OpenAI recently brought back the 5-hour usage limit for Codex on Plus, and after actually running into it today, I have to say: this is awful.
Not because I expect unlimited access. I don't.
I completely understand that these models are expensive to run and there need to be usage limits somewhere.
The problem is that the 5-hour window makes it incredibly difficult to actually use Codex for the thing it is designed to do: sit down and work on software.
Today I gave Codex one fairly simple task in my Markdown notes app.
Add Find and Replace.
That was it.
I wanted Find in the Edit menu, keyboard shortcuts, match navigation, Replace, Replace All, and the usual behavior you would expect from a basic editor.
Codex implemented most of it, wrote tests, ran the build and linting, and started doing browser verification of the actual UI.
Then my entire "5-hour" limit was gone.
Actual elapsed time: about 24 minutes.
One task.
Twenty-four minutes.
Five-hour allowance gone.
And the best part? It didn't even completely finish.
The search highlighting still isn't finished, and Codex hit the usage limit while it was wrapping up verification. Now I have an implementation that is almost done, Codex knows what it changed, Codex knows what it was checking, and I can't even tell the agent to finish its own work until my limit resets.
That is the part that makes this system feel fundamentally broken.
Obviously I don't expect "5 hours" to literally mean five hours of wall-clock time.
Agentic coding usage depends on context size, reasoning, tool calls, tests, commands, and a bunch of other factors.
That's fine.
But there is a pretty enormous difference between "this might not last exactly five hours" and "one normal coding task can consume the entire allowance in less than half an hour."
At that point, calling it a 5-hour limit is almost meaningless from the user's perspective.
I can't reasonably plan around it.
I don't know whether opening Codex means I have an afternoon of coding available or one 20-minute task.
And that uncertainty completely changes how useful the product is.
For a short time, Plus users effectively had the weekly limit without this restrictive 5-hour window getting in the way. I actually preferred that.
I could sit down on a Saturday or Sunday and work.
Maybe I burned through a large chunk of my weekly allowance in one weekend. Fine. That's my choice.
If I use my week's allocation in two days, I understand that I'm not going to have much left for the rest of the week. But at least I could decide when I wanted to be productive.
The current system gives me a weekly allowance and then prevents me from actually using it when I want to.
That's the frustrating part.
I can still have plenty of weekly usage remaining and be locked out because one agentic task demolished the smaller 5-hour bucket.
So what exactly is the weekly allowance accomplishing at that point?
Software development is not evenly distributed throughout the week.
Sometimes I barely touch a project for several days.
Then I have an afternoon free and want to spend four or five hours working through features, bugs, refactoring, and testing.
That is exactly where an agent like Codex is useful.
You give it a task. It explores the codebase. It implements something. It tests it. Maybe it finds a bug. It fixes the bug. It runs the browser. It verifies the behavior.
That process can be expensive, but that's also the entire point of using an agent instead of a chatbot that spits out a code snippet.
If the more thoroughly Codex works, the more likely it is to lock me out halfway through a coding session, we're creating a really weird incentive. I don't want it to skip tests to save my quota.
I don't want it to avoid inspecting the codebase.
I don't want it to stop verifying its changes.
Those are some of the biggest reasons I use a coding agent in the first place.
Looking through the OpenAI community, there are already multiple Plus users describing almost exactly the same experience.
People are reporting ordinary development tasks consuming most or all of the 5-hour allowance while they still have significant weekly usage remaining. Some have also had Codex stop before the work was finished.
That's pretty telling.
This isn't someone trying to run an autonomous agent against a million-line repository for eight hours straight.
These are normal coding sessions.
Interestingly, OpenAI's current documentation says usage depends on things like the model, task complexity, context, reasoning, speed, and tools, which makes sense. It also says an active Codex turn may continue after reaching a limit, subject to fair-use restrictions.
That certainly wasn't a very useful distinction in my case. My session still ended with a usage-limit error before the overall implementation was where I wanted it.
I am completely fine with having limits on Plus.
Give me a weekly compute allowance.
Give me a generous rolling limit designed to prevent extreme abuse.
Throttle extraordinarily expensive workloads if necessary.
Whatever.
But the current setup, where one completely reasonable development task can consume an entire 5-hour allowance in about 24 minutes, is not a useful coding workflow.
If I still have weekly quota available, I would much rather be allowed to decide when I use it. Maybe I blow through the whole thing in one marathon weekend.
That's okay.
At least I got something done.
Right now I have most of a Find and Replace feature, missing search highlighting, an agent that stopped before completely finishing what it was doing, and a several-hour timeout before I can ask it to continue.
That's not really a productivity tool anymore.
It's a productivity tool with a mandatory intermission halfway through and to me that makes it almost useless.