OpenAI Just Reminded Us Why We Can't Trust Codex Limits

OpenAI removed Codex's 5-hour usage limit on July 12, 2026. For about two weeks, heavy users like me got a reprieve — no forced stops mid-session, no watching the counter tick toward zero while you were still in the middle of something that mattered.

Then I saw the clarification post on the OpenAI community forum.

"...the five-hour limit display was temporarily removed as part of an incident response and should return once the incident is resolved."

That word "temporarily" is doing all the work. And it's the most important thing to understand about this whole situation.

What Actually Happened

Here's the timeline nobody's talking about clearly:

April 2026: OpenAI introduces the 5-hour rolling window limit for Codex and ChatGPT. Plus users get roughly 40 minutes of actual reasoning time per window. Pro users get more, but the ratio is similar.

April–June: Users immediately start reporting the limits are too restrictive. Real coding sessions burn through the window in under two hours. Heavy users are hitting the wall mid-task, not at the start of one.

June 28–29: OpenAI discovers Codex usage is being consumed faster than expected. Multiple emergency resets. Full usage banks reset for all users. Tibo Sottiaux (OpenAI product lead) posts that they've "fully reset usage limits again for all users" and "added more detailed monitoring."

July 12: OpenAI temporarily removes the 5-hour window limit entirely for all Plus, Business, and Pro plans. Usage limits are reset across the board. Everyone mid-cycle gets a fresh runway.

July 21–25: OpenAI clarifies: removing the limit was "part of an incident response." It will return.

That's the story. OpenAI had a metering problem, the limits were broken, so they took them off. They called it temporary. It was always temporary.

Why They Removed It (And Why It Wasn't Generosity)

The 5-hour limit was introduced alongside a new pricing model in April — replacing the old unlimited-within-plan model with a credit-based token system. The idea was to make costs predictable. The reality was:

• A single complex coding task could burn 5–7% of your weekly allowance in one shot
• The "5 hours" was measured in wall-clock time, not actual reasoning time
• OpenAI's own tracking was inconsistent — usage was consumed faster than expected, multiple times, for reasons nobody fully explained

The limit removal in July wasn't OpenAI being customer-friendly. It was OpenAI fixing a broken incident. When the incident is resolved, the limits come back. That was always the plan.

What the Return Actually Means for Heavy Users

If you're using Codex casually — a few queries a day, nothing intensive — the limits returning probably doesn't affect you much. You'll hit them occasionally, buy some credits, move on.

If you're using Codex the way I am — heavy, daily, across multiple projects — the return of the 5-hour window changes your workflow in a specific way I hadn't fully appreciated until I thought about it.

You research before you build. That's the workflow that actually works with AI-assisted coding: you do the research, understand the problem, figure out your approach, then hand it to the AI to execute. The research phase is where you figure out what you're actually trying to build.

The 5-hour window limits how much research you can do before each session reset. Right now, without the limit, I can research as long as I need to before hitting the build phase. When the limit comes back, the window shrinks again. I can do my research — but then I have less runway left for the actual execution phase when the model matters most.

And the reset timing? Nobody knows exactly when. The community forum posts suggest it could be soon. Maybe even before the end of July 2026. OpenAI has given no specific date.

Why They Might Remove It Again

Here's the part that actually gives me some optimism — and some anxiety.

The removal happened because the limits were broken. The metering didn't work. OpenAI had to take them off to stop the bleeding.

When they put them back, they're putting back a system that has already failed twice. If it fails again — if users start reporting that usage is still being consumed faster than expected, if the resets keep happening — OpenAI will have to choose between fixing it properly and just removing the limit again.

My take: they know heavy users are watching. They know the community forum has threads with 10,000+ views on the "5 hour limit feels misleading" post. If the limits come back and immediately break again, they will have to act. The reputational cost of a third incident is higher than the compute cost of another removal.

That means: enjoy the window while it's open. Use it heavily. Document what you're doing. If they take the limits off again, you want to be in a position to use that runway immediately — not scramble to figure out your workflow.

The Signal in "Temporarily"

I keep coming back to that word. OpenAI didn't say "we're rethinking the limits." They didn't say "we're redesigning the pricing model." They said "temporarily removed as part of an incident response."

That means the underlying model — rolling window, token metering, the April 2026 pricing structure — is still the plan. "Temporary" means they're going to put the same broken system back in place once they think it's fixed.

It might hold this time. It probably won't.

If you depend on Codex heavily, plan for the limits to return. And plan for the possibility that they might not hold, and you might get another window — whether that's weeks from now or months.

The one thing you shouldn't do is assume the reprieve is permanent. OpenAI's own words tell you it isn't.

Use the window.