codex is really too slow
Open 💬 35 comments Opened May 7, 2026 by s7monk
💡 Likely answer: A maintainer (github-actions[bot], contributor)
responded on this thread — see the highlighted reply below.
What version of the Codex App are you using (From “About Codex” dialog)?
26.429.61741
What subscription do you have?
pro
What platform is your computer?
win
What issue are you seeing?
codex is really too slow, whether it's the vs code plugin or the codex app, the model responses are way too slow, the model being used is gpt 5.5 fast, it's really slow, can't compare to claude code at all
What steps can reproduce the bug?
codex is really too slow, whether it's the vs code plugin or the codex app, the model responses are way too slow, the model being used is gpt 5.5 fast, it's really slow, can't compare to claude code at all
What is the expected behavior?
_No response_
Additional information
_No response_
35 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
Sorry for asking? But whats the codebase size? do you notice it stalling on a certain topic ? does it repeat itself?
does it take too many steps ? or it is per say slowness?
I can confirm this issue.
I was using Codex app fine until approximately yesterday, when switching between threads became very slow, starting new thread is also slow. When I write something simple to model (I have session "ping" where I'm just checking if it's available since I have internet issues), so when I write "ping" in my simple session, model responses, but when I'm entering session with long context and write something, it start reconnecting.
But when not reconnecting, Codex app just freezes when I write something and model starts to think and answer. In this moment app just freezes, like windows shows that popup "app is not responding" for nearly 10-20 seconds. After waiting, app unfreezes and model answers.
I have such issue only on windows, on mac Codex app works fine.
So I dunno if it's because some of Codex updates or what, but on windows it started working really slow and with freezes nearly 2 days ago
My perception is that it's purely slow. The codebase isn't large, but it responds very slowly to questions. Even for what I consider simple problems, using fast mode still takes a long time to complete tasks.
you think you can share prompts you are having issues with ?
did you had issues like in the computer-use area?
like "go make me a postgress database, migrate my code "
(would be cool if you are working on a open source project , and can share which prompt casuing issues to you)
Have you tried using the plan-mode (ctrl + shift . or ctrl+shift+p)
https://developers.openai.com/codex/learn/best-practices
alkhamdulilah , thank you
I can confirm the issue. I don't know what happened, but even with the fast mode, it is extremely slow and it doesn't matter which model you actually selected. I tested all models 5.3 codex 5.2,5.5..(not mini one..). I don't know what OpenAI is doing, but it wasn't like this before!
I can confirm the issue too, with or without fast mode, 5.4 and 5.5 extra high used during 5h hours, rate remaining for the 5 hours window : 85% ... pro subscription
The same, GPT 5.5 fast high effort become EXTREMELY slow, not usable. $200 for waiting all day for one simple task. Something is deeply wrong.
a. please use the /feedback option within codex
aa. Can you replicate the issue on some open source project and share the prompt?
ab. you think making a new chat , and specifing the issue/ prompt (or even tell the ai the end goal you try to achieve)
high-effort
gpt5.5-x-high
is extremely over kill , but everyone and their use case , have you tried expirementing with 5.5 medium
from personal expirence, I usally scream at the AI , all day, not phrasing what I want correctly
b. sometimes slowness , can be happen after chatting a lot , and code-base increase
c. the way I feel , ai right now has two issues
Whats your simple task? Does it include changing many files at once?
Its not what we seen with gpt5.2 and lower codex models , a lot has improvement
@Mahkhmood9
Still not sure how slow work of Codex app and Earth's gravity are connected
Adding a more precise update from my side, including the full workaround that fixed it locally.
The issue was present across several recent Codex Windows app versions, up to
26.506.31421.Tests performed while keeping my existing local user data/configuration:
The issue did not look like a project-size problem, rate-limit problem, or model-only problem.
What finally fixed it was not just reinstalling. The full workaround was:
codex-26-422-8496-0.msix/ app version26.422.8496.0.prompt-historywas stored as an object, while the rolled-back app expected it to be an array.~/.codex/.codex-global-state.jsonMore specifically, under:
electron-persisted-atom-state.prompt-historyI converted/cleared the value so it matched what the rolled-back app expected: an array.
node,npm, andnpxwere resolving through broken mise shims that depended on a bad/missingNODE_BASE.I realigned WSL so that:
noderesolved correctlynpmresolved correctlynpxresolved correctlycodexCLI still resolved correctlyNODE_BASEdependency was removed/unsetAfter those steps, the same user data, same conversations, same config, and same projects worked again on
26.422.8496.0.So the workaround that fixed it locally was:
26.422.8496.0prompt-historystateHow does anyone work with Codex? It’s extremely slow and almost impossible to use productively
I can confirm the issue too. It would be great if you can take a look and fix it. Thanx!
@davlola90
I worked with it all the time , it gotten a bit slower , but honestly I use the free tier--- and barely hit ratelimits --- alltho I do mostly use the gpt 5.4mini , I really dont care about 5.5 --- 5.5 can understnad you better,,, but you if you format your langauge and speak to 5.4mini , its just as productive
@Toufik1247 I do understand where you comming from ... and personally there is an appearance like "codex gets slower with bigger chats",, restarting the chat (click "Start New Chat" ) and then asking what I want it to do and It does appear to stall after a while
`
this is what I want to do [Here Is mention a function or class or file@ ] ,,, [mention the functionality I want]``and codex really magically does it..
I do understand where it can feel slow ,,, like when it appears to stop but is actually writing and thinking about making 900-3000, lines code change .... but the model actually preforming in a very healthy way considering all of whats happening in the background its very healthy --alltho compute is an issue its a crisis -- and indursy still face the problem of having to queue many chats into a gpu... Yet server uptime seems to be 95% sla ,,, no kernel crashes no nothing
Unrelated mumbeling
Lets look at it this way,, codex always completed my prompts - I made over 4000 prompts (part of the expense is me the user slowly learning how to correctly use the technology---- but now the dataset are fully synthetic --- aka merging users chat ,,, like 1 big chat ,, the user made 10 prompts ,,, took the user a lot of time to get it working ---- future ai models know to inject ideas from when it started working ,,, amazing using human language and lingusitic logic).... which is could the next generation of ai models would already know what the user would ask in the future ... Making things needing 1 prompt,, and the user doesnt know why his 1st prompt was bad.... "But the AI got the VIBE".... which is very interesting , how the magic for the people who now need to build AI also now to think about philisophy --- should the cat fly ,, and we pretend its a tiger .... now we ran out of human dataset and the dataset .... and llm chatlogs dataset insnt enough.... in some way its a statstical issue ,,, ""how likely the user is to format his prompt correctly?? should the model inject ideas that help it in the future [eg this function is crashing... but it worked only 19 shots later] How does the model learns to memorize and predict the future,, like a human does --- oh I seen this issue before --- I know which website or documentation to use"" but in many ways it can do it now ,,, like very rare programming languages and it just need documetation refrence ( and why coding agents,,, treat code in a stastical cluster of words) ---- and to be fair a lot of expirements are being done and a lot of things likely proven to be very very good in a short time ... which is beautiful,.... for now the model keeps doing things like "Oh lets see, I tried this apporch, now next I do this and this,,, " which is all and good, but like the idea of injecting the future --- in the synthetic dataset way of things
But the model is fast...
a. it appears to be slow , but its actually healthy - but to be fair I also notice it a bit slower --- you can see on chatgpt in the past that the response can slowly generate --- we seen for example chatgpt also being slow-er
b. My personal opnion codebase size --- or you are in a big task
c. you could be using extra-high ,, which is slow for everyone --- personally its pointless the extra-high its totally meaning less for me atleast
@Mahkhmood9 thanks for the detailed reply. I understand your point about long chats, large codebases, and
xhighmaking Codex appear slower while it is actually thinking or preparing a large patch.But my case seems different.
I am on the $200/month plan, not the free tier. The issue happened across the recent Codex Windows versions I mentioned earlier, up to
26.506.31421.I reproduced it across 3 different conversations:
With the recent version(s), after intensive usage from about 8 AM to 6 PM:
That was not normal from a paid-plan/user-experience perspective, because the app was freezing/lagging heavily while the usage window was barely being consumed.
After downgrading to
26.422.8496.0, while keeping the exact same local Codex config, projects, workspace setup, and environment:So the behavior changed massively with the Codex app/runtime version, not with my local config, project size, chat size, or prompting style.
I agree that model choice and reasoning level can matter. But this looks more like a regression in the desktop/runtime behavior: the newer version feels stalled/frozen and barely consumes the usage window, while the downgraded version actually runs work normally under the same conditions.
same issue, GPT pro on 100$.
it might be just OpenAI pulling a slow down rug scheme to users instead of a faster token consumption to differentiate the rug pull from anthropic.
I’ve been clearing cache, deleting the .codex folder, reading through forums and trying every suggested fix. It always comes back to the same state — extremely slow, almost unusable.
Tasks that used to complete almost instantly now take ages. I currently have a single process running for over 1h 30min and it still hasn’t finished.
<img width="192" height="47" alt="Image" src="https://github.com/user-attachments/assets/64cce5e9-fcf1-4b5e-aa63-462470b38407" />
Is there any way to downgrade the Codex VS Code extension to an older version?
In the meantime, I’m speaking with support about a refund, because I’m really not happy with the level of service recently.
same for me on 5.5 medium
related (?): on startup I get:
I contacted OpenAI support, raised a claim. One of the things they shared is below:
I actually did some of these steps, and it finally works for me as it was earlier, fast !
rm -rf ~/.codexand thenrm -rf ~/Library/Application\ Support/Code/User/workspaceStorage/*) - this removes installed VS Code extensions as well, projects/code remain untouched, need to reinstall extensions afterwardYes it seems to have some ratemiting has increased to be more often? (But its still very generous , and useful)
as per slowness, clearly there is a lot of server load
Its not perfect by any mean, I personally do not like the 5.5 release ,mostly because its a bigger model ,
they are very clear signs that gpt 5.4 was a smaller model , and that 5.3 even smaller (numbers are based on mil/$~price)
~~ but think about this way you get a team of 40 employees you invest in all of them so 5 are productive alltho at first glance its not a currect idea, and a bad practice , if you not a pioneer or attempt to optimize and beat limitation.. (speaking of AIM alliance by IBM,,, risc and arm design on modern times)
we should enjoy things , while understanding nothing is unlimited - alltho what we seen in server rooms in 2005 was pertty much perfect --- in the company I worked for as a networking person (routing ruleset ) we have hardware lasting for years and years ---- to a level we replaced hardware based on software (I was contacted to help because of legacy reasons) and it just lasts. Was Alexnet really big deal or not a big deal? was the apporch correct, or older papers and research focused on improveving and debating current research, they are likely tricks and methods and creativeity applied to other things, or moderate your life style because you are in central asia and export and import is expensive here - so you live a peaceful respected life.... you think of education of our youth --- you should give them or put them back in time like all of the research before Alexnet --- and make them thing of a way to recreate the Alexnet moment (where it crashed scores by 80x) --- like humans^ and models* you can get crushed by all of the noise and losse attention -- you think of our youth , and you ask how we make them acctually prodcutive (more than you were), how do we make good universities... instead of giving them degrees and having them needed years to be somewat prodcutive ... and you need to be productive because there is a lot of science undiscovered ,, like you can think of 1700s sailors discovering land and islands , but now there is nothing to rediscover -- alltho much cheaper ----
Sweden is big , and swedish is big but not so big for language models, so is kyrgyz .... Alltho its russian with a lot more curse-words...
you want to buy a billion dollars worth of tomatoes, you don't show up to the grocery store and hire and then and then and then they don't have a billion dollars of tomatoes and you go, "Aha, you're withholding tomatoes from me. "
my personal choise is 5.4mini ,
its really powerful,
its also way better on ratelimits --
to a level that on weekly limit (free tier) I get about 30-40 messages?
and you know , you think you get a better bang for the buck , if you use 5.5 , and its true --- but not always... so you should priotize when you use the bigger model or not ---- but again (your mileage may vary)
For example they are 2 type of Coding Agent - users ..
2.1 (prompts that are too big , that sometimes cause the model to aim on the wrong things, instead of stopping and ask the user what he means)
honestly codex seems to do it the right way, because differnet people mean other things (its also higher quality dataset -- and I guess they can synthesis dataset with cross refrence to q&a logic )
2.2 I be honest I like to hammer codex with big prompts my self too , but small prompts seems to be the best
Again, the model that has so much more edge than anyone else , is able to serve it so often to so many people --- so clearly I wont be surprise if 5.5 is smaller than what we saw in 4o
understand how powerful the 5.4mini model is , I think they designed it for many many agents to work togther so this is the idea , and the reason why gt5.4 and gpt5.5 considered being so optimized
when 5.2 was released I was very existed to see something from openai but then it became like "What you want to do a b c " and it repeat itself... So I was wondering if OpenAI have the wrong approch to things--- I was thinking not --- but its likely proven to be very correct.
For exmaple with 5.5 and 5.4 , you would ask the model to train for example a yolo model for you -- or help you label a dataset --- or create a self-attention model --- maybe using DinoV3 or DinoV1 -- to create a CRNN or a Transformer (but then maybe its the wrong prompt -- but its lucky a vision issue... First see the task, what you trying to achieve -- thinking of refrencing older tasks ---- view 10-30 samples [but uhh you see , its not like a human , that look to a small computer screen and already understand the issue from looking at a windows folder, and 20 other small thupnails ] )..... And you know its kindda failing the task,,, when its the issue is simple --- But if you have a small task for the ai --- it can do it... I dont know
you are tinkering with pm2 -- or give a another user access to user when he runs pm2 (or not try to insert gnu style linux commands into a macos computer? or find universial commands - yet needing to use some edge cases -- orbut simplicity mean because its a statstical reason?) - and its so hard for a human to find some obscure documentation? no problem you can ask CODEX to do it --- and it does it in a second --- when it take you 20 minutes.. to fix this minor problem---- this problem fixed, as agentic coding, fixed those type of computer usage task[image related] (notice the period where agentic-coding took , meaning people dont have a repeated problem. How can you know what you dont know? you expirement if you are able to reinforce?
thinking about economy look at AV1 being funded by Alphabet, and AV1 historically can only be encoded on a CPU, not a TPU or GPU... yet its still economical to encode a lot of youtube videos to AV1, but I am pretty sure with AV2 , hardware encoding would come much more natrually and wont take years... its amazing to think that GPU you dont think of them like in graphics per say, but more of a computing thing with backword compablity -- like you cant play a video game on a hopper cpu right? because a lot of the graphics part are stripped - yet it still needs to support things like 3d Medical Imaging -
without disrespecting ideas 30 years ago and calling it obsulete -- because clearly things are being retested just now.
Think how much of the GPU are limited by artitechture , transistor density, how they gonna write one transsitor to another --- yet saw 9x boost in preformance? between a100 to hopper
and from hopper to Blackwell, is a 2x boost?
think how amazing it is , that a lot of chip designers , can actually manage to design chips to ever evovling technology - like a developer uses an sdk or a game developer a game library
its not my field of expiertee , but it comes to design and limitation that they need to fight
<img width="1211" height="400" alt="Image" src="https://github.com/user-attachments/assets/c9ab597d-5004-4497-9c8e-f7e26d5424ac" />
I guess I am really happy, that I get to use it , consdiering how much demand there is for such models --- and how much of a precious thing it is --- so think of it 2024 chatgpt consumed a lot
[modern times as of today?] now a prompt is a serious of 10prompts (and maybe a smalelr model , along with many optimizations) , so about 5x the compute
honestly its a mircle we even see the age where you can run a server for 10 years , and it wont degrade much ---
I am optimistic about the future , because there is a way to optimize , keyed on the fact , that good code would be produced in the same way language is statstical , and we human are statstical biological models .... like when you think of a song you also visualize an image... you can see it as a template? and more and more language models seems to be able to use templates or whatever .... Alltho I did really enjoy the tempare and randomization of models ... and it seems to ask too much for guildlines -- when it should act like a divinci , creative and random --- instead of specified system prompts .... but for example
openai does things in a really amazing things ,,,
and humans can be overfited like ai models.. think of a chess gamer , learning the wrong books , and do the wrong repitiroe. Chess engines are also complex for example , I tried making a few , I never managed to optimize them -- when I was at university , and we asked to innovate or rewrite .... But you know its the friends the long the way
im pretty sure right now they are lesser neuroens in people brains , than they are transistors in computers (or ever had) likely
like past-memories in other models --- chatgpt doesnt seems to be poisioned a lot by it --- other do.... or they inejct the past memorizes too heavily -- when it should have a very low context idea or "influence"
and to be frank its pretty interesting , like in china you got Everything apps ... and then comes AI , and this is the first time you see an everything app attempt , that has any chances. that being said , chinese are trust based people despite being untrusted between themselve very often.... i myself a Koryo-saram so I understad this type of untrust very well))
I can resolv this problem， Maybe you use vpn, and then you have five reconnecting, after that, is thinking. but it is too slow,
so you need to use TUN Mode clash. It can work to me
Are you trying to access codex from China / Hong Kong / Macauw ?
You speaking of a buissness type of VPN , or an actual VPN in the sense of routing all of your network somewhere else
I know chinese people use VPNs often, and the latency there is really awful , and the vpn from the chinese sites can get very slow
I use both Codex and Opencode with a chatgpt subscription. Codex is at least 3-4 times slower
I personnally bought a macbook pro and swiitched to macOS, i can now use ALL my tokens, all is working well, latest version with latest features
I don't know how many times I've had to completely reinstall the Codex app over the past few months because of exactly these kinds of issues. How can they develop something so unstable? The coding models are actually really good by now. OpenAI needs to set a different standard if it wants to attract business customers. It can't be right that a small coding task takes 20–30 minutes. Generating an image? That can easily take 20–30 minutes, too.
Code base size?
@Mahkhmood9 Doesn't matter. Even with new projects that have zero files, Codex is extremely slow.
If I disable WSL, it works. So it seems that something in the WSL environment got messed up during a Codex update.
This guy is so funny. I never seen such chaotic speech by anyone. But I guess people can express opinions, considering some of the comments here are event more low quality than his posts… whether if you read and understand , or don’t understand but read this pile of *** , and it caused you to compile a correct idea, or or less
If it wasn’t for his spelling errors every sentence I was sure this was ai generated, as you may know AI doesn’t make spelling errors
He is like John Butler teaching you how to use Windows 2.0
Anyways are you using Sandbox or not ? habibi
I’m pretty sure there was a changelog regarding WSL try to update your codex
@888luck8 You’re apparently posting anonymously for a reason? Would you still express yourself like that if you had to stand behind your statements with your face and name?
Even if you start Codex in a new project folder with no files at all and ask it to create a simple .htaccess, HTML or JS file for that test, it is extremely slow. It takes >8 minutes to write just 1 or 2 lines of code.
Disabling WSL sped up the process for me. I’m always on the latest Codex version (Version 26.609.41114 • Released 12.06.2026), but I also tested this by completely setting up the entire WSL environment from scratch, including the WSL default system (Ubuntu 24.04 and Ubuntu 26.04). No difference whatsoever. WSL lags enormously. Sandbox is always active.
And the Codex Windows app itself is also significantly more sluggish compared to OpenCode, even on a modern machine with 32 GB of RAM and a CPU that is basically "on vacation".
The germ of thiz would be locked up in the original manuscripts
בתאריך יום ב׳, 15 ביוני 2026 ב-1:15 מאת wbdb @.***>:
maybe you already chat with it for a while - and ur context window is big - and it tries not to go too fast -- I recomend expirementing with using [medium or low] in the reasoning for small tasks
make sure to find an example you wise to become the same --- and try to describe it - atleast by words
if you are sastify with the results --- but unhappy by the speed - lets think about it more boradly
/new]speaking of .htaccess (apache http right) is really outdated
but if you like .htaccess LiteSpeed / OpenLiteSpeed (The Drop-In Upgrade)
If you love the convenience of .htaccess
alternativly there is
Caddy Servertoo which is very lean----
I also recomend if you got slop generated from models 4 months ago I recomend also review your code base and attempt to make code generated from the past smaller - as it does eat a lot of precious resources
//and its a ongoing issue// AI will never tell you have so much bloat in your codebase -- when this is what you should do (literally can be solved by
/plan debloat my codebase refactour it attempt to remove lines --- simplify the logic) (because you are on web development keep in mind not to cause "link rot" or change URL paths , just bad code --- )second - I recomend relaying on frameworks as a lot of it in web developement comes for it to ease development (allowing you to make your codebase lean -- and for you easy to design -- if you want something novel)
alternatively --- I recomend --
/goal is really good considering it can stop fast too (if it goes for too long and you are unhappy ,, kill it and rephrase it what you want,,)
thanks for reading)))
@Mahkhmood9 I don’t think the core issue is related to an existing long-running chat, a large project, or a huge codebase.
Even when I start a completely new chat with GPT-5.5 using medium or low reasoning, the very first request can already take several minutes while it only says that it is “thinking”. There's definitely a WSL issue with Codex here. Even in completely fresh installations. And many people are reporting it: https://github.com/openai/codex/issues/27240.
I already use Caddy and I also operate servers with Apache and nginx reverse proxies. Each setup has its own valid use case. Just because many people use Node.js does not automatically make it the right choice for every company, especially considering topics like supply chain attacks (Golang or PHP has a clear advantage here). In the business world, other factors matter more than always having the latest technology (which isn't necessarily the best choice for every individual company) and the fragmentation that comes with it.
It is absolutely terrible for me the last couple of weeks. I gave them 100 dollars to wait for hours for a simple task? Fix your service, OpenAI, and be quick or people will start switching to Claude.
I am unable to use codex at this point. It does not matter which model for me. Either it's just frozen the whole day when I use it on the cluster with vscode or just stuck at thinking mode if I do manage to send a prompt. it has been super useful for me the past months but I have not been able to use it this week..
speaking of the model itself? are you not using the Codex Desktop right?
whats the task, whats the codebase size?
do not use Extra-High or High
I recomend you Use 5.6Terra-Medium
I like the new models as it needed to get better...
)))