Definition
In OpenAI's model catalog, GPT identifies its Generative Pre-trained Transformer lineage. The expansion also describes the generative pretraining approach more broadly. OpenAI's lineage began with language models that learned from token sequences before being adapted to downstream tasks; later releases have different reasoning and modality capabilities.
GPT-6 Astra and GPT-6 Luna are distinct models. The older GPT-4 report describes a model that accepts images and text. Check supported inputs, outputs, and interfaces for the exact release.
Origin
OpenAI published its generative pretraining work on June 11, 2018. Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever authored the accompanying paper, which combined a left-to-right transformer language model with supervised adaptation. That paper documents the lineage's starting point. It builds on earlier transformer and pretraining research.
Family, model, and application
GPT is a lineage. Sol, Luna, Astra, and Terra are names OpenAI has used for tiers within later generations. A complete model identifier specifies more than the tier name, such as gpt-6.1-sol.
ChatGPT is an application that can use GPT models. Codex is an agent environment that supplies tools and execution controls around selected models. A custom GPT in ChatGPT is an application configuration; creating one does not, by itself, train new model weights.
Operational significance
Record the model identifier, endpoint, reasoning settings, instructions, and enabled tools when evaluating a GPT-based system. Model results and application results answer different questions: the surrounding system determines which files, services, and actions are available.
The GPT-4 technical report documents factual errors and limitations in checking benchmark contamination. Later branding does not remove the need to verify outputs or measure performance on the intended workload. Claims about one release should retain that release's name and evaluation conditions.
Distinguish it from nearby terms
- A transformer is an architecture family used by many developers. GPT identifies OpenAI's lineage.
- A large language model is a broader model category; it includes models developed outside OpenAI.
- Sol and Astra identify different named tiers within the GPT family.
- An AI agent combines a model with an action loop, state, and permissions.
Check your understanding
A report says that "GPT completed the task" but gives no model identifier, reasoning setting, or tool list. Which details would you need to reproduce the result and determine whether the model or the surrounding application supplied the required capability?