Tips: Long Sessions Waste Tokens and Degrade Model Performance

hudsonhateno · reddit · 2026-09-02

A Reddit user explains two main issues with long-running LLM sessions: token consumption grows linearly with context processing, and model output quality degrades significantly once context usage exceeds 40%. The post lists context windows and max output limits for various Codex models and recommends "handing off" sessions—summarizing the current state and starting a new one—when capacity reaches 40%-50%. The author also shares two custom skills: $context-capacity to calculate current context load and cumulative usage, and $handoff to assist in the session handoff process.

Original post →

More from coding & agent

coding & agent channel →