What's this about?
After weeks of complaints that Codex was burning quota too fast, OpenAI reset usage for all paid Codex and ChatGPT Work users and published what it found. Depending on usage patterns, quota should now go 10–50% further. The specifics are unusually candid: compaction was keeping old images, sometimes leaving the context large enough to immediately trigger compaction again — about 10% saved for image-heavy users. Background memory workers were inheriting Stop hooks, so when a hook blocked them from finishing they kept spinning; under 1% of users affected, but one thread was observed checking whether it could stop 15,000 times.
What is "compaction" and why did it loop?
When a long session fills the context window, the harness summarises older content to make room. The bug was that images were being carried through the summary rather than dropped, so the compacted context was still nearly full — which triggered compaction again, which cost another expensive model call, and so on. You paid for the summarising, repeatedly, to make room that was never actually made.
What's the community saying?
- The reception is unusually warm for a quota post, mostly because the fixes are specific and verifiable rather than "we've improved efficiency".
- The 15,000-iteration stop check is the detail everyone quotes — a reminder that agent harnesses have loops with real costs and few people are watching them.
- The follow-up, "I challenge you to burn as much usage as you can", reads as confidence that the leaks are actually closed.
- Both compaction and memory appear here as *billing* bugs, which is a useful reframing: harness inefficiency is not an abstraction, it is your quota.