Back to full roadmap
topicfoundation
Context Window
The total tokens the model can hold in one turn. Must-know.
2 hours2 resources1 prereqs
Each model has a context limit: GPT-4o ≈ 128K, Claude 4.7 (1M) → 1M, Gemini 2.x ≈ 2M. Exceeding it triggers a rejection or truncation.
Important nuance: Long context suffers from the "lost in the middle" effect — info at the beginning and end is recalled best, the middle fades. So:
- Place critical instructions at the start AND end
- Add structure (XML / markdown sections) in long contexts
- Split into multi-turn or RAG when needed