Codex Repeatedly Compresses and Burns Tokens in Long Tasks
aigclink · x · 2026-07-19
A post pointed out that Codex can insanely burn through tokens in long tasks due to a flaw in its compression design: after compression, the model only sees the initial instructions, causing it to repeatedly enter an "initial instructions -> compress -> execute initial instructions" loop. It might only complete the task by chance if it skips compression; otherwise, it drains the entire quota.
The accompanying image further illustrates this runaway loop: once the context is compressed, the working memory is repeatedly reset, leading to redundant execution and uncontrolled token usage. The poster tagged relevant personnel, hoping they would look into the issue.
Related event: Alibaba Announces 2.4T Open-Weight Model Qwen3.8(22 posts)→
More from coding & agent
- Alex Townsend posts 200 open problems in numerical linear algebra for humans and AI agents — IgorCarron · 2026-09-11
- Kimi K2.8 Preview rolls out: near-K3 coding performance, 1M context for all tiers — teortaxesTex · 2026-09-11
- Looking for a classifier of software engineering task shapes to pick models per task — StewartalsopIII · 2026-09-11
- Steal this idea: prompt-to-hardware where agents assemble custom devices — paraschopra · 2026-09-11
- Model Is the Least Interesting Part: A Guide to Six Core AI Architectures from RAG to Multi-Agent — goyalshaliniuk · 2026-09-11
- Non-coder builds layered memory architecture: 20k tokens tracks a year of agent conversations — matteoianni · 2026-09-11