TL;DR: GPT is OpenAI's family of large language models and the one that started the modern chatbot era through ChatGPT. The flagship models are closed-weight and hosted: you reach them through OpenAI's API, through cloud platforms such as Azure OpenAI and Amazon Bedrock, or inside the ChatGPT product line. The family's defining trait is breadth - general capability plus the largest third-party ecosystem.
What it is and how you use it
GPT stands for "generative pre-trained transformer," and the name is literal: these are transformer models pre-trained on large text corpora, a recipe OpenAI's GPT line did more than any other to make mainstream. OpenAI ships the family in generations, each spanning a flagship tier and smaller, cheaper tiers. Spending extra compute thinking before answering - once a separate reasoning model line, the o-series - is now a dial on the main models rather than a different product: you set a reasoning effort per request, from none up to maximum, and the o-series models are being retired. Treat the specific version numbers as perishable; the tiered shape of the lineup is what persists.
What the family is known for is being the default. ChatGPT made it the model most people met first, so tutorials, libraries, and third-party integrations tend to target OpenAI's API shape before any other - many competitors even offer OpenAI-compatible endpoints so existing code works unchanged. The models themselves are strong generalists across writing, coding, analysis, and multimodal input, which is why "just use GPT" remains a common starting position before teams profile their real needs against alternatives like Claude or Gemini.
Access is hosted. The flagship GPT line is closed-weight, reached through OpenAI's own API or through a cloud platform - Azure OpenAI, now served inside Microsoft Foundry, or Amazon Bedrock inside AWS - which serve the same models with enterprise networking and compliance attached. OpenAI's own SDKs ship provider hooks for both. OpenAI has also published open-weight models - the Apache-2.0 gpt-oss line and its gpt-oss-safeguard safety classifiers, which you can download and run yourself - but those are the exception, and the family's center of gravity is the API. Consumers and businesses without integration needs use the ChatGPT apps directly, which bundle the models with tools, memory, and file handling.
Where it typically fits: general-purpose assistants, content and analysis workloads, and any project that benefits from the deepest pool of examples, libraries, and hosting options. The ecosystem advantage is real but not decisive - the frontier families trade places on specific tasks often enough that the sensible move is to shortlist GPT alongside its peers and let your own evals settle it.
Where it sits in the AI stack
GPT sits at the model layer: your code sends a prompt to a hosted endpoint, OpenAI or a cloud partner runs the inference, and tokens stream back. The infrastructure is never yours to run:
Key tools and implementations
-
OpenAI API
The first-party endpoint whose request format became a de facto industry standard - now the Responses API, with the older Chat Completions shape still supported.
-
Cloud platforms
The same models served through Azure OpenAI in Microsoft Foundry and through Amazon Bedrock, for teams that need their existing cloud's billing, networking, and compliance.
-
Official SDKs
Python and TypeScript client libraries that wrap the API, widely imitated by compatible providers.
-
ChatGPT products
The consumer and enterprise apps that bundle the model family with tools, memory, and integrations.
Related entries
- Claude Anthropic's family of language models, accessed through an API and major cloud platforms rather than downloadable weights.
- Gemini Google's multimodal model family, accessed through the Gemini API, Google Cloud, and Google's own products.
- Grok xAI's closed-weight model family, known for live X and web search tools, served through its own API, cloud platforms, and the Grok apps.
- Model benchmarks Standardized tests that score model capabilities, useful for rough comparison but easy to overfit and game.
- LLM (large language model) A neural network trained on huge amounts of text to predict the next token, which lets it generate and understand language.