Context Window
The maximum amount of text an LLM can consider at once. Measured in tokens. Larger windows enable longer documents and conversations.
An LLM's context window is the maximum number of tokens (input + output combined, in some models) that it can process at once. Anything beyond the context window is forgotten or truncated.
Context windows have grown dramatically: GPT-3 had 4,096 tokens. GPT-4 reached 128K. Claude 4 supports 200K. Gemini 2.5 supports 1M+ tokens. A 1M-token window can hold roughly 750,000 words — equivalent to most novels or 500+ pages of documents.
Larger context enables: processing entire codebases, long PDF analysis, extended conversations without forgetting earlier context, and more detailed in-context learning. But longer contexts also cost more (linear in input length) and can degrade quality on very long inputs ("lost in the middle" effect). Best practice: only include relevant context, not everything you have.