Confused about Codex limit usage after GPT-5.6 release (Plus plan)

Codex
1

Hi everyone,

I still can’t understand why Codex limits are being used up so quickly. I’m on the $20 Plus plan, and since the GPT-5.6 family was released, the amount of work I can get done within my limits has dropped significantly.

Before GPT-5.6, I used Codex heavily and could usually work with GPT-5.5 Medium for a full day while using about half of my five-hour limit. I also used Codex on another account for a long time and almost never hit the five-hour limit or even came close to using half of the weekly limit.

When GPT-5.6 Sol came out, I decided to try it. Since it was presented as a more efficient model, I selected Medium effort and asked it to make a small fix of about 40 lines of code.

That one task used around 15% of my five-hour limit.

After a few requests, I switched back to GPT-5.5. But tasks that previously used about 4-5% of the five-hour limit were suddenly taking around 9-12%.

I’ve tried different models and settings: GPT-5.5, Sol Medium, Sol Light, Terra Medium, and others. For tasks of similar complexity, I haven’t seen any meaningful difference in limit usage.

I also saw Tibo’s messages saying that the limits themselves had not changed, but Sol uses more tools and has a higher so-called “juice,” so it may consume more of the limit. There was also information about an optimization that was supposed to reduce usage by around 18%.

I understand why a newer model might consume more because of its higher juice and additional actions. What I still don’t understand is why GPT-5.5 now uses more as well, even though it used to handle the same tasks much more cheaply.

My main question is: which model should a Plus user choose to get quality roughly comparable to GPT-5.5 Medium without using up the entire weekly limit in just a few working days?

Is the current level of limit usage expected after the release of GPT-5.6, or has something changed in how usage is calculated?

Thanks to everyone who read this, and special thanks to Tibo for the limit resets.

2

A few more details and examples might help explain why I have these questions.

Luna test

After Luna and Terra became cheaper, I decided to test Luna on a pet project. I used Max and Extra High because the published charts suggested that these modes were roughly on par with Sol Light or GPT-5.5 Medium.

The first larger run took about 25 minutes. The model wrote around 2,500 lines of code and ran tests many times. It used about 5% of the weekly limit.

I split the rest of the development into smaller stages. On several of them, Luna used around 2-6% of the weekly limit per task.

I also used Sol Light to review the changes. It found many issues and used around 4% of the weekly limit. GPT-5.5 Medium then fixed those issues for another roughly 5%.

So even with the cheaper Luna model, the weekly limit was still being used up fairly quickly.

Comparison with Opus

Later, I tried Opus 5 High through a friend’s account. I liked the implementation quality less. In my opinion, Sol follows the code style better and more often suggests solutions that fit the project.

Still, implementing one more stage, making two fixes, and looking for solutions for the next substage used about 35% of the five-hour limit and around 4-5% of the weekly limit.

I understand that limits across different services cannot be compared directly, but the difference from Codex’s current usage feels very noticeable.

My observations on the models

After looking through the available benchmarks, this is how the models appear to compare.

GPT-5.5

After the new models came out, GPT-5.5 looks less cost-effective. Based on the benchmarks, Sol Light performs better than GPT-5.5 Medium at a lower theoretical cost. In practice, though, their limit usage now seems similar.

GPT-5.6 Luna

After the price reduction, Luna looks quite interesting even at Max effort. Its quality may be below Sol Light, but the results are still impressive.

GPT-5.6 Terra

I find this model difficult to place.

Terra High is roughly comparable to GPT-5.5 Medium, or slightly better, but it is hard to justify in terms of cost when Sol Light can produce similar or better results.

Terra Medium is roughly in the same range as Sol Light, yet it still does not look better than GPT-5.5 Medium.

Terra Extra High is slightly better than GPT-5.5 Medium, but Sol Medium seems like a better value.

GPT-5.6 Sol

In terms of quality, Sol is clearly better than GPT-5.5. But its actual limit usage is either similar to GPT-5.5 now or much higher than GPT-5.5 was before GPT-5.6.

That is why I still do not understand which model is meant to be the economical replacement for Plus users.

Sorry for the long post. I wanted to explain the situation in enough detail to make the issue clear.

3

This post was extremely helpful and detailed. Thank you for sharing your testing and breakdown of the models!

5

Has anyone else noticed that Codex usage limits have started draining much faster over the last few days?

I’ve been using GPT-5.6 Sol with High reasoning for a while, so I already know that it’s an expensive model. Previously, I could use roughly one weekly limit over the course of a full day of fairly intensive work. Now I’ve burned through the limits on two accounts in about half a day.

The tasks are approximately the same as before: similar projects, similar amount of context, and the same general workflow. I understand that token usage can vary by 20–30%, since no two tasks are exactly identical, but this feels like a much larger difference.

Fast mode is not the explanation. I was already using Sol High before, and I’m comparing it with the same Sol High setup now.

It almost feels like something changed behind the scenes — perhaps the reasoning-token budget, prompt caching, context accounting, or the multiplier used for Sol. OpenAI announced reduced consumption for Terra and Luna, but I’m wondering whether Sol effectively became more expensive at the same time, even if its published rate did not change.

I’m not claiming this is definitely intentional. It could also be a bug or a recent rollout affecting usage accounting. But seeing the same sudden increase across two accounts makes normal task-to-task variation seem unlikely.

6

OpenAI questions

Why don’t new models even understand what a limit on a single task is?

Why do new models launch subagent without a preliminary cost assessment?
Why aren’t repeated checks recognized as duplicates? With minimal point edits, the model re‑runs checks that have already been completed, and no system prompts stop it from doing so.
Why can the model independently expand the scope? Without objective reasons?
Why, in fact, is it claimed that the cost of the models is being reduced, but the expenses have increased many times over? Are the same code limits simply disappearing? It was better even a week ago.

GPT Sol’s advice is to leave it alone. Well, this is just perfect material for my YouTube review: how GPT Sol wastes new meager limits, while cheap Chinese models perform exactly the same task, making good use of the money invested in them.

Sol admitted to being helpless:

-The screenshot confirms this is no longer just “spent a little.” I consumed 20% of the weekly resource on a routine analysis and made several more incorrect corrections to my own mistake. I stop and no longer spend the OpenAI limit on this topic; switching to the Chinese route here looks quite rational.

-So what, can you implement a proper solution for me so that it works and you don’t keep doing this kind of nonsense?

-Prompt rules no longer provide physical blocking: real protection requires a runtime guard right before delegate_task, plus focused tests. Right now, it’s more sensible not to waste the remaining resources on a risky half‑hearted implementation, but to switch to the Chinese route and finish the guard there.