NOOSIGMA AI / TRIAD / 02OpenAI and GPT: who it is, history, products and architecture
OpenAI created the GPT model series and ChatGPT, the assistant that brought generative AI to the general public. This article covers who the company is, the path from GPT-1 to GPT-6, its products and what is public about the architecture.
- Founded in December 2015
- ChatGPT launched on 30 Nov. 2022
- Current series: GPT-6
Reviewed on
Who it is
OpenAI is an AI research and products company based in San Francisco. It was created on 8 December 2015 as a nonprofit, with Sam Altman and Elon Musk as co-chairs and names such as Greg Brockman and Ilya Sutskever among the founders. Sam Altman is the CEO.
In 2019 the company adopted a capped-profit model under a nonprofit parent, and Microsoft invested US$ 1 billion. According to Wikipedia, the restructuring completed on 28 October 2025 turned the for-profit part into OpenAI Group PBC (a public benefit corporation) and the nonprofit into the OpenAI Foundation.
OpenAI became known worldwide with ChatGPT, launched on 30 November 2022. Reports say it reached 100 million monthly users in about two months, which accelerated the race among AI companies.
Timeline
- 2015
Founded as a nonprofit, on 8 December.
- 2018
GPT-1 (11 June), the first model in the series. Elon Musk leaves the board.
- 2019
GPT-2, with 1.5 billion parameters. OpenAI adopts the capped-profit model and partners with Microsoft, which starts hosting its services on Azure.
- 2020
GPT-3, with 175 billion parameters.
- 2022
InstructGPT, tuned with RLHF, and ChatGPT (30 November), initially based on GPT-3.5.
- 2023
GPT-4 (14 March) and custom GPTs (November). On 17 November the board removes Sam Altman, who is reinstated days later.
- 2024
GPT-4o (May), which handles text, image and audio, and o1 (December), the first of the reasoning models.
- 2025
Operator (January), Deep Research (February), Codex (May), ChatGPT agent (July), GPT-5 (August), ChatGPT Atlas (October) and the corporate restructuring (28 October). GPT-5.2 in December.
- 2026
A GPT-5.x sequence through the year and, in September, GPT-6: Astra (3 and 4 September), then Sol and Luna (22 September) and GPT-6 Instant (early October), with the Intelligent UI feature.
- Oct. 2026
OpenAI announces that its models and Codex are coming to Google Cloud (Model Garden) and presents ChatGPT Slides, to create and edit presentations as a team.
Products
ChatGPT
The general-purpose assistant, on web, desktop and mobile, with web search, in-depth research (Deep Research), agents and its own browser (Atlas).
OpenAI API
Access to the models for developers and companies, and now also through Google Cloud, as announced.
Codex
A coding agent, with a command-line version.
GPTs and GPT Store
Custom versions of ChatGPT created by users, launched in 2023 and 2024.
GPT Image and Whisper
Models for generating images and recognizing speech.
ChatGPT Slides
Creating, editing and sharing presentations as a team, rolling out by plan.
Scientific research
Together with PerturbAI and Scripps Research, OpenAI used thousands of agents with GPT-Rosalind in a project on mouse brain cells; the hypotheses generated still need experimental validation.
Architecture and training
GPT models use only the decoder part of the transformer architecture, introduced in 2017, and are pretrained to predict the next token in a sequence. The acronym stands for “generative pre-trained transformer”.
Training has two stages: generative pretraining on large amounts of unlabeled text and a later tuning step. After GPT-3, OpenAI started using reinforcement learning from human feedback (RLHF) to align behavior with people’s preferences; that is how InstructGPT and ChatGPT came about.
Reasoning models, such as o1 and o3, were trained with reinforcement learning to produce multi-step chains of reasoning before the final answer. GPT-4o processes text, image and audio. With GPT-5, in 2025, ChatGPT began choosing between models through a router.
About GPT-6, reports say the Astra series was the company’s largest training run to date, on more than 100,000 GPUs at the Stargate complex in Texas. This comes from a secondary source, and OpenAI has not confirmed the architecture itself.
What to look at
ChatGPT is the best-known general-purpose assistant and receives features in stages, depending on plan, region and account. The GPT-6 Instant announcement, for example, is about the ChatGPT experience, not about any app built with the API.
The company and the press also report controversies, such as lawsuits and agent safety issues. For professional use, compare cost, privacy and quality on real tasks, and check the terms of your plan.