GPT (Generative Pre-trained Transformer)

GPT — short for Generative Pre-trained Transformer — is the family of large language models from OpenAI. It's the family that brought LLMs into the mainstream: the public launch of ChatGPT in late 2022 (built on a GPT model) is what turned LLMs from a research curiosity into an everyday tool used by hundreds of millions of people.

💡 In one line: GPT is OpenAI's family of decoder-only LLMs — the models behind ChatGPT and the spark of the modern AI boom.

What is GPT?

GPT models are decoder-only Transformers (see the GPT subtopic in the Transformer series) trained with next-token prediction. They generate text autoregressively — one token at a time — and the larger versions are multimodal, handling images and audio alongside text. Each new generation has scaled up and added capabilities like deeper reasoning and tool use.

GPT vs. ChatGPT

A common point of confusion:

  • GPT = the underlying models (the AI).
  • ChatGPT = the assistant product that uses GPT models through a chat interface, with extra features like memory, tools, and search.

So GPT is the engine; ChatGPT is the car.

The GPT Lineage

GPT has evolved through several landmark generations:

  • GPT-1 (2018) — the original proof of concept.
  • GPT-2 (2019) — bigger, noticeably more fluent.
  • GPT-3 (2020) — huge scale-up; strong few-shot learning.
  • ChatGPT / GPT-3.5 (2022) — the mainstream breakthrough.
  • GPT-4 (2023) — much stronger reasoning, multimodal (image input).
  • GPT-4o (2024) — an "omni" model unifying text, audio, and image.
  • o-series (2024–25) — separate reasoning models that "think before answering."
  • GPT-5 (2025) — a unified system that pairs a fast model with a deeper reasoning model behind a router that decides how much to think.

As of mid-2026, the lineup is the GPT-5 generation (with rapid point releases like GPT-5.5 and a GPT-5.6 preview). The exact version names change often.

Key Characteristics (Current Generation)

  • Unified design — a fast path ("Instant") and a deeper reasoning path ("Thinking"), with a router choosing per request.
  • Tiers — typically an instant model, a reasoning model, and a top-end "Pro" variant.
  • Multimodal — text, images, and audio.
  • Large context windows and strong performance on coding, reasoning, and agentic tasks.

What GPT is Known For

  • Mainstreaming LLMs via ChatGPT.
  • General-purpose strength across writing, Q&A, and analysis.
  • Coding (including OpenAI's Codex coding agent).
  • A broad ecosystem of related models (image generation, video, speech).

Open or Closed?

GPT's flagship models are proprietary — accessed through ChatGPT or the OpenAI API, not downloadable. OpenAI has, however, released some open-weight models separately. (See the Open-source subtopic for how this contrasts with fully open families like Llama and DeepSeek.)

A Note on Versions

The GPT lineup changes rapidly — new versions, tiers, and capabilities arrive every few months. Treat specific version names as a snapshot; always check OpenAI's site for the current models.

Summary

  • GPT is OpenAI's family of decoder-only LLMs, behind ChatGPT.
  • It evolved from GPT-1 (2018) through GPT-3, GPT-4, and the GPT-5 generation (2025–26).
  • GPT = the models; ChatGPT = the product built on them.
  • The current generation unifies fast and reasoning models behind a router, and is multimodal.
  • Flagship GPT models are closed/proprietary (API + ChatGPT), and the lineup evolves quickly.