The context window determines how much information a model can 'see' at once. Early GPT models had 4K token windows (~3,000 words). Modern models offer 200K+ tokens (~150,000 words) — enough to process entire codebases or books.
Larger context windows enable: longer conversations, more context for RAG, full-document analysis, and complex multi-step reasoning. However, model performance can degrade with very long contexts (the 'lost in the middle' problem).