Repository navigation
Work mode used ~90% of 5-hour allowance in 13m24s on GPT-5.6 Sol Medium without producing requested artifact #41330
Description
What issue are you seeing?
I used ChatGPT Work mode with GPT-5.6 Sol Medium for a No Man’s Sky save-editing/reverse-engineering project.
The run lasted 13 minutes 24 seconds and afterward my Work usage meter showed only about 10% remaining in the 5-hour window.
The prompt explicitly asked Work mode to continue an existing project and build/extend a reusable mission cleaner/database workflow, not just perform one-off research.
The run produced some groundwork:
- identified a few mission IDs / terminal progress values;
- recreated or validated save/meta rebuild tooling;
- performed a zero-change validation test;
- produced a Phase 1 report.
However, it did not produce the requested reusable program/tool/database. It stopped after claiming the available save attachment was stale and asked for a fresh upload.
A second concern is file/context grounding. In that Work conversation, I had not uploaded any fresh save.hg or mf_save.hg files yet, but the response stated that it had recovered attached save files and reported specific mission-state values from them. If those came from an older referenced conversation, that should have been stated clearly as stale inherited data rather than presented like current workspace evidence.
The main bug I am reporting is the combination of: - extremely high usage consumption for a short Medium run;
- incomplete delivery of the requested artifact;
- and potentially misleading handling of inherited/stale file context.
What steps can reproduce the bug?
- Start a ChatGPT Work task using GPT-5.6 Sol Medium.
- Continue a technical project from a referenced prior chat containing file-based reverse-engineering context.
- Ask Work mode to:
- inspect the existing methodology;
- build a reusable tool/program around it;
- only patch current files if current files are actually available;
- otherwise request fresh files.
- Do not upload fresh files into the new Work thread.
- Let the Work run execute.
Observed result in my case:
- visible runtime: 13m24s;
- usage meter afterward: about 10% remaining from the 5-hour allowance;
- Work reported stale/recovered save data and specific mission values;
- no reusable mission cleaner/program was delivered.
What is the expected behavior?
A 13-minute Medium Work run should not consume nearly the entire 5-hour allowance unless there is genuinely extreme underlying usage, and if it does, the UI should make that resource usage understandable.
The model should also clearly distinguish:
- files uploaded in the current Work thread;
- stale files inherited/recovered from a referenced conversation;
- and current live user data.
Most importantly, if the task asks for a reusable program/tool and current live files are unavailable, Work should continue building the reusable tool architecture rather than spending most of the run on stale-data analysis and stopping before the requested artifact is produced.
Additional information
I have a direct comparison from the same project in normal ChatGPT chat.
Normal ChatGPT, with tighter file-delivery constraints, successfully:
- decoded the native No Man’s Sky save;
- diagnosed a failed first patch using fresh game behavior;
- identified the required parent/child mission-state dependency;
- rebuilt the native .hg save and matching metadata;
- validated the output;
- and returned working downloadable patched files.
By contrast, Work mode on Sol Medium consumed almost the entire 5-hour allowance while delivering substantially less practical output.
I can provide: - the Work transcript;
- the screenshot showing “Worked for 13m24s”;
- the screenshot showing ~10% usage remaining;
- and the original ChatGPT conversation for comparison.
I am a paying ChatGPT Plus subscriber in the UK. If this level of Work-mode usage consumption is expected behaviour rather than a bug, I would like clarification on how users are informed of it in advance and how a paying user is supposed to reasonably predict that a single 13m24s Medium-mode run could consume almost an entire five-hour allowance.
Activity
github-actions commentedon Aug 28, 2026
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
motion-bot-byte commentedon Aug 28, 2026
I reviewed #40880, #41131, and #40965. They overlap on quota consumption and workflow regressions, but this report is not a duplicate. My case is specifically a single 13m24s GPT-5.6 Sol Medium Work run consuming ~90% of the 5-hour allowance while failing to produce the requested artifact and operating on stale inherited file context despite no fresh upload in the new Work thread. I also want to stress that the referenced normal ChatGPT conversation had already handled the same class of problem successfully without consuming any of my Codex/Work allowance. That chat decoded and patched the save, learned from a failed first attempt, identified the actual parent/child mission-state dependency, rebuilt the native save pair, and returned working files. Work mode then inherited that methodology and context, consumed ~90% of the 5-hour allowance in a single 13m24s Sol Medium run, and still delivered less usable output. That makes the usage discrepancy itself part of the bug: the more heavily metered environment performed materially worse on the same problem than the unmetered chat it inherited from.
bjorkian commentedon Sep 8, 2026
Update: the same Work task is still running and has now reduced my 5-hour allowance from 100% to 1% remaining, while the task is still far from complete.
A substantial part of the remaining work involves organising photographs on the Mac.
So a single GPT-5.6 Sol Medium Work task has effectively consumed ~99% of an entire freshly reset 5-hour allowance without completing the requested task.
My weekly allowance remains comparatively high (previously 85% remaining), which makes the discrepancy between the short-term and weekly limits particularly striking.
I am stopping the task before the 5-hour allowance reaches 0%, to avoid it being interrupted during file operations.
Final update: the task has now exhausted 100% of my freshly reset 5-hour allowance and is still not complete.
This was a single ChatGPT Work task using GPT-5.6 Sol at Medium reasoning. The task was organising files and photographs on my Mac.
So, in this case, one unfinished Work task consumed the entire 5-hour allowance.