OpenAI restores ChatGPT Work / Codex 5-hour limits after fixing a GPT-5.6 Sol token bug

The rate limit holiday ends. OpenAI turns the 5-hour caps back on 29 July after patching a GPT-5.6 Sol overconsumption bug that had been running since 12 July - with a claimed 18% net capacity uplift.

The fact

OpenAI's Tibault Sottio said on 28 July that the temporary lifting of the 5-hour usage limits on ChatGPT Work and Codex - in place since 12 July for paid subscribers - would end at 05:00 on 29 July 2026. The cap was lifted in the first place because GPT-5.6 Sol was over-consuming tokens on certain task types, and OpenAI opted to eat the abuse rather than degrade the paying experience. After a series of fixes, Sottio says the effective available time should now rise by "about 18%" versus the pre-holiday baseline, and per-user rate counters have been reset.

Our read

The story is one paragraph long in the trade press, but it says two things worth marking. First, this is the second time in six weeks OpenAI has visibly lost cost control on a shipped model - precisely what our token-budget-caps thread has been tracking. Second, the fix pattern - pause the meter, patch, resume with a small capacity bump - is being normalized. It is a soft admission that per-user token economics on frontier models remain volatile even in production, and that the "unlimited during incident" default is now part of the SLA rhythm for premium seats.

What to watch

Whether OpenAI publishes the specific Sol regression (task class, average tokens/turn) - currently opaque. And whether Anthropic sees a mirror bug in Fable 5 as it moves into Enterprise plans (thread anthropic-fable5-shipping). Two data points make a pattern.