Foundations

Context Window

The maximum number of tokens a model can process in a single request (input + output).

The context window is the total token capacity a model can attend to at once. In 2026, frontier models support 200k–2M token windows — enough for entire codebases or hundreds of pages of documents in a single call.

Larger windows enable long-document reasoning and multi-file code understanding without retrieval. But they come with real costs: quadratic attention (in practice, tuned to be near-linear), higher per-query price, and the 'lost in the middle' phenomenon where facts placed mid-prompt are recalled less reliably.

For most enterprise workloads, targeted retrieval into an 8–32k window still wins on cost, latency and accuracy over stuffing large context. Long windows are a tool for specific escalations, not a default.

Key points

  • Frontier windows in 2026: 200k–2M tokens
  • Cost and latency scale with input tokens — big prompts get expensive fast
  • 'Lost in the middle' — recall drops on facts placed mid-prompt
  • Retrieval into a smaller window still wins economics for most workloads

Common use cases

Whole-document analysis (contracts, filings, research papers)
Multi-file code understanding
Video transcript reasoning
One-off analyst workflows where cost doesn't matter

Frequently asked

Related terms

Applying Context Window in your business?

We help enterprises design, ship, and govern AI systems end-to-end.