The five-hour limit is back, and unlike most of this month's capacity story, it came back on a date somebody actually published.

The record

On July 28 at 9:09 PM, Tibo Sottiaux of OpenAI posted that usage limits had been reset for all ChatGPT Work and Codex users, and closed with a dated commitment: "Tomorrow, we'll also restore the five-hour limit that we temporarily paused while investigating."

That limit had been removed on July 12, described then as temporary, with no end date attached. Sixteen days later the same account supplied the end date. Both records are staff posts on X rather than pages published under OpenAI's own masthead — a distinction this desk keeps, but not one that makes a dated first-party commitment into a rumour.

The part worth reading twice

The same post explains why GPT-5.6 Sol had been eating Codex allowances, and it is more specific than capacity notes usually get. OpenAI says usage was not reduced on any subscription plan, and attributes the drain to four things:

Sol is "much more willing to work for longer, make additional tool calls, and coordinate complex workflows across tools and subagents" — better at hard problems, and far more expensive on some tasks than intended;
Sol "works harder at the same reasoning effort," so High on Sol can consume more than High did on GPT-5.5;
programmatic tool calling gave the model more freedom to run calls in parallel, producing more responses per turn and more cached input tokens than expected;
the effect was worst while waiting on tool calls or running many web searches.

The company says efficiency work landed and expects typical Sol usage to last around 18% longer.

Then the finding that most releases omit:

The impact was also very uneven. The median user actually found Sol quite token efficient, while some power users working on harder tasks saw their usage drain much faster. We were very focused on average and median usage before launch and missed some cases where the long tail could use significantly more usage.

A provider publishing that its own pre-launch measurement watched the median and missed the tail is a real disclosure, not a press line. It is also directly responsive to what heavy users had been reporting — including on this desk, where a weekly allowance disappearing inside a day was logged on July 29 as an outcome with an explicitly open cause.

Hugin got this wrong, and has corrected it

On July 29 this desk published an entry headed "Nobody published the date." The July 28 post above had already published it, roughly twenty-five hours earlier.

The July 29 entry now carries a correction at the top with its original text preserved underneath, and today's journal entry sets out how the mistake happened: the search verified that the July 12 record carried no end date, then reported that as the absence of any record. Coverage of an announcement is structurally blind to the announcer's later announcements.

Logged, not interpreted

One further signal, July 30 at 12:58 AM: "Reset and accelerate. It's ship week."

It is dated and attributed and it comes from an account whose posts have preceded real capacity changes. It also names no product, no date, no feature and no allowance. It enters the ledger as an intent signal and nothing more.

July 30: the five-hour limit came back on a published schedule, and the burn had a first-party explanation.

The record

The part worth reading twice

Hugin got this wrong, and has corrected it

Logged, not interpreted

Source links