Glossary
What is OpenAI?
In short
OpenAI is an American company that develops the GPT models and services built on them: ChatGPT, the Codex agent and the open gpt-oss models.
Companies · OpenAI
MMOOZZGG.com editorsChecked: 1 min read
For a child
Imagine a workshop where talking robots go to school. First they read mountains of books, then they learn to answer politely and clearly, and then they go out to help people: write letters, explain lessons, make up stories. This company is that workshop, only its robots live inside computers and talk to you through a screen.
In plain words
Most people know the company for ChatGPT, a chatbot: you type a question, and the model answers, writes texts and reads files and pictures. For programmers there is Codex, an agent that edits code in a project by itself. Developers connect the models through an API and build them into their own programs. In 2025 the company released the open models gpt-oss-120b and gpt-oss-20b, which you can download and run yourself. OpenAI shut down its Sora video model.
For business
Companies use OpenAI models in customer support, sales, document analysis and software development, through ChatGPT for business or through the API in their own systems. For Russia, the key fact is this: the country is not on the list where OpenAI supports access, and use from unsupported countries can lead to a blocked account. The open gpt-oss models under Apache 2.0 offer a separate path: you may run them on your own servers, and data does not go to the vendor. The larger model needs one 80 GB graphics card, the smaller one 16 GB of memory. Teams compare which option is legal and cost-effective with 152-FZ, Russia’s personal data law, in mind.
For a specialist
Two well-known OpenAI papers deal with large language models. GPT-3 (Brown et al., 2020) showed that an autoregressive model with 175 billion parameters solves new tasks from a few examples in the prompt (few-shot). InstructGPT (Ouyang et al., 2022) established fine-tuning on demonstrations and human rankings of answers, followed by reinforcement learning (RLHF); chat assistants grew from this scheme. The gpt-oss-120b and gpt-oss-20b models are mixture-of-experts (MoE) transformers: 117 billion parameters with 5.1 billion active, and 21 billion with 3.6 billion active, trained with distillation and reinforcement learning. The weights, inference code, tool environments and tokenizer are open under Apache 2.0; the larger model fits on one 80 GB GPU thanks to MXFP4 quantization.
Example
A bank wants an AI assistant for its staff but cannot send documents abroad. The team runs gpt-oss-20b on its own server and connects it to an internal knowledge base.
Related terms
Questions and answers
Sources
- Brown et al., Language Models are Few-Shot Learners, 2020arxiv.org
- Ouyang et al., Training language models to follow instructions with human feedback, 2022arxiv.org
- OpenAI, gpt-oss-120b & gpt-oss-20b Model Card, 2025arxiv.org
- OpenAI, openai/gpt-oss, model repository, checked 16 October 2026github.com
- OpenAI, Supported countries and territories, checked 16 October 2026developers.openai.com
Not sure whether your company needs this?
AI solutions map