Update — August 25, 2026: Separate from this weekend's bug reset — Tibo says the rolling 5-hour cap returns tomorrow for Plus on Work and Codex; Pro $100/$200 stay uncapped for months. Plus 5h limit returns.
If your Codex or ChatGPT Work bar melted this weekend, the useful question is not "did OpenAI secretly shrink Plus." It is which background work was billing you, and whether the model even finished the list you gave it.
On August 23, 2026, Codex and ChatGPT lead Tibo Sottiaux named three drains, promised a tiger team, and said paid accounts would get a full usage reset with the next day's fixes. On August 24 he posted that the reset had propagated and some of those fixes were live. Plus users immediately reported Sol Max sessions that barely moved the meter. That is the opposite of the Friday–Saturday experience.
This is a new incident, not the July 29 Sol appetite reset and not the August 21 20M banked credit that slipped past its 8pm PST deadline. Same engineer, same meter, different bug list.
TL;DR
| Question | Direct answer |
|---|---|
| Did OpenAI cut Plus/Pro quotas? | Tibo says no. He listed product inefficiencies, then reset paid usage |
| What was actually burning the bar? | Images + repeated compaction, Computer History at the tail, a title generator |
| Cache? | Also real for some users that week — worse hit rate means the same turn costs more |
| Reset type | Hard refill for paid subscriptions on Aug 24, not a banked reset |
| Continue after 0%? | Community reports say that quirk is gone. Do not chase it |
| Did the model finish your ticket? | Not always — and it often will not tell you |
The week in three posts
Saturday, August 22. Tibo's first diagnosis was caching: for some users, cache hit rate was worse than the stable weeks before, and cache misses make Codex usage drain faster. That is a real mechanism. It is also an incomplete one. Replies asked for the actual hit-rate numbers. They did not get a dashboard. They got "we are investigating."
Sunday morning, August 23. explainx.ai founder Yash Thakker (@goyashy) replied that he was still hitting limits faster than before. That was the right read of the cache-only story. A worse hit rate can explain a few extra percent. It does not, by itself, explain a session that feels like it is billing you for work you did not ask for. Hours later Tibo posted the three-item list: images in long compacted sessions, Computer History at high p95, and a conversation-title feature that was "draining a bit more usage than intended." He also said they had found a separate efficiency approach for the following week, and that paid subscriptions would get a full reset with the next day's patches.
Sunday, August 24. Tibo: reset propagated, some of those fixes landed, more the next day, "you should feel a positive difference." One Plus user reported running Sol Max for about half an hour and losing about 2% — versus a single Sol Medium task eating 3–5% before the patch. Treat that as one account, not a new rate card. Treat the direction as confirmation that the meter was wrong, not that Sol got cheaper as a model.
If you already spent the weekend fighting the bar, the sub2api fraud-flag post is still the first check when your account looks uniquely starved. This week's story is the opposite shape: lots of official Sign-in-With-ChatGPT users, same complaint, then a named bug list.
What Tibo actually blamed
Three product surfaces, not a mysterious Sol tax.
1. Images in long sessions with multiple compactions. Screenshots, UI captures, and diagrams in a thread that has already been compacted more than once. Compaction is lossy by design — explainx.ai covered that mechanic when Codex's 1M-token window started auto-summarizing older history. Re-encoding images across those summaries is the kind of cost that never shows up as "I asked one question." Practical move: new thread after a visual debugging binge. Do not keep pasting screenshots into a day-old session.
2. Computer History at high p95. Tibo did not publish a spec. The name matches Codex / ChatGPT computer-use keeping a history of desktop or browser sessions so the agent can resume. Tail latency and tail cost are the same family of bug: average users are fine, a slice of power users pay for a feature they did not know was on the meter. If you use computer-use or long family-IT setup loops, watch the bar after those sessions specifically.
3. Conversation titles. A background job that was supposed to name the thread, billing more than intended. This is the most embarrassing item and the most believable. Silent auxiliary calls are how auto-review panic threads start. Difference this time: OpenAI named it.
He also kept the cache-hit diagnosis from Saturday. Those stack. A title-gen call on a cache miss in a compacted image thread is four problems in one turn.
This is the same pattern as July: Sol's appetite was real, and so were the harness costs around it. The fix is still "meter the side work," not "hope the model uses fewer tokens."
The continue-after-zero quirk is closed — do not look for it
On August 21, a public thread described a long-running Codex /goal that kept going after the usage meter hit 0%. OpenAI did not document that as a feature. By August 24, users reported it no longer continues once you are out of quota.
Do not try to reproduce it. Do not build a tool around it. OpenAI closed a hole in the limiter. That is the whole story that belongs on this page.
GPT will skip a task and not tell you
Quota is half of this week's complaint. Completeness is the other half, and it does not show up on the usage bar.
In explainx.ai's own Codex and GPT testing, the model sometimes leaves items off a multi-step brief and never mentions the omission. You get a confident wrap-up. The original list still has an unchecked box. There is no "I skipped step 4 because of context" line. The session looks successful until you diff it against the prompt you actually sent.
That failure mode is worse in the exact sessions Tibo just patched: long, compacted, image-heavy threads where earlier constraints have been summarized away. A bigger window delays compaction; it does not make the model honest about what it dropped. The same gap is why trust, not generation, is the job once answers are cheap.
What we do in practice:
- Keep the source checklist outside the chat — a file, a ticket, a test list.
- Ask for a coverage table (requested vs done vs skipped) at the end, then verify it yourself.
- Start a fresh session when the work changes shape, instead of compacting a kitchen-sink thread.
- Do not treat Sol Max as a completeness upgrade. It is a longer worker. It can still omit.
If you forwarded that wrap-up to a teammate without checking, you were the meat proxy in the loop. The model did not owe them a status. You did.
What to do on your account today
- Look at the meter after the August 24 reset. Paid plans should show a hard refill. If you were waiting on the August 21 banked credit, this is a different lever — it should already be applied.
- Default to Medium unless the task is actually hard. Sol Max is for the jobs that were blocked, not for renaming a function.
- New session for images. Screenshots belong in a short thread, not a compacted epic.
- Sign in With ChatGPT on official or OSS clients you control. Shared subscription proxies are still unsupported.
- Verify the work, not the wrap-up. GPT will skip a bullet and smile.
Tibo said more efficiency work is coming that is "completely unrelated" to this week's three bugs. Until that ships, assume the meter can lie again, and assume the model can stay quiet about what it did not do.
Related on explainx.ai
- Codex 20M users — the banked reset that slipped
- Codex limits vs sub2api
- GPT-5.6 Sol usage: ~18% longer, 5-hour cap returns (Jul 29)
- Enable 1M context in Codex CLI — compaction tradeoffs
- Codex auto-review usage controversy
- ChatGPT Work vs Codex
- OpenAI $80 usage reset on Pro
- Loop engineering for coding agents
- Meat proxy: don't forward unread AI
Sources
- Tibo Sottiaux, August 23, 2026 — three drains + paid reset
- Tibo Sottiaux on X — August 22 cache-hit note and August 24 "reset has been propagated" follow-up
- Yash Thakker (@goyashy) — August 23 reply on still hitting limits faster
Quota language, reset scope, and the three named drains reflect Tibo Sottiaux's public posts through August 24, 2026. Confirm your own meter in Codex or ChatGPT Work. The continue-after-zero behavior is described from community reports, not an OpenAI changelog. Follow @explainx_ai for the next patch.