I was really happy with Codex a month ago on my Plus subscription, and I even recommended it to my friends because it was much more cost-effective than Claude Code.
However, since last week, I’ve noticed the token usage has increased dramatically. A simple task that takes only one or two minutes now consumes around 11% of my 5-hour quota, whereas the exact same type of task used much less just a few weeks ago.
At this point, it doesn’t feel much different from the free tier anymore. I’m wondering if something has changed recently, and if there are any ways to reduce token usage or make my quota last longer.
I tried GPT-5.6 Luna Light, which is supposed to be the cheapest model. However, for a very simple command like “Initialize the folder as a Git repository,” it consumed 5% of my 5-hour quota. Based on the rate card, GPT-5.6 Luna Light should use only about one-fifth the quota of GPT-5.6 Sol, but in my experience, it consumed almost the same amount. Is this expected, or could something else be affecting the quota usage?
By the way, I compared the same prompt in Claude Code and found that it only consumed about half the quota compared to Codex.
For the same task, I used Fable High, which consumed 32% of my 5-hour quota. However, in Codex, using GPT-5.6 Luna Light, the exact same task consumed 75% of my 5-hour quota.
This doesn’t seem right. Based on the expected usage, Luna Light should be significantly more efficient, so I’m wondering if there’s an issue with how the quota is being calculated.
OMG it is sooo expensive. I have the $100 pro account and the credits are dwindling away faster then I have ever seen before. Essentially, a $100 pro account today is no different then a plus account 1 year ago. I can’t throw numbers at you but this is how it feels.
At these level it would be impossible for individuals to run multi agent setup.
Regardless of what is being said, the cost of 5.6 sol high is far more then 5.5 high (when it came out).
I might have hit a limit 2 or 3 times running 3 consecutive agents for days with 5.5 high. Today, I hit my limit on the first day. This is a Pro account. This is insane.
Chat GPT 5.6 Sol (High) burned through my 5 hour usage and left me with 63% with a small session over night i did. I thought it was going to be comparible cheaper with regards to my awsome usage with 5.5. Whats happened Open AI? Have you got a bug in your usage in codex because it seems to be eating tokens like no tomorrow much faster.
I have the 5x Anthropic Plan and the 5x Open AI PRO plan but im having to use Open 4.8 (High to Max) / Fable 5 High since my token usage is lasting much longer. This is not what the token usage numbers tell us should be happening. In fact 5.5 used to provide me substantial more usage on these plans than anthropic did. Is it a bug in your usage metering?
GPT-5.6 drains so much tokens. I tried to change UI on some pages about layout, refresh UI a bit by using GPT-5.6 Sol extra high. I was surprised when the model did very detailed. And also got surprised when it hit the 5h quota limit after running over 2 hours. So in my opinion it should be increase more the 5h quota limit. I’m in $100 plan, it is not the most expensive plan, but it can only do one task per 5 hours even not using maximum effort. You guys can see too much agents spawn just for review.
Is this a glitch or is that how 5.6 is? I finished my weekly quota in a single research task, something I never faced with 5.5 xhigh(sol is also xhigh).
1 task, 56 minutes (a lot of waiting for codex during this task without doing any work, as zip’s where auto created etc), Terra High, 100% of the 5 hour limit used, 82% left in the weekly limit. Compared to 5.5 High this is extreme, and i usually don’t hit the 5h limit with continuous work of exactly this sort, working on the same app code. Not a nice surprise
I found the biggest reduction in token use came from changing the workflow, not changing models.
I stopped asking coding agents to take broad tasks and work autonomously until completion. That usually leads to repeated repository scans, speculative edits, regressions, repairs, and more repairs.
Instead, I use what I call mechanical coding.
The core principles are:
One bounded task per turn.
Inspect the current state before changing anything.
Define exactly which files may be modified.
Treat every other file as protected.
Back up files before mutation.
Let terminal output, tests, hashes, and generated artifacts determine whether the work passed.
Never accept “this should work” as validation.
Stop immediately when validation fails.
Repair only the exact proven failure.
Do not allow the agent to continue into the next task automatically.
Keep a written plan and a queue of small implementation tasks.
Start a fresh AI session when context becomes large, using a structured handoff instead of rebuilding context from memory.
Here’s a template to get you all started.
You are coordinating a controlled software development workflow.
Project:
[PROJECT NAME]
Repository:
[REPOSITORY PATH]
Current objective:
[ONE SMALL, PRECISE TASK]
Allowed files:
[EXACT FILES THAT MAY BE MODIFIED]
Protected files:
Everything not explicitly listed above.
Rules:
Inspect the current implementation before proposing changes.
Make the smallest change necessary.
Do not perform unrelated cleanup or refactoring.
Do not add dependencies unless explicitly authorised.
Back up every file before modifying it.
Run the relevant formatter, tests, static checks, and regression checks.
Terminal output is the source of truth.
Do not claim success unless validation proves it.
If validation fails, stop and report the exact failure.
Do not continue to another task.
Response format:
Return one copy and paste ready terminal command block.
The terminal output must clearly report:
RESULT=PASS or RESULT=INCOMPLETE
Files inspected
Files modified
Backups created
Validation commands run
Validation results
Next recommended task
There’s a major issue with the 5-hour limit—it’s too short, and it’s impossible to work this way. Please, OpenAI, do something about it; we spend all our time watching the reset timer, which isn’t viable for getting work done—and I have a Pro subscription.
Facing same issue since updating the app. 5.5 used to feel like unlimited for my tasks in Plus plan. However for similar tasks now 5 hour limits are getting consumed within a few prompts and half an hour.
There used to be a commercial where someone gets into a Ferrari, revs the engine, and realizes he has run out of gas just from revving it. That is exactly what using Ultra feels like on a Pro plan.
There is, at least implicitly, a community expectation that you can continue paying what you have been paying $100 per month for Pro in my case and, as the models improve over time, continue working more or less as you did before.
But that is not what is happening.
Simply put, the cost of using the frontier model has gone up, and not by a little. That is the betrayal.
Just as a restaurant is expected to remain consistent in what it serves, OpenAI has to remain consistent in what its subscription tiers actually mean. Telling me that the previous model is still available and cheaper is not part of the deal. The frontier is the deal. We are paying so that we can use the best model when it becomes available not so that, when a new model arrives, we are effectively pushed back onto the older one because meaningful use of the new model burns through the allowance at an absurd rate.
The increasingly disingenuous plan structure and opaque token accounting are devoid of integrity and steadily erode community trust. A subscription tier cannot retain the same name and price while its practical value deteriorates every time the underlying model improves.
The new model is amazing, and of course everyone wants to use it. That is precisely the point. You cannot market access to the frontier and then make the frontier functionally inaccessible to the very customers already paying for it.
Just like a hotel during a hurricane, this is a moment to do the right thing not exploit demand.
