Users kept reporting that Codex usage was dropping faster than usual. In response, Tibo Sottiaux shared the causes and a plan of action on X.
There was more than one confirmed cause. Three things were named together: the process of compressing context several times while handling images in long sessions, Computer History's high p95 usage, and the feature that generates conversation titles.
Cache efficiency wobbled with images and long sessions
The post explains that there were inefficiencies in long sessions that included images and in repeated compaction. Compaction is the process of compressing the key content of a long conversation to carry it into the next task.
This doesn't mean the price of a single image went up in a fixed way. It's an observation that usage accounting ran less efficiently when images were part of long tasks and repeated compression.
This also ties in with OpenAI's help article, which explains that usage varies with the size and complexity of the work.
Computer History and conversation title generation had an effect too
The second cause was Computer History. The post says p95+ usage was high in that feature. p95 is a metric that looks at the heavy upper end of observed values, so it doesn't mean the average usage of all sessions.
The third is the conversation title generation feature. It was originally a helper for telling conversations apart, but it was consuming more usage than expected. It's an example of how even a feature that looks small can affect total usage when it's called repeatedly.
The explanation that Computer History is a feature that records interaction events from selected apps and websites can be found in the ChatGPT Business release notes.
When looking at usage issues, separate session conditions first
Even with the same model, results can differ depending on whether the session is long, whether it includes images, whether context compression repeated, and whether Computer History is on. If you record only the model name when comparing usage, it's easy to miss the cause.
In practice, it helps to split work into shorter units and record whether images were used and how long the session was. If you can see usage by feature, also look separately at calls that repeat automatically, such as title generation and history features.
A full reset is different from a fix
Sottiaux said that along with the fix, usage would be reset across all paid subscriptions. This is a measure that corrects limits already consumed, and should be kept separate from meaning that the cause of usage continuing to shrink has been removed.
OpenAI's help article says Codex usage varies with the size and complexity of the task, the model, and where it runs. So even after the reset, you should compare changes in usage under the same task conditions.
The point of this case isn't simply raising limits. It's that you should treat usage not as a fixed model price but as an operating cost created by sessions, features, and context management.