Limit hit very early in GPT 5.5
Open 💬 16 comments Opened Apr 23, 2026 by SirTyson
💡 Likely answer: A maintainer (github-actions[bot], contributor)
responded on this thread — see the highlighted reply below.
What version of Codex CLI is running?
v0.124.0
What subscription do you have?
Bussiness
Which model were you using?
gpt 5.5
What platform is your computer?
_No response_
What terminal emulator and version are you using (if applicable)?
_No response_
What issue are you seeing?
I'm hitting limits very early in gpt 5.5 usage, even though /status shows that much of my limits remain available:
■ You've hit your usage limit. To get more access now, send a request to your admin or try again at 3:51 PM.
/status
╭─────────────────────────────────────────────────────────────────────────────────╮
│ >_ OpenAI Codex (v0.124.0) │
│ │
│ Account: garand@stellar.org (Business) │
│ Collaboration mode: Default │
│ Session: 019dbc0d-9976-73d3-8825-dc7f6617aeaf │
│ │
│ Context window: 68% left (90.7K used / 258K) │
│ 5h limit: [████████████████░░░░] 78% left (resets 15:51) │
│ Weekly limit: [██████████████░░░░░░] 68% left (resets 11:20 on 28 Apr) │
│ premium limit: │
╰─────────────────────────────────────────────────────────────────────────────────╯
What steps can reproduce the bug?
Uploaded thread: 019dbc0d-9976-73d3-8825-dc7f6617aeaf
What is the expected behavior?
_No response_
Additional information
_No response_
16 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
Yeah. 5h limit hit in a single prompt (24 minutes of work) on plus plan. This is using 5.5 medium.
This works out to 2.5h of work per week. Not really good enough - there is no value there so I may as well save the money.
Same here, I had ~65% weekly limit left and in a few prompts it says I hit my weekly limit.
Same here on Business, barely got a single change through today before it ran out of credits (on medium, fast mode off), also seemed alot worse (or ignoring agents.md) at following your prompts, asked it to work on specific version/part of our monorepo of something, it went and did code for some legacy parts in a wrong place, that it was asked specifically not to touch, one follow up, directing it to the specific path in the source (monorepo) and all credits was gone, by it fixing its mistakes alone.
Similar changes many times in the past, would have taken at most 10% of the 5h credits.
Same here. I have both Business and Plus accounts and both hit limits super fast in comparison to my earlier work with 5.4.
even two plus subs are cooked on 5.5 low due to 5h limit, insane
This is absurd.
I just ran 3 short simple prompts on 5.4 Medium and maxed out the $20/month plan, what on earth...
If openai wants to destroy the value of chinese models they should make new models that are more than usable on plus plans while being better than 5.3-codex (and chinese models) in either speed or intelligence for coding. Currently there is no progress for plus plans past 5.3-codex because 5.4 has alzheimer and 5.5 drains 5h limit from a single prompt (are we claude yet?)
All this while sama is talking about "lots of compute" and efficiency
I'm pretty sure the issue is related to this write_stdin polling that 5.5 is very aggressive with:
https://github.com/openai/codex/issues/13733#issuecomment-4322785225
@jusava I think it's just the input pricing that "reasoning efficiency" doesn't even out at all (unlike output pricing)
I'm not that inclined to believe it is explained by reasoning efficiency. For some some tasks your reasoning tokens are within some 10k token range, whereas cached input token ranges are in millions. Maybe higher reasoning effort results in more turns, and that consumes more cached tokens. But if you were to look at it just from basic reasoning token count perspective, that's kinda irrelevant. It really is the total turn count (aka how many times you submit full context), that seems to be driving this. Whether this be model feature or harness feature and what changes are more impactful is hard to say. But I'm seeing 5.5 now costing me net 4-5x in either outright api costs or token quotas as running similar tasks in gpt 5.4 (again not sure how much harness updates contribute to this).
This goes entirely too fast.. Basically trying to force everyone to go to the higher plan. I used to do 5x the work in one day vs what I am able to do now. And even when it drifts off or hallucinates, it says "Yeah, that was my bad" and charges you again for the work it was supposed to do correctly the first time.
@jusava this is exactly what I am saying
it's the insane input pricing
reasoning efficiency evens out output pricing increase, but doesn't even out input pricing increase compared to 5.3
and they are telling us their pricing is on par
100%. I bought £10 of credits. Asked a few simple queries and poof the credit was gone. Even on medium intelligence it seems to race through credits. In comparison, another tool seems to go much further for less. :/ Shame, I like the experience overall.
This issue is the main reason why most devs are switching to chinese models. The current feasible & affordable fix for this is to override the codex API in
~/.codex/config.tomlto other API (e.g. deepseek). reference: https://zeabur.com/en-US/blogs/use-codex-with-ai-hub