From GPT-1 to GPT-5: The Complete History of OpenAI’s GPT and ChatGPT

From GPT-1 to GPT-5: The Complete History of OpenAI’s GPT and ChatGPT featured illustration
⏱ 3 min readUpdated 28 September 2026

GPT stands for Generative Pre-trained Transformer. Each version showed that scaling up the same basic idea — predict the next token — unlocks new abilities. (How that works: Lesson 1 of our LLM course.)

In this article
  1. The timeline
  2. Three ideas that explain the progress
  3. Free and paid access
  4. Continue learning

The timeline

Model When Why it mattered
GPT-1 2018 ~117M parameters. Proved “pre-train on lots of text, then fine-tune” works across tasks.
GPT-2 2019 1.5B parameters. Wrote surprisingly fluent paragraphs; released in stages over misuse concerns.
GPT-3 2020 175B parameters. Could follow instructions from a few examples in the prompt (“few-shot”) — no fine-tuning needed.
ChatGPT (GPT-3.5) Nov 2022 Chat format plus training from human feedback made it helpful for ordinary people; reached an estimated 100 million users in about two months.
GPT-4 Mar 2023 Big jump in reasoning and exam-style tests; could read images.
GPT-4 Turbo Nov 2023 128K context, cheaper.
GPT-4o May 2024 “Omni”: text, images and audio in one fast model; real-time voice conversations.
o1 and the o-series From Sep 2024 Reasoning models that “think” (spend extra computation) before answering — much stronger at maths and code.
GPT-4.1 Apr 2025 Developer-focused: better coding and instruction following, very long context.
gpt-oss-20b / 120b Aug 2025 OpenAI’s first open-weight models since GPT-2 — downloadable and free to run. Both are available in this site’s AI Helper.
GPT-5 Aug 2025 Unified fast answers and deeper reasoning, routing between them automatically.

OpenAI continues to release updated versions; check the official model list for the latest names, prices and limits.

Advertisement

Three ideas that explain the progress

  1. Scale: more parameters, data and compute kept improving results.
  2. Alignment: training with human (and AI) feedback turned a text predictor into an assistant.
  3. Test-time reasoning: letting a model think longer on hard questions improves accuracy — at the cost of speed.

Free and paid access

Option Cost Notes
ChatGPT free tier Free Usage limits on the strongest models; limits change over time.
ChatGPT Plus / Pro Monthly Higher limits and newest features.
API Pay per token For building your own apps.
gpt-oss Free to run On your own hardware or free cloud tiers.
⚠️ Whatever the version, verify facts, numbers and code. Newer models hallucinate less, not never.
✨ Ask AI about this article

Stuck on a step? Ask a question and the AI answers using this article.

Free · AI can be wrong

Leave a Reply

Your email address will not be published. Required fields are marked *