A very large, general-purpose model trained on broad data. The base layer that specific applications build on.
Foundation models are the big ones: GPT-5, Claude Opus, Gemini, Llama. They cost hundreds of millions to train. They're designed to be broadly capable - handling writing, code, math, reasoning, and more. Everything downstream (fine-tuned models, agents, RAG pipelines, ChatGPT itself) is built on top of a foundation model. The term 'foundation' is literal: without them, nothing else in the AI product layer exists.
Picking the right foundation model is the single biggest product decision you'll make. The rest - prompts, tools, UI - can be rebuilt. The model choice determines what's even possible.