LLMs with the largest context windows

A context window is a ceiling, not a promise. Every model degrades before it reaches its advertised limit, and the shape of that degradation matters more than the headline number. The well-documented pattern is that material at the very start and very end of a long prompt is recalled reliably while material in the middle is often missed — the lost-in-the-middle effect — which means a million-token window does not mean a million tokens of dependable attention.

Retrieval benchmarks flatter models here. Finding one planted sentence in a long document is far easier than reasoning across twelve facts scattered through it, and most published long-context results measure the former. If your task needs synthesis across a whole corpus rather than lookup within it, expect real performance well below the advertised capability.

Long prompts are also expensive and slow. Input tokens are billed on every call, so a large stable prefix multiplies cost across a conversation unless the provider caches it, and time to first token grows with prompt length. A retrieval step that sends four relevant pages instead of four hundred is usually faster, cheaper and more accurate than relying on the window.

Where large windows genuinely earn their place: whole-repository code understanding, long legal or financial documents that resist clean chunking, multi-hour transcripts, and agent loops that accumulate long tool-call histories. In those cases the window removes an entire class of chunking bugs, and that is worth paying for.

Practical rule — use retrieval by default, use the long window when chunking would destroy the structure you need, and always test recall on your own documents at the length you actually intend to use, not at the length the specification allows.

More on how these numbers are produced in the methodology and what each evaluation measures in the benchmark guide. Spotted a score that disagrees with its source? Tell us.