Context Window
Maximum tokens an LLM can consider in one request: prompt, history, files, and response combined.
- Frontier models advertise 200K-2M tokens.
- Real recall degrades before the stated limit.
- Long-context models cost more and respond slower.
When to use: When deciding between long context vs retrieval (RAG, file-mention).
Example: Claude Code's 1M-token context can load a mid-size codebase wholesale; 200K window holds ~500 pages of text.