I had 91% of my Codex usage left when Tibo Sottiaux announced another full reset.

My first reaction was simple: turn on Ultra, use /fast, and queue the expensive tasks.

I posted that on X. The replies quickly turned into a quota planning session. Some builders wanted to burn everything before the reset. Others were already comparing effort levels and usage speed.

That reaction says a lot about how people use Codex now. Quota is no longer an account setting you check once a week. It affects which model you choose, how many agents you run, and when you start a large task.

What OpenAI Found

Some Codex users reported faster usage drain this week. Tibo first pointed to lower cache hit rates for affected users. When Codex misses the cache, it has to process more of the same context again.

The team later named three more problems:

  • Images inside long sessions with multiple compactions created inefficiencies.
  • Computer History produced high usage at the p95 and above.
  • A conversation-title feature consumed more usage than intended.

OpenAI plans to ship fixes on August 24. Tibo also said the team found another way to improve efficiency and will work on it next week.

Paid subscriptions will receive a full usage reset around 2 PM Pacific on August 24, 2026.

Tibo called this a full reset. He did not call it another banked reset.

What Builders Said

My first post said I still had 91% left. One reader had less:

"45% left 😂"

@tongzhenbao

Others had already decided what to do:

"Gonna use it all"

@notjazii

"time to max out fast mode"

@0xRemakw

One builder recommended a slower burn:

"Nah.. just use medium or high non-stop. That should do. Right now, the token usage goes down fast."

@tinkerersanky

Another reader asked why OpenAI would not send this as a banked reset. That would let users save the extra capacity instead of receiving it at a fixed time.

One user also reported that their Sol Ultra workflow with three Luna sub-agents was using about one percent per hour. That is a useful early observation, but it is not confirmation that the fixes have landed for everyone.

What I Would Do Before the Reset

I would not burn quota for the sake of seeing the meter move.

I would queue work that benefits from more reasoning or parallel execution:

  • Architecture reviews
  • Large refactors
  • Multi-agent research
  • Full codebase audits

Ultra and /fast make sense when the task can use the extra capacity. File discovery, summaries, and narrow fixes can stay on Luna or a lower effort level.

I would also avoid putting many screenshots into one giant session until the fixes land. Start a fresh session after several compactions, especially when the conversation contains many images. OpenAI named that pattern as one source of wasted usage.

After the reset, I will compare usage by completed task rather than by time. A full meter tells me nothing about efficiency. I want to know whether the same coding task consumes less quota than it did this week.

The reset gives builders another window for demanding work. The usage curve after the fixes will show whether Codex feels predictable again.

Sources