A week later, on 29 September, GPT-6.1 Sol “nearly matches GPT-6 Astra’s intelligence” at “one-fifth of Astra’s standard input and output token prices”. On 7 October 2026 GPT-6 reached ChatGPT. OpenAI said at the time that more than 1.2 billion people use ChatGPT each week.

What “GPTs” are in ChatGPT

In ChatGPT, “GPTs” means something different from the model family. OpenAI announced them on 6 November 2023 as “custom versions of ChatGPT that combine instructions, extra knowledge, and any combination of skills”. You can give a GPT a fixed set of instructions, upload reference files and switch on tools, all without writing code. You can then keep it private or share it. The GPT Store opened on 10 January 2024; by then users had created more than 3 million custom GPTs. Every custom GPT still runs on one of OpenAI’s underlying GPT models. The customisation sits on top through prompts, files and tools.

GPT-style models beyond OpenAI

The architecture behind GPT is not exclusive to OpenAI.

gpt-oss (5 August 2025): OpenAI’s own open-weight models, gpt-oss-120b and gpt-oss-20b, under the Apache 2.0 licence. Both are mixture-of-experts models: 117B and 21B total parameters, with 5.1B and 3.6B active per token. The 120b runs on a single 80 GB GPU; the 20b needs 16 GB of memory. You can run them yourself, for example locally with Ollama.
EleutherAI: GPT-J-6B (mid-2021) and GPT-NeoX-20B (February 2022) are open GPT-style models, also under Apache 2.0.
Other chat LLMs: Claude, Gemini, Llama, DeepSeek and Qwen also use decoder-only transformers. They differ from GPT in training data, scale and fine-tuning, but not in the basic idea. Only OpenAI’s models carry the GPT name.

Is “GPT” a trademark?

Not in the United States. In February 2024 the US Patent and Trademark Office refused OpenAI’s application to register “GPT”. It said the term “merely describes a feature, function, or characteristic”. It refused the application again in 2025 as likely generic. OpenAI still restricts the name through its own brand guidelines, which say it does “not permit our GPT brand to be used in app, product, developer or company names”. So “GPT” is free to use as a technical description, but OpenAI does not allow developers on its platform to put it in their product names.

Limitations to keep in mind

Hallucinations. GPT models can state false things with full confidence. They predict likely text, not verified facts, so any figure that matters needs checking. See LLM hallucinations for causes and fixes.
Knowledge cutoff. A model only knows what was in its training data. Without web search or retrieval, it cannot know about later events.
Cost by tier. Prices differ by model. On the API, GPT-6 Astra costs five times as much as GPT-6 Sol per token ($10 / $50 vs $2 / $10 per million tokens). Our guide to reducing LLM API costs covers how to choose.
Opacity. OpenAI has not disclosed model size, training data or compute for GPT-4 or any later model. That makes independent comparison harder.

FAQ

What does GPT stand for?

GPT stands for Generative Pre-trained Transformer. “Generative” means it produces new content. “Pre-trained” means it first learns from a large body of unlabeled text and is then adapted. “Transformer” is the neural-network architecture, introduced in 2017, that uses self-attention to process text.

What does GPT stand for in ChatGPT?

The same thing: Generative Pre-trained Transformer. ChatGPT is OpenAI’s chat application, and “GPT” in its name refers to the model family that powers it. At launch in November 2022 it ran on a model from the GPT-3.5 series. Since 7 October 2026 it runs on GPT-6 Sol for paid plans and GPT-6 Luna for Free and Go.

What is the difference between GPT and ChatGPT?

GPT is the model; ChatGPT is the app. Developers reach GPT models through OpenAI’s API and pay per token. ChatGPT wraps a GPT model in a chat interface with conversation history, file uploads, tools and custom GPTs.

Is GPT a large language model?

Yes. Every GPT model is a large language model: a transformer trained on very large amounts of text to understand and generate language. GPT is one family of LLMs; others include Claude, Gemini, Llama, DeepSeek and Qwen.

What is the latest GPT version?

As of 9 October 2026, the newest release is GPT-6.1 Sol, launched on 29 September 2026. GPT-6 Astra (3 September 2026) remains OpenAI’s most capable model, and GPT-6 Luna is the version used on ChatGPT’s Free and Go plans.

Who created GPT?

OpenAI. The first model, GPT-1, was published in June 2018, followed by GPT-2 (2019), GPT-3 (2020), GPT-4 (2023), GPT-5 (2025) and GPT-6 (2026). The transformer architecture itself came from a 2017 paper by Google researchers. Other labs have since built GPT-style models, such as EleutherAI’s GPT-J and GPT-NeoX.

Is GPT free to use?

Partly. ChatGPT has a free plan, which runs on GPT-6 Luna as of October 2026; paid plans run on GPT-6 Sol. The API is paid per token: GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens, and GPT-6 Astra $10 and $50. OpenAI’s gpt-oss models are free to download and run under the Apache 2.0 licence.