Does 80% of a 256k context window degrade the same as 80% of 1M+?

BriefServe453 · reddit · 2026-09-11

A user asks whether retrieval and formatting accuracy at 80% utilization of a 256k window (200k tokens) matches 80% of a 1.05M window (840k tokens), i.e., whether context degradation scales by fill percentage or absolute token mass. Calls for benchmarks comparing deep retrieval at both scales.

Original post →

More from Models

Models channel →