codex-discussions

[frosty zealot]

Sucks to suck get miniaturized fool

[kind jay]

Parameters are wrong now

[frosty zealot]

Idc it’s a cute little telescope

[exotic cave]

5.4 is an amazing model what are you on about?

[brave gazelle]

define amazing

[kind jay]

You're cute and little :(

All that typing for a 2/10 ragebait

[brave gazelle]

can it hypothetically generate good code, yes i agree with that. but the UX regression on 5.4 is horrendously bad

[brave gazelle]

you have to tell it what u want it to do like 5 times before it does it, then it stops halfway through, ya know just to make sure thats still what u want

[kind jay]

This is how they did the Artemis launch!

[kind jay]

Skill issue tbh

[lean lark]

flat Earthers are going "I told you they used AI in 1969!"

[potent mason]

Guys this is how they did it back in 1969

[brave gazelle]

didnt have this issue with 5.2 or 5.3 or claude models

[potent mason]

(Joking btw)

[kind jay]

I saw it in person

[brave gazelle]

5.2 was a work horse

i miss that model but its too slow

[frosty zealot]

I don’t know what horses have to do with this

[kind jay]

#gpt-models

[lean lark]

Actually I saw it live. 👴

[kind jay]

We're discussing codex here

[brave gazelle]

i paste screenshot of me use codex

[kind jay]

And? Wrong channel

[kind jay]

You

[frosty zealot]

Me core

[kind jay]

New pfp

[kind jay]

DM

[brave gazelle]

hmm i disagree i was discussing workflow using 5.4 inside the codex app but i suppose its a matter of opinion

[potent mason]

4k

let's goooo

[lean lark]

SMH - images are amazing

[frosty zealot]

I feel bad for my dad he already believes every AI thing he sees on the internet he’s always sending me stuff like 14’ people being like wow you see this?

[lean lark]

World isn't ready for really high quality deepfakes. 🙁

[frosty zealot]

Rather than make him feel bad I just say like wow neat

[ornate jolt]

new image model sucks

[potent mason]

The world isn't ready for a lot of what's happening with AI because we are moving so fast.

Security wise text models are going to get us all hacked, image wise and videos deepfakes, audio scams, etc

[lost drum]

yo how to use that gpt 2.0 on pro plan?

[ornate jolt]

cant even generate assets for my game

wont listen at all

[potent mason]

Put it in thinking

Well that was boring

[lean lark]

Looking forward to generating consistent game tiles

[exotic cave]

No new model. @kind jay lies...

[frosty zealot]

My bet is safe

[kind jay]

I said 23rd lil bro

I'm still up 80%

[lean lark]

no one knows when products will come from OpenAI. not worth the words.

[orchid plume]

images v2 looks impressive, but it looks like it's not in Codex yet

[kind jay]

I know and I told you, 23rd, quote me on that

[frosty zealot]

How dare you take that away from me

Hopefully my gif still came through

[boreal holly]

Probably will be in v0.123.0

[lost drum]

is not codex smarter today? no issues I had at all

[lean lark]

I've been running on 5.4-low just to save tokens, no issues so far on this lighter work.

(It's re-writing JaneBot)

[frosty zealot]

What if I use image 2 to generate pictures of the code I need is that the new gpt-6?

[true bane]

wow another shadow reset for usage, 2 in 2 days. Thanks.

[warm pilot]

what's a shadow reset? The last one was pre-announced

[frosty zealot]

They both were

One was from yesterday outage and today is from 4m codex users

[kind jay]

We hit 4m? I have a bet on that

[frosty zealot]

Yep

[kind jay]

+2k nice

Pay for my codex usage for the week

[lean lark]

Oh, this is so bad....

[lusty nimbus]

does codex desktop have access to image 2.0 yet?

[kind jay]

Yes

[deft sable]

GPT Image 2 is also available in Codex
https://x.com/OpenAIDevs/status/2046671238534496259?s=20

[latent tiger]

Can we use images in codex for design or is that not a thing

[trail marlin]

I have before

[kind jay]

Okay and?

[cobalt junco]

what an anticlimax

Thursday better be good

[stoic bay]

so instant ai is available in codex for a free user because l can't find it dispite l have all the GPT models 5.4 and 5.3 codex and so on

[lean lark]

Yeah, and don't forget to disable CC spending limits.

[lost drum]

hmm

ye right

[stoic bay]

so instant ai is available in codex for a free user because l can't find it dispite l have all the GPT models 5.4 and 5.3 codex and so on

[lean lark]

Please repeat this one more time, I think the third time I hear something that doesn't make sense, it makes more sense. Does that make sense?

[stoic bay]

l mean the gpt Instant is available for free users in codex app or not

because l can't find it but l have the rest of the models

[stoic bay]

.

[lean lark]

I see, you're asking if instant is or should be a model in Codex app. Anyone?

[orchid plume]

https://x.com/thsottiaux/status/2046678531695104187 - Codex is now using Images v2 🙂

[lean lark]

That's crazy ... I thought it already did all of those things.

[unique spade]

There is no gpt instant in codex. That is basically the chatbot you have in chatgpt

As a free user you have some free codex qouta but that's it. There is no free model in codex you can use, no one gets that not even pro users

[lean lark]

Be a troll, dude ... Post this in #openai-chatter . It'll go viral and OpenAI staff will have a big laugh.

[frosty zealot]

I posted mine in the Claude code discord

[lean lark]

Now THAT is bein a troll! Well done! 🤝

[boreal holly]

That's one way to get permabanned by A/

[frosty zealot]

Sometimes the risk is worth the reward

What’s funny is the mod replied to somebody else below and ignored my message lol

[lean lark]

Meh, they can use it to sue OpenAI for false advertising ... the joke'll be on them.

[trail marlin]

Has Codex started doing the reconnecting loop for anyone else or just me?

[lean lark]

not here SS

[kind jay]

Is this channel the 4chan of OAI discord server?

[lean lark]

nah, but this is one of the Fun channels here because we're creative, knowledgeable, and we have fun with the tech. Other channels here are fun but I think this one is a little cut above the riff-raff. 🙂
And sure, giving other platforms a little hell is just a welcome side-effect.

[kind jay]

Crazy tbh

[lean lark]

HA - Kris thinks it's real.

[unique spade]

Image 2.0 combined with gpt thinking is quite awesome

[loud dragon]

anyone else get their swag today?

[boreal holly]

we don't even need programming to make a killer website 😭

[lean lark]

OMG this is too good. Dude, you need to sell some marketing services with that. Or find some kid that goes "I dunno how to make money" and tell them "shutup and do this!"

[unique spade]

Let s see poeple complaining about codex sucking at design anymore 😂

[lean lark]

Paste GaaS page to #codex-show-and-tell and #openai-chatter

This is what @boreal holly is working on!

[frosty zealot]

My kink

[vital hill]

i saw this and had an inspiration, thank you

[ruby lynx]

do they offer in australia?

[lean lark]

Brilliant - the world is a different place today.

[loud dragon]

it was a limited giveaway for people that found the hidden skill reference during the super bowl commerical and invoked it, then went to a secret webpage to enter their info

[ruby lynx]

nvm

:(

nooooooooooooooooooooooooooooooooooooooooooooo

bruhh

[lean lark]

https://tenor.com/view/next-time-terrence-howard-war-machine-iron-man-james-rhodes-gif-6162269

[teal harness]

After updating Codex, the app frequently becomes unresponsive and eventually cannot be used at all.
This issue severely affects usability. Please investigate and fix it as soon as possible.

[lean lark]

I'm hoping anyone who generates those awesome fake/parody website images will put a link to themselves at the bottom to get proper attribution. I expect stuff like this to viral quickly.

[boreal holly]

It cut development time down from 12 months to 12 minutes 😩

[lean lark]

I want to share the cat one with a cat friend but I don't trust her not to publish in facebook.

[vital hill]

it's fine, i just did it for fun

but yall seen it here first

[lean lark]

wuts with the bot invasion?

[frosty zealot]

They gotta eat

[kind jay]

Yoo look what I made

[lean lark]

brilliant. Love it

[dusk thorn]

When I saw I got 2 rate limit resets same day (we all had to have run multi sub agent 5.4 xhigh chats bro in between times)

[blissful basin]

Looks like anthropic remov3d claude code for plans below 100usd

[dusk thorn]

You fr?

[ornate jolt]

i mean...it was useless on that low tier anyways

[high girder]

They're giving indirect claude code access from cowork

[frail meadow]

This changes everything

[blissful basin]

Nope

[high girder]

not 100% confirmed, but it has been swapped over on the website. Let's see them not say anything about it for ages.

[blissful basin]

Unless you meant for real, not french.

[frail meadow]

Probably because it’s too dangerous to use if you don’t make enough money

[blissful basin]

Tbf there are still advantages of using it in IDE

But imo they will loose a lot of people who would upgrade in the future, not everyone will go directly for 100usd plan

[high girder]

That's probably their goal still. Clear out people that don't WANT it, free up compute for those that stay

[blissful basin]

I still believe they have serious compute issues and all of this is just mitigation

[dusk thorn]

Are you kidding me bro

If it actually happens I’m cancelling the plan and going full codex

I got both 20 dollar plans for each

[blissful basin]

I had 2 months of Claude 200plan and canceled it, it was barely usable (didn't hit limits) but quality was very low, so it wasn't really productivity boost. Codex on other hand handles every task I need

[trail marlin]

go and look if you dont believe me

[chrome raven]

https://x.com/petergostev/status/2046720618566242657?s=20

[dusk thorn]

Hmm

It’s on website

But not updated in subscriptions

[elfin holly]

Yoo look what I made

[compact dawn]

i feel like they increased usage again?

on codex

i get so much more out of each % now

[unique spade]

out of curiosity anyone knows if you get more claude usage on anthropic sub as compared to google sub via antigravity?

i don t have claude sub, but i do test opus from time to time as part of my 20$ google plan in antigravity and it has decent quota

[frosty zealot]

Wow it’s so life like

[fallow steppe]

I'm using Codex, but it just can't seem to do a good job with UI/UX design. Is there anything I can do to fix this in one go? Does anyone know?

[ruby path]

I wish credits weren’t such an issue

I used 2000 credits in a night

[lime inlet]

I am using codex (cli, in xfce-terminal, Linux). How can I input a newline? I tried Shift-Enter, Ctrl-Enter... it must be possible somehow because if I paste an enter, it works.

[unique spade]

no. first of all i have no clue what "good UI/UX means for you". second i have no clue how exactly codex fails in your case. 3rd i have no clue what exactly are you telling it to do....so i dunno 🙂

[potent mason]

Has anyone used Chronicle?

Just from seeing it, it seems a little bit useless (+ after subsidation ends it will truly be useless because it will be too expensive to use), has anyone who used it have a different perspective about it?

[plush nymph]

thats not a flex

thats a disaster

[paper tartan]

I'm sure this was asked 100x so apologies in advance: is the new Codex IDE-style app update (in-app browser, etc.) coming to Windows? Mine's been on the same version since the announcement, and while I see a "toggle browser panel" option in the View tab, nothing happens. I'm working in a WSL environment if that's relevant.

[potent mason]

It doesn't seem like an IDE to me, the main feature of an IDE is literally just being able to modify files. I think they said everything except computer use is already on Windows too

[paper tartan]

yeah i hear you, it's not a real IDE (yet, haha), but I wonder if I'm just not seeing the new features yet? I'm a Pro sub, version is 26.409.61251

[magic moat]

I use GitHub copilot with Codex extension because it's easier on me

[paper tartan]

I've used both too, I want to stick with the Codex app, but it's been challenging at times, mainly with WSL issues that aren't present in VSCode's WSL setup

[dusk thorn]

bruh

[civic viper]

Hey guys ! Thinking about switching from Claude Code to Codex. Can I seamlessly use Claude's skills with Codex?

[plush nymph]

it's close

[nocturne folio]

why is everyones images on image v2 so good but mine are just genuinely terrible

[plush nymph]

some skills dont work butyou can usually have codex port skills

[tropic karma]

you can at least ask codex to import them for you

[plush nymph]

its not hard to have codex take a claude skill and rewrite it for codex

[civic viper]

So skills in Codex are also simple MD files?

[dusk thorn]

you can try, but some skills arent available

[nocturne folio]

theyre the same exact thing

[civic viper]

Ok so why wouldnt they be available in Codex?

[nocturne folio]

they are

[plush nymph]

im thinking more of tools/mcps. i havent used skills much.

[potent mason]

Different locations

[civic viper]

yeah so just a matter of moving files in a different folder I guess

[high girder]

are you doing image gen in GPT or codex?

[whole estuary]

Probably i made the right decision to switch to Codex this week

[ruby path]

Lowk

It kept messing up

[plush nymph]

highkey i knew.

like people need to accept that these are chatbots. they're text generators. furthermore the people operating these companies are saving the resoning power for themselves.

they give you trash for you to sift through, because you are telling the system what was a good or bad output

you are the computational power.

its impossible to get anything done right now without heavy oversight

they ahve extremely nerfed codex's abilities. if you are trying to vibe code something because everybody keeps telling you coding/ software engineering is dead, you're screwed.

[high girder]

Codex has gotten better. So much better. It does what you tell it to do

Ive been using codex to clean up after opus 4.7

[cedar skiff]

It's always been the same misconception propagated by ill informed hype train riders.

[frail meadow]

Does anyone have experience with building deep research agents?

[marble gazelle]

guys any idea why my codex wont open? it just stuck in here and never open...

[plush nymph]

im spending far more time babysitting. 5.4 overengineers/overthinks, writes a ton of crap i never asked for, doesnt seem to even understand english that well

I'll ask it to organize docs and it'll just write more docs

right now im in the process of going through all the docs i didnt even realize it wrote and its in the hundreds.

those are just the ones it wrote "by hand" theres thousands of files it generated with scripts.

its an absolute f-ing mess and this is while I was monitoring a single terminal

[high girder]

This is probably the only compliment I'll give Gemini outside of nano banana. The free tier Gemini does pretty good deep research if you are detailed enough for what you're looking for

[marble gazelle]

any idea on what to do?

[nocturne folio]

hoth

[plush nymph]

gemini is maybe the best CLI agent out there for web searching, as you'd hope from google

[high girder]

GPT definitely is using image2. Codex in the VS Code extension was still hitting 1.5 for a bit.

[plush nymph]

but you're still far better off using either gemini pro browser version or chatgpt pro with webseearch.

[unique spade]

update it from windows store

[unique spade]

a skill is literally an md with some nl instructions for the model that gets loaded in context

[unique spade]

educate your codex better then, you are responsible for what it does 🙂

i have to admit it's a lot of fun crafting my own UX for my personal workflow

codex on the left, chatgpt on the right. all the extras i need in the middle plane, including various automations between the 2, like linking 2 threads, sending with one button work/plan done by codex to gpt for review etc

that's why i love codex is oss , you can own the UX

[dusk thorn]

That’s sick

[unique spade]

that can be done on full vanilla btw. just use an electron shell in which you load web page of chatgpt and make whatever ux you want for codex and call codex backend via app-server

and whatever crazyness you want in the middle. have a separate codex thread act as your youtube/spotify DJ or whatever 😂

[dusk thorn]

I need pro plan for codex

Would be like infinite messages for me

Does the pro plan have codex security?

[unique spade]

i think only the 200$ one

i had it there a few weeks ago when i was on the 200$

now i'm on 100$ and don t have it

someone on 200$ can confirm though

[unkempt patio]

Claude Opus and GPT-5.4 xHigh feel basically tied to me. On hard tasks, I just switch between them and one usually solves it. The real difference isn’t the model—it’s the harness.

Claude Code wins there, mainly on planning and parallel execution. It breaks tasks into clear subtasks with a visible to-do list, so when it says “done,” you can actually trust it. And it can run things in parallel—if I interrupt, it keeps the first task going and starts the next.

Codex is still great (sometimes even better on specific problems), but the workflow is more serial: one task at a time, interruptions cost you, and you have to be more deliberate. It works—just not as fluid.

[unique spade]

well, yea but since codex is oss you can customize the harness...if you're bold enough
with claude code you can't

with claude everyone is forced to be a user of it
with codex the doors to be a dev are open

big difference

plus no one was born or trained to be an "ai dev" or whatever you wanna call it before 1 year ago, soooo the sky is open for everyone, and only limited by your own understanding of the AI phenomenon that is taking the world by storm

[unkempt patio]

Yeah, fair — and honestly, I wish there was a way to get this feedback directly to the Codex team. It feels like they’re really close to surpassing the competition, and what’s missing isn’t the model, it’s the harness (planning + parallelism).

[unique spade]

you can go open a issue
here is the guidline

https://github.com/openai/codex/blob/main/docs/contributing.md

[frosty zealot]

meeeooowwwww

[plush nymph]

lol

you cant educate codex. its going to drift regardless.

[lime inlet]

I just tried to update codex (cli) from 0.121.0 to 0.122.0 (sudo npm i -g @openai/codex@latest) on openSUSE, aarch64. Install broke for me. First time this happened to me. Anyone else using Linux (aarch64) and had that happen to them? Downgrading back to 0.121.0 fixed it for me.

[plush nymph]

you can give it the same prompt 100 times and you'll get 50 different responses.

this is how openai develops agents. you write the prompt, you get a response, you respond with good bad or otherwise, it keeps iterating on the weights model after model based on user inputs.

which i think is going to collapse the quality because as LLMs get better, the users get dumber and dontt know a good response from a bad one, which will affect the weights

[lean lark]

I think about this kind of thing a lot. As developers we just need to use the tools as best we can, learn from others, teach others what we know, create new mechanisms, try others, share, communicate. Most of the problems we recite come back to people operating in isolation. Recognize these challenges, don't just use the tools. The few people who break away from the prompt of the moment are the ones who shape the way the world uses this stuff.

you can give it the same prompt 100 times and you'll get 50 different responses.

That sir, is temperature. This isn't just OpenAI, it's the entire industry. We don't want temperature 1.0 for 100% determinant responses. We want the model to sway just a little to be creative without swaying too much into hallucinations. THAT'S the game.

[cedar skiff]

We have a convergence of evolution happening with the paradigm. There is no solid - This works and This doesnt guide and as a conscience many power users are making the same mistakes and coming to the same conclusions independently over and over again.
The simple example is as a newbie, people got the idea - I can use agents.md/claud.md to manage all the different ideas for the project. Then a month later they realise that is hurting them. Then they start to think of memory systems, or they start to think about skills etc.
All steps power users already visited and came to (mostly) the same conclusions.

What works is generally the same and takes a similar shape, it changes so fast though.

[lime inlet]

what is the issue with using AGENTS.md to manage different ideas for a project, how is this hurting users in the long run?

[unkempt patio]

https://github.com/openai/codex/issues/18940

[lime inlet]

I hear MDPs and skills mentioned a lot but simply do not understand the problem they solve. I definitely use AGENTS.md though.

For context I mostly do relatively small, standalone machine learning scripts to predict stuff and explain predictions.

[cedar skiff]

If you treat it like a living growing document it starts to have contradictions and ideas that hamstring the model for doing what it needs. It overloads the model with information it doesnt need for every task. It's not suited to be a large growing document, it should instead be solid rules that the directory needs and follows every time.

[lean lark]

There is a confusing stay of "solutions" and we're so weighing how the do and don't for specific problems.
AGENTS.md "mostly" contains directives that the assistant must follow. They cascade like CSS to assert priority. Other markdown files are merely suggestions, guidance, preferences. The trick before skills and still to an extent is to carefully word these files to assert directives with words like "must" and "will" rather than "should" or "can". Getting these models to follow instructions is extremely dependent on prompting skillsv and careful use of language, conveying specificity, intent, and limits.
All of this is in the air now with all of these .md file protocols. And it's especially precarious when every model and ever setting affects this stuff. There's no right or wrong, it moves too fast for that. People need to understand the concepts, the flow, and how words really matter, especially here.

And I didn't even mention the concept of tensions yet, ambiguities, contradictions, and subtle influence of poor grammar and even spelling errors. All of this needs to be recognized by professionals. If you aren't paying attention to this stuff, you get out what you put in, and it's not gonna be pretty.

[cedar skiff]

It really is an art

[lean lark]

Maybe I should have just said that. LOLz

[cedar skiff]

I'm a little excited for the new model coming up.

[frosty zealot]

Me too

[frail meadow]

rawr

[unique spade]

will probably be badass, at least based on how fast gptpro is lately

[frosty zealot]

I wonder if things like Expo/RN will go the way of the dodo bird, since you could easily build for multiple native apps at the same time

[lean lark]

Everything we do now has the longevity of a cup of warm coffee. We see it every day. I think it'll be like this for a long time, assuming the mundane types don't try to legislate it.

[frosty zealot]

I could see them doing it to save the economy some how lol

[lean lark]

they don't understand it enough to make better choices. they operate on fear.

[ivory crane]

?

What

[dusk thorn]

is this ai generated

what

OH MY GOD

WHART

[ivory crane]

No

[dusk thorn]

WHAT T

HOLY CRAP

[ivory crane]

Bruh

Dammm

[dusk thorn]

this a bug?

what is 2.1

[frosty zealot]

O.O

He aint trolling

[dusk thorn]

bro what the genuinely f is going on

[ivory crane]

Damm

[frosty zealot]

The question is, is it ACTUALLY routing to a 5.5 backend, or is this just a ui thing right now

[twilit bluff]

are these new models?

[visual canyon]

what happening ??

[frosty zealot]

I dont know what oai-2.1 is

[ivory crane]

Crazy

[twilit bluff]

holy...

[frosty zealot]

oh what i didnt even scroll or notice i just hyperfixated on 5.5

Maybe somebody screwed up lol

[ivory crane]

“Praise OAI.”

[frosty zealot]

@cedar skiff @plucky halo @lean lark

You guys seeing this?

Not a troll I repeat, not a troll

that or im about to wake up and ill have work

[lean lark]

gotta go back n read

[dusk thorn]

or rolling out for a release today?

[frosty zealot]

i want to know what the backend is actually routing to

[dusk thorn]

well couldnt be

[dusk thorn]

NUH UH

oh nvm

read wrong

mb

lmao

[frosty zealot]

Did you read back to bed

thats what i originally read lol

[dusk thorn]

yes

[twilit bluff]

lmao

have you tried the 5.5 yet?

[dusk thorn]

ima try rn for s and g

lmao its uh

[frosty zealot]

Not seeing anything yet on X

[ivory crane]

Maybe somebody do the vivecoding

[frosty zealot]

@ivory zodiac dude

[dusk thorn]

a few but not much

[twilit bluff]

oof, they probably noticed now

[frosty zealot]

@ivory zodiac I think this is the will I'm always spamming on X, you got the inside scoop on what this is all about?

[dusk thorn]

codex team tries 5.4 mini for dev work what happens next is shocking

[paper tartan]

Ohhh good looks, thank you!

[frosty zealot]

lol I think they didnt intend to push all those models to the front end like that

[dusk thorn]

no fun

[frosty zealot]

Just a guess, not sure why all those like arcanine etc. would be showing up

[dusk thorn]

yeah i assume some of the glacier ones maybe are for the 24/7 agents?

or just random crap

[visual canyon]

noo~~~

[dusk thorn]

booo

[frosty zealot]

Thats good though, I need it to release tomorrow or Friday so I can hit big on polymarket... if it 'technically' releases today I will win nothing lol @kind jay wake up wake up LOCK IN

[dusk thorn]

lmao

[twilit bluff]

so 5.5 this midnight?

or maybe next week

[dusk thorn]

smh

i got atleast 3 messages into 5.5 before most

[frosty zealot]

prob tomorrow, I cant imagine they'd just drop it and not a peep from Tibo, or OpenaI official account, especially at like 10 on a work night when they're likely all sleeping lol

[vernal comet]

why there is a model named heisenberg, I'm starting to feel fear.

[ivory crane]

Give me 5.5

[visual canyon]

Were the people who just saw the 5.5 model on the Pro plan?

[ivory crane]

[visual canyon]

My friend is on Plus, and they said it didn’t show up for them.

[dusk thorn]

im on plus

and i have it rn

[twilit bluff]

anddddd it's gone

[visual canyon]

nooooo~~

[frosty zealot]

Figured it was too weird, but, gave a look at some other internal models

Some intern sweating rn they just dumped their models

[visual canyon]

It was only brief, but I turned on Fast Mode and asked for a bit of DB tuning lol.

[frosty zealot]

like I said, it might not even route to 5.5 on the back end, was hard to tell

[dusk thorn]

well better get ready

[idle chasm]

"GPT, commit it yourself when you're done."

[lean lark]

looks like Pro is getting 5.5 A/B in the field.

[frosty zealot]

Noooo not thursday

[dusk thorn]

it says thursday

?

[frosty zealot]

nooooo

[dusk thorn]

you bet friday right?

[frosty zealot]

I bet it all, but I only win a $1 for thursday lol

[dusk thorn]

codex+big openai stuff is always on thursday

[frosty zealot]

Wednesday is like $80, Friday is like $250

[dusk thorn]

yeah now just says custom for me

some codex agent running on tibos computer had a little fun tonight

[lean lark]

enough with the poly already

[frosty zealot]

LOL

[rare olive]

• dont have nay of those models, which IDE are u using

[frosty zealot]

Its like a fever dream, nobody outside of this channel is gonna believe us, they will think they're getting Image-2'd

[visual canyon]

it blocked now

[frosty zealot]

"yeah last night at 10pm I had gpt-5.5"

and arcanine

and rosalind

[dusk thorn]

this is like the leak of the ai world rn lmao

[visual canyon]

Do OpenAI employees have access to all of these models?
That’s amazing.

[frosty zealot]

No spud in the list, interesting

[frosty zealot]

Guaranteed most likely, or at least a certain set of responsible ones

[glossy patrol]

FWIW I recorded a video. So people don't think it's GPT image generated.

[sacred plume]

I used arcanine for a second and it was amazing, some of the best code ever

[vernal comet]

seems so.

[dusk thorn]

fr?

[frosty zealot]

It solved world hunger

[dusk thorn]

oh damn

[sacred plume]

Yeah, it did it well. Won't show the rest as it'll leak some things.

[dusk thorn]

you asked it good questions, i said "yo"

[frosty zealot]

LOL

[vernal comet]

I'm wondering what's oai, is it a new model with something totally different to GPT like some new inference paradigm？

[frosty zealot]

I uploaded my resume and said please distribute this as a mass email to all openai emails

I should get a call tomorrow I would think

[glossy patrol]

Was using codex-cyber. So close to getting Mythos leaked.

[sacred plume]

Spud is GPT-5.5, Arcanine is an internal test of recurrent-depth transformers instead of COT from what I could tell. Glacier-Alpha is GPT's competitor to Sonnet, and Alpha-block was Mini and Nano.

pretty much adding a new model to compete with sonnet

[lean lark]

I believe that's correct, Spud=5.5

[visual canyon]

It’s hilarious how you can immediately tell it’s about to end, and then you frantically run it like an addict, haha.

[frosty zealot]

If spud is such a massive leap I'm amazed they dont release it under GPT-6

[glossy patrol]

GPT-26

[dusk thorn]

we shoulda spent more time figuring out which one is for unrestricted government access lmao

[sacred plume]

Heisenberg is internally the new version of Rosalind that just released. I got no idea what OAI-2.1 was

[frosty zealot]

I just bit my cheek somebody call the police

Maybe their OSS model or something?

[sacred plume]

Arcanine refused to show thinking, which is evidence of RDT instead of chain of thought

[dusk thorn]

maybe

[lean lark]

Gov has some version of 5.5. For better or worse i think they're the alpha.

[sacred plume]

no Arcanine seems to be their true tried and tested mythos level model, 5.5 will compete in Chain of Thought performance but not be the same tech (reccurrent depth transformers) as Arcanine

[glossy patrol]

https://tenor.com/view/frodo-keep-your-secrets-alright-then-lotr-gif-12907121

[dusk thorn]

local model intergration as a downloadable model in codex would be nice

maybe something they should explore

make them small like gemma and have them as nano or mini tiers

[glossy patrol]

Huh

🤔

[lean lark]

Would be funny if Arcanine is Mythos equivalent ... if Mythos was a stunt.

[glossy patrol]

I wonder if it'll respond.

[dusk thorn]

how did you get that

is it back?

[lean lark]

Meh, so speculation at this point. I just wait for announcements. When they say it's here ... it's here ... until then it ain't. 🤷🏼‍♀️

[glossy patrol]

codex -m gpt-5.5 or any other model name

Doesn't seem to be working of course. But worth a try.

■ {"type":"error","status":400,"error":{"type":"invalid_request_error","message":"The 'gpt-5.5' model is not supported when using Codex with a ChatGPT account."}}

[dusk thorn]

yeahh

[brave badge]

👀

[frosty zealot]

I keep getting re-edged

by these late screenshots

ive checked it like 9 times

[lean lark]

going back to whatever i was doing...

[frosty zealot]

Shouldnt you be asleep

You get a little senile when you're awake too long

[lean lark]

why i oughtta...

[frosty zealot]

🧓

[vernal comet]

It was said that the API endpoint has been closed.

[glossy patrol]

If we had Mythos, we could reopen it.

[dusk thorn]

claude, use mythos to get me the true and better model from openai. make no mistakes

[frail meadow]

so anyone that was able to test it while it was open, anything interesting discovered?

[dusk thorn]

yeah

[dusk thorn]

read from here and down a bit

[ivory crane]

Giv me back 5.5

TT

[fallow steppe]

uhh guys i cant load the previous vscode codex conversations, anyone can know how to fix this

[dusk thorn]

yeah its so over for anthropic 😭

now we know who dropped the models @frosty zealot 🤣

[fallow steppe]

gpt 5.5 is coming?

[dusk thorn]

sam is just going off on x rn

its so funny omfg

[fallow steppe]

oh

[lost stag]

codex is great

[cedar skiff]

You have to give it the tools to handle it.
My process is:

• make sure it has all of the linting tools its needs and has instructions to use them to see and fix mistakes. essentially i give it all the feed back and tools i have in the ide - autofix, or format, cli linting access etc.
• Any gaps it still has add concise instructions to a skill that is always loaded when dealing with the language/framework.
  It will be writing idiomatic code very quickly.

[tiny fulcrum]

are they now just ressetting the codex limits daily? is this like 2 days in a row they did it?

[turbid axle]

if we get 1 million new users every day, sure

[cedar skiff]

I have the opposite experience.

[rugged siren]

wow

[turbid axle]

tbf, at 1 million a day it would still take about 20 years for all of humanity to use codex.

we gonna need more gpus fam

[lusty nimbus]

is anybody able to achieve good UI results with Codex, or is Claude still the move?

[high girder]

Codex got better if you give it enough context. The images from image gen 2 that you can generate inside of codex now as well, can be good UI anchors for it.
If you have access to claude design, use that to wire it up

[orchid plume]

yes, I'm currently working on my own skill (UI from image) for this exact purpose. Feed it an image, let GPT-5.4 do 99% of the work with maybe one or two minor changes after it's made it into code

it works impressively well

[lusty nimbus]

gotcha yeah i've been using Opus 4.7 for most UI stuff and it seems to kind of guess or interpret what I'm asking of it better than codex

[orchid plume]

GPT-Image-2/Images v2 obviously makes the UI concept image

[lusty nimbus]

lwk prob just a skill issue tbh tho

[high girder]

Honestly. Opus is good, but codex is excellent at catching what opus misses

and now we wait for 5.4-cyber access <_<

[lusty nimbus]

yeah I usually plan things out with Opus then have codex implement or look over and fix what claude implements

[orchid plume]

rumor has it that GPT-5.5 will be another step up in frontend though, presumably coming Thursday. Someone had a brief opportunity to use the model when it was "accidentally" leaked last night in the early hours of the morning briefly 😄

[high girder]

ideal workflow

[lusty nimbus]

fr

[high girder]

You mean when sam was drunk lol

[orchid plume]

haha, nah, I suspect it was a deliberate teaser rather than an accident. They did similar "leaks" for 5.4 too

[lusty nimbus]

I really wanna upgrade to the claude max plan but if gpt 5.5 drops that could make it useless lol

[cedar skiff]

There is a few a/b testing instances around where ppl hae had access (probably to pro) and they did some nice stuff.

Still remains to be seen if it will be 5.5 or 6

[orchid plume]

true, looking forward to the launch hopefully Thursday. Based on last night's brief leak it was named GPT-5.5 at the time

along with some other codenamed models

[cedar skiff]

Since it's a new base model i would expect it to be 6

[high girder]

Oh, that would make sense on why I had one of those then

[orchid plume]

agreed, it does seem odd that it was 5.5

[cedar skiff]

oh ok, nice. Is this on reddit?

[orchid plume]

I found it mentioned on X, not sure about reddit, I usually avoid that place 😄

let me find the link

[orchid plume]

https://x.com/marmaduke091/status/2046803980089536718 and also scroll down for a potential frontend preview someone got out of it before the model disappeared apparently

[cedar skiff]

I guess why would you do a huge step if the competition didnt iterate in a strong way.

both opus 4.6 and 4.7 where pretty weak iterations for coders

[orchid plume]

yeah, curious what the benchmarks will say, of course can't go heavily by them but just curious 😛

[cedar skiff]

bench marks really dont mean a lot

But like you say we always have a look at them

[orchid plume]

yeah, sadly so, especially in Gemini's case 😄

[supple perch]

incoming rest

reset sorry lol, I need rest

[high girder]

I was working on a project with opus 4.7 in claude code, gpt 5.4 in codex, and the free tier gemini. Claude essentially told Gemini to screw off.

[oak trellis]

wanna invest some time into building this: User Input ↓ LLM layer ↓ LeWM / JEPA core ↓ Memory + tools ↓ LLM response layer

[cedar skiff]

just let the big bois do that

[turbid axle]

I just make whatever I want/need, knowing that in a year or so an AI can probably do in a day what now takes me months

I mean, what else am I gonna do

sisyphusmaxxing

[alpine wasp]

Hey guys, is anyone else getting a server_error when using OpenCode? Keep getting the "An error occurred while processing your request" message. I've tried retrying but it keeps happening. Is the API down for anyone else or is it just me?

[dusky spade]

+1

[cedar skiff]

Something i avoid is trying to build for the SWE domain. Since we are experts here it's where we see opportunity. But like you say it will be trampled by ai progress.

[alpine wasp]

is it related to the GPT-5.5 model that just rolled out?

[cedar skiff]

Imagine if they rolled out the model and didnt say anything.

[dusky spade]

no, I was using gpt 5.4 for planning and well...

[turbid axle]

yeh, I just build tools for myself for real life improvements etc now.
things that help me on the short term.

just had gpt delete the entire memory system I had it create a couple months ago because now codex has its own 🙂

[quick ermine]

+1

[alpine wasp]

I also use the same model as you, my question is, is it because they are rolling out the latest model so their API endpoint is broken?

[dusky spade]

I think probably because of the Image 2 model released recently

[cedar skiff]

I did many things like this when i first got claude code, i saw so many improvments, but they never paid off. Just cost me time and effort

[dusky spade]

Do you guys use the codex cli or opencode?

[cedar skiff]

codex app, i use droid for minimax

i havent touched the cli in about 2 weeks now

[turbid axle]

cli and app for me. mostly app these days

[dusky spade]

I am on linux, so the app isn't available to me yet

[kindred moss]

codex was unstable now!

[cedar skiff]

I have droid on linux, i like it more than opencode

[dusky spade]

I actually never used the codex cli and github copilot cli

[dusky spade]

tbh never heard of droid

not a hardcore vibe coder

[kindred moss]

what provider did you use on droid
there subs?

[dusky spade]

would love to hear your opinions

is it the factory ai thing?

[kindred moss]

droid is a clone of claude code with better UX

[cedar skiff]

I use minimax on it. They have byok without needing a sub.

[alpine wasp]

really? I only know about opencode because I previously used Droid

[dusky spade]

😅

I have been using opencode for a while now

I use other subs via opencode as well

and not directly via their own clis

[cedar skiff]

I have open code, droid and codex on the linux box. I mostly use codex and droid to manage the openclaw instance i have on there.

[kindred moss]

@cedar skiff did droid has an run commands in background like codex did

[cedar skiff]

you mean like codex exec? - yeah it has this

[dusky spade]

Ok so let's cut to the chase,

droid or opencode?

I have my projects setup using the global agents framework .agents, AGENTS.md and so on, will droid be plug and play or would require more config?

[kindred moss]

Codex can run commands in the background during the session. It keeps waiting for them and does something else while the background commands finish

[cedar skiff]

droid supports everything skills exec running etc, it has spec/plan mode and so on. One of the main draw cards i liek about it is has it's own inbuilt web search that works, unlike opencode.

[bright swift]

codex cli custom fork is the way to go. app only wastes resources and tokens

[alpine wasp]

I’m pretty sure opencode is the way to go. It’s just way better, especially when you factor in all those community plugins that make it super powerful

[cedar skiff]

not sure tbh

[dusky spade]

seems appealing

but web search has never been a big deal for my usecases so far

[bright swift]

opencode has so many bugs, couldnt believe it when i tried it out

[dusky spade]

I got flashbanged yesterday 😂

[bright swift]

read tool hallucinating paths etc

[bright swift]

like this should be solved, its mid 2026 almost

and with the github copilot gpt 5.4 + opencode "hack" now ending, i dont see the point of opencode

[dusky spade]

Oh btw, when I used copilot cli a few times.

the task completion and everything was way different. I tell the models to one shot the plan, opencode with copilot sub does it in one shot.

copilot always stops in between no matter what I do

[dusky spade]

wait what?

[bright swift]

https://github.blog/changelog/2026-04-20-changes-to-github-copilot-plans-for-individuals/

rip 12h long work with 200 subagents for 1 request

[alpine wasp]

Yeah, true, it's still super buggy. One of the reasons people even use Opencode is because of that oh-my-opencode orchestration. Even Anthropic blocked Opencode 'cause so many people were cheesing their Pro Plan through it lol

[dusky spade]

The status pages still say fully operational 😂

[lime inlet]

My reading is that this is only about Opus 4.5/6/7 ?! Where does it say anything about GPT?

[dusky spade]

actually same question

I didn't find anything in there that mentions about opencode

[bright swift]

Gpt 5.4 via copilot is still great value

[dusky spade]

It does say ratelimits

[bright swift]

But they are introducing token based rate limits

[dusky spade]

Oh that is already there

[bright swift]

So not request based only anymore

[dusky spade]

was ratelimited for the first time in my life with copilot yesterday

[wicked briar]

can you do this with claude

[dusky spade]

for way less work than i usually do

[lime inlet]

My reading is that the GPT series can still be used with opencode. Generally codex is the better harness for GPT I feel, but I also use GPT+opencode to get those tokens as well through my Copilot Pro account.

[bright swift]

Sure its still usable

Just not 20 parallel sessions that last for 12h each with hundreds of millions of tokens used each through heavy subagent use

For one request

My guess is they will switch to token based plans instead of request based soon

Might still be good value. But I cant get the same codex behavior out of opencode so far despite a lot of system prompt tweaks for some reason

[wicked briar]

I mean copilot is basically gpt models running on azure hardware

[bright swift]

And it doesnt have the magic server side compaction that codex has..

[wicked briar]

microsoft losing money

[lime inlet]

You mean you are making the opencode prompt more like the codex prompt? I was thinking about doing that.

[bright swift]

yea i basically started with the codex 5.4 system prompt, then removed/adjusted things to make it more suited to autonomous work instead of interactive with the user etc. but unfortunately it still tends to overthink a lot

And opencode tool usage has a lot of bugs that get in the way, so needed to build custom plugins to get around that.

Of course once i had it mostly working the github plan changes are announced fml

The "exploit early and often" from WoW comes to mind 😄

[orchid plume]

I used to switch between Copilot CLI and Droid (both setup as BYOK), but now that Codex itself has improved quite a bit over time, and of course now has image generation/edit support, I just use the Codex app now

[bright swift]

I'm paranoid to install the codex app because it seems to affect usage consumption in the cli

[cedar skiff]

codex app is really great, now that it has forking at any conversation point, and has nice ui for plugins and skills etc.

[orchid plume]

I think Codex is more greedier with input token usage, it likes to read more which means context can fill up faster too, when compared to Copilot CLI or Droid

[cedar skiff]

I just really jump in the deep end on it and turned on the memory and chronicle today

[orchid plume]

I wish Chronicle was available here

[bright swift]

Yea buts thats what makes it work so great compared to claude code for example. It makes sure to get the right context

[orchid plume]

definitely, it has a much better understanding of the context 🙂

[bright swift]

Implemented sth in my codex fork, then after i was done i realized theres a feature flag for it lol

[cedar skiff]

claude code is just lazy, i don't get it. It can do the work, if you add prompting layers and force it to do whats needed it does. It's like they are trying to conserve tokens by having it aggressively infer and make balanced assumptions as soon as it can, but that causes mistakes. So in the end it cost more tokens because you have to redo stuff anyway.

It keeps iterating in this direction more and more over the last year. In the end i just got rid of it.

[orchid plume]

when CC's code leaked, I was astonished at the amount of issues Codex found in its code, and how many secret things existed in CC that potentially made the UX worse

[cedar skiff]

One of the telling things was how they were injecting the claude.md over and over to make it remember.

[bright swift]

Codex does that too though, agents.md is added to the system prompt which is a seperate field in the api and gets higher attention (at least thats what codex told me after doing some investigation on it)

[cedar skiff]

Right but turns out claude code was putting in each turn

[bright swift]

Oh ok

[delicate sandal]

Guys, can someone tell me that why when I use the Codex then it sometimes uses some of my PC storage?

[bright swift]

Its cozying up in there

[delicate sandal]

Why

But it sometimes takes too much of my free space

[bright swift]

Probably stuff like caches, node_modules folders etc

[delicate sandal]

Like once i had 30gb left of free space before using, and after using it, it got left me with like 15-14GB of free space

[bright swift]

Just ask it to analyze disk storage and come up with a cleanup plan, but approve each one, dont let it clean up on its own or you might end up with a "clean" (= empty) disk

[cedar skiff]

I cross my fingers for a bigger context window

[delicate sandal]

I just give it full control, and when I ask it to analyze it, then it tells me that the taken up storage comes from other files or folders or etc.

[bright swift]

the 200k to 1m window jump in opus made things worse. i think 200k context window + magic auto compact to let it run "forever" is the better approach

or 270k or whatever it is.. interestingly via gh copilot the gpt 5.4 context is 400k..

[wicked briar]

do you know how to turn that feature off

[cedar skiff]

If we can get higher context without problems I'm all for it.
The limitation is in planning, you can end up in a spot where there is not enough context after planning and it hits compaction too early, so you get real context diffusion and loss that actually hurts.

[bright swift]

The new 0.122.0 has the "clear context before plan exec" thing

[cedar skiff]

Why, if you dont use it its like it doesnt exist?

[bright swift]

But honestly i mostly use the "just talk to it" approach for planning and at some point i say "ok LFG" and dont pay much attention to when compaction happens

[cedar skiff]

I found compaction deeper into implmentation doesnt hurt at all. But if you get it too soon OR even before implmentation it hurts outcomes, i think this is why 5.4 extra high can have bad out comes compared to high.

An extra 100k would be enough

[wicked briar]

hmm im working on main branch only rn to avoid this

[plush nymph]

lol 5.3 codex is at capacity. forced to use 5.4

guess 5.3 codex is popular

[steady vigil]

yeah I find it much better then 5.4 for most coding tasks

[ivory crane]

Oai2.1 is the best

[plain crescent]

Any of you also make use of only Codex, and not other features like normal chat, video generation, voice mode etc?

[cedar skiff]

oh this isnt work tree fork its just local convo fork

[wicked briar]

really?

[tawny turret]

I found a company that sells codex only

[wicked briar]

5.4 high is good

[wicked briar]

okay yes

[tawny turret]

I get 2x 20$ accounts worth of codex for the price of one, but like no access to regular chatgpt and video generation etc, not using it anyway so im fine. can send u link if u want

[bright swift]

I use mainly codex but outsource some heavy reviews or planning to gpt 5.4 pro via oracle cli

[fallow steppe]

bro i try to enter my old conversations in vscode and its not loading

[kind jay]

Send me link pleaseweee

[bright swift]

sooo creative with the subagent names...

[tawny turret]

uh

[tawny turret]

LOL

[tawny turret]

its lowk limited seats i need to ask them how many ppl they allow rn, its early access

[kind jay]

Pleaseeee

[ivory crane]

What.

[tawny turret]

LOL

[ivory crane]

What is this model 😭

[tawny turret]

check ur dm. i told u ill let u know

[ivory crane]

Bruh

[cedar skiff]

It's the model equivalent of doing alt+f4 to get goodies. PPL stopped falling for it 23 years ago.

[wicked briar]

what are you working on

👀

[bright swift]

data analysis system with custom ML

[wicked briar]

hmm

[brave kettle]

hi everyone, whats the best interface for Codex in your opinion and why?
a. CLI
b. Desktop App
c. IDE Extension

[orchid plume]

I prefer Desktop App now, easy to handle multiple sessions at the same time, can see images it generates or edits, clean UI too

[bright swift]

cli, performance, control & parallelism, easy to run headless in remote envs etc

cant fork the desktop app to fix things OAI is slacking on

[cedar skiff]

i like the app as well, one click commits and push and pr etc, just point and click to fork local, everything is easy to access. Voice mode is flawless i hardly ever type prompts now.

[brave kettle]

i really love Codex that's why i wanna maximizes the usages 💖

i never tried desktop app, this is good insights.

[brave kettle]

is git diff included in desktop app? if yes, is it beautifully presented?

[cedar skiff]

yeah

[cedar bear]

Tbh I use both the app and the CLI from time to time
But if I had to decode between the two it would be the CLI

[brave kettle]

is both interfaces shares the same conversation history?

[cedar bear]

If I understood it correctly, yes

[bright swift]

the app uses the cli under the hood through the app server

[brave kettle]

i see

[gusty basalt]

Why are my buttons to run my applications and to open them in VS Code gone?

Where did they go

[hard scarab]

im having this codex version and how to get remote connections here

[glass furnace]

[features]
remote_connections = true
in your toml if you want the connections

as. long as you are up to date and in the US, that will make it show up

[hard scarab]

already did that but didnt get and if change this file how to connection menu will come in app?

[hard scarab]

did that but in codex app im not getting connectins yet

[cedar skiff]

i think you need to have

[glass furnace]

^

its not a remote control. It will not turn your tv on and off. That comes with gpt 5.6

[hard scarab]

can i do alternatively via cli or any other to reflect in codex app

[glass furnace]

if you have the cli on your mac, just run it and say:

Please update my config.toml for codex to have the following so i can use the new connections feature and be great like Not Whale:

[features]
remote_connections = true

but it should be simple enough that you can just edit it yourself.

[hard scarab]

this one ?

[glass furnace]

yes but it shouldnt say "remote_control" it should say "remote_connections"

[glacial shadow]

Does anyone know if they’re working on a remote_control feature? I want to talk to the CLI while hiking and running

[hard scarab]

changed now

[glass furnace]

you can already do that.. download termius and ssh

[hard scarab]

now this error

[nocturne folio]

so excited, gpt5.5 is dropping

[hard scarab]

where?

[glass furnace]

you need to blow into it

[cedar skiff]

ask it nicely

[glass furnace]

tell codex to fix it for you. You are literally using a expert coding slave app. You can do this!

[hard scarab]

codex is now on free chatgpt subscription also? that means we can use free model request there?

[turbid axle]

'expert code slave'

[glacial shadow]

OK, yes I didn’t want to do it that way

[glass furnace]

am i wrong?

[turbid axle]

slave seems a bit anthropic perhaps

[stiff leaf]

just asked codex to review my code & look for bugs & it's taking more than an hour is this normal ?

[dusk thorn]

Holy crap 😭

[lost drum]

do you use review feature? cause I ve never used it idk how it works

[dusk thorn]

That exists?

I just ask for a security review

[lost drum]

ye I just ask to validate if eveything ok

[stiff leaf]

no just normal GPT 5.4 & it's still working smh & running commands & asking quetions so it's not a bug lol

[lost drum]

but I see in codex menu that I ve never used a feature code review and idk whats that

[lost drum]

if he soing something then give him 1 mor hour haha

[stiff leaf]

idk how it's still working even tough token are out for a long time

[dusk thorn]

Reviews code for security and cleanliness ig

[lost drum]

oh

[dusk thorn]

Once you run out rate limits it still finishes the task

[stiff leaf]

lol

[dusk thorn]

Because openai isn’t anthropic

And actually wants your task to finish

[stiff leaf]

what is the limit for execution of a task ? is it unlimited ?

I don't think he is going to stop anytime soon lol he is even editing my code even tough i didn't ask him to write code, he already wrote around 1400 line

[lost drum]

so he found something probbably

[gentle harbor]

whats the point of cloud ?

[stiff leaf]

Yes clearly he found flaws also, from commands he is running, he is litteraly building the software & reading log alone, idk how he did even find the builder path while running in a sandbox.

[lost drum]

AGI

[gentle harbor]

makes sense that the model wouldnt say the exact version but a bit of a shame

[stiff leaf]

i tough AGI does not exist lol

[lost drum]

its just controlled

[unique spade]

5h limit usually let s you complete ongoing turn

[ivory crane]

[unique spade]

If you don t have a computer 😂

[stiff leaf]

im just gonna let it work at this point.

[gentle harbor]

but you cant use codex on the web or mobile so what do you mean ?

[glacial shadow]

What are you guys coding?

[gentle harbor]

custom plugin to test codex

[unique spade]

You can use codex cloud obviously even from them moon. Are you new?

[gentle harbor]

but you still need the codex app thats on a pc only last i checked

[unique spade]

No

Codex cloud has 0 to do with you computer or desktop app. It s on web

https://chatgpt.com/codex/cloud

You give tasks here from web to cloud

[left quarry]

Good day. Does anyone here use the Codex agent inside Copilot in VS? I'm on the Pro+ plan, but nothing happens when I select Codex. GH Support is silent, so I'm trying my luck here.

[glacial shadow]

To test it for what

[gentle harbor]

to see how good it is and what task it can do, failing so far sadly same with claude

[glacial shadow]

What. Makes no sense. It works fine

You’re wasting time coding a plug-in into test codeX, but you’re not actually using it for anything

[unique spade]

Only human fails. Machine works 😂

[gentle harbor]

works fine for what you do, not for what id use it for 😅 its good at most task though !

[stiff leaf]

Yes it's not perfect, the more i read the messages & console commands log the more i get this, he is rebuilding multiple times & failing because of errors related to inventing functions that does not exist, i don't understand why codex does this lol

[unique spade]

Mine makes money trading. I'd say it does everything it needs to do 😂

[glacial shadow]

And what would you use it for?

[gentle harbor]

yea so it works fine for what you need it to do as i said ..

[glacial shadow]

What instrument?

[unique spade]

If you don't tell anything about your use case. What s the point of the convo though

[glacial shadow]

He’s a China bot account

All they do is spam USA social media to make us think our AI isn’t good

[gentle harbor]

using zig for decompiling and fixing compiler errors, not that hard to do but the safety restraints always think its something bad for some reason 😆

[unique spade]

Lol yea makes sense and I noticed a few of those too

The pattern is "codex is bad" but never go into a convo like an actual human wanting some help

[gentle harbor]

i didnt say it wasnt good ? i said it didnt work for what i need it to do

[unique spade]

Didn't let it yet to actually do it. But I'm buikding my trading system with it. So far I still do the trades manually till I'm comfortable I have the entire picture clear to actually automate my trading strategies

[glacial shadow]

They create millions of accounts virtually using AI, and they spam on discord Reddit, Instagram, TikTok, and Twitter. There’s videos showing the warehouses they used to do this

[glacial shadow]

Crypto, equities, commodities?

[unique spade]

I imagine. And it's annoying. Because they leech emotions of actual humans

[glacial shadow]

Yes, CHYNA is trying to beat us in the AI race, and if they win, they want to take over the whole planet

[unique spade]

Crypto, easiest because all the crowd there copy influencers or use dumb bots. It's so predictable and has very good volatility

[turbid axle]

hey, the reset is fake. only yesterday or so my reset date was 26th, now it moved 28th.

boooo

[cedar skiff]

This guy has called everyone who disagreed with him a bot, and he probably believes it.

[cedar skiff]

Yeah the date moves to 7 days from the reset

[glacial shadow]

This one is a bot too

[gentle harbor]

Wasnt it just a bunch of stripped down phones they only have the mobo and one giant screen to use it ?, i dont see why they would focus on that if they have no better versions, k imi and de ep seek are both awful even for general use, anyway gemini seems to know zig better then codex and claude for now, i hope spud can fix that

[unique spade]

Yea but Eric is an advanced version 😂

[turbid axle]

but then what is the point. its only useful for people who ate all their tokens and are waiting for more

[cedar skiff]

yeah, i agree. often i get less usage becuase of it

[glacial shadow]

Part of it entices people to sign up because it’s essentially getting more than what you paid for

Most people only want to pay 20 bucks, when an actuality this stuff cost way more

[turbid axle]

Guess I need to take their 'we will reset soon' way more seriously and go full ham with agents to eat as much tokens as possible

[glacial shadow]

I wouldn’t do that because that’s not honest usage practices I would buy the 100 $ plan

[unique spade]

It s useful for anyone who used more than the average for the period already passed.
So it s basically incentivizing to work towards your limit earlier rather than later (and hoping for a reset)

[cedar skiff]

I like to use my extra usage in the last few days of the month.

[turbid axle]

I have that plan..

[lost drum]

Do they plan another credit reset?

[turbid axle]

no

they just did 2 in a row

well I guess resets are a scam 🙁

[gentle harbor]

maybe when spud comes out 🫣

[lost drum]

Ye that was insane I mean the worse part is that it actually gave me less usage

[lost drum]

This week?

[turbid axle]

I guess they are riding the mythos wave with spud now though. probably going to see some 'upgraded' models until that ordeal settles

[gentle harbor]

people are saying it comes out the 23rd so tmr

[turbid axle]

maybe tomorrow, but I think its just 5.5 upgrade, not gpt6 epic AGI megamodel

[lost drum]

Idk what would it change gpt 5.4 xhigh already a big brain

[unique spade]

The 2nd gave few hours heads up. So if you had your 10 projects ready and lined up, all specs prepared, you could have just hit the ALL IN button
And you re gonna use most of it in a few hours

[lost drum]

Ye that agi model could get us all working as plumbers soon

Nothing wrong in it but

[turbid axle]

I dont mind working with my hands tbh. its just that pay is garbage, otherwise I would prefer it ez.

[unique spade]

You guys should stop wishing for official AGI... when that happens you're gonna get eaten by bigger predators who just start using agi

[lost drum]

Ye the ones with more budget wins

[turbid axle]

im hoping for art utopia though. no more work

[still trellis]

I got my merch! Thank you OpenAI team.

[lost drum]

Merch to be friends with gpt terminator

[unique spade]

That s the plan. You can read openai recent policy recommendations for industrial Ai era.

[still trellis]

When in Rome I guess

[unique spade]

Sam has always been into ubi and related concepts, see his separate worldcoin project

[turbid axle]

I'm aware yes

[lost drum]

Gotta preapare when they come

Say thank you even if he messes up

[unique spade]

Might take some years though, and not because we lack the tech, but because of the larger societal global legacy issues

At societal level we re still collectively behaving like apes with clubs unfortunately

[unique spade]

Larger context without loss.
Speed. Cheaper inference

There s a lot of directions to improve

Raw intelligence is already strong, gptpro is solving open math conjectures since late 2025 sooo

[turbid axle]

yeh, not too worries about this transition period though, its gonna be wild, but whatev. far more worried about what come after this utopia

[unique spade]

Why worry? What you're afraid of

Some form of blob hive mind?

[turbid axle]

not really fear. more just a big ball of 'unknown' coming our way.

our entire identity is our intelligence. I wonder what happens when every thought, every datapoint, was generated by AI, not humans, or animals, or everything else we evolved together with.

we are summoning a gigantic mega planet sized meme whale. it will wipe out every single human generated meme in no time.

might be good and cool, no idea. it will be hella weird though, of that I am pretty sure.

[cedar skiff]

I want to a see robot in the nba

[unique spade]

Personally i always found legacy society boring af.
I trained as a philosopher all my life

[turbid axle]

'legacy society' haha, what a term XD Elon deffo an effective meme machine XD

[unique spade]

Lol. It's just easier than to write "wars and all the dumb stuff happening cause humans still have their brains wired to stone age scarcity"

[gentle harbor]

is there any reason why i shouldnt use full access ? i hate it asking me to approve everything

[unique spade]

I am using full access from day 1. Never had a problem with it

[turbid axle]

war is simply physics. ain't nothing to it.
space and time are a limited resource. when there is conflict, one will win over the other, like it or not. ai won't solve that, nothing will.

probably what ai will do, is win.

[blissful basin]

i need some codex stickers so people cant see in public im using mac 🙁

[unique spade]

That's some middle ages world view you have right there. But this is not a philosophy channel. So we can leave it at that

[half frigate]

would be nice if codex app could preview jupyter notebooks in-app better so i don't have to open in an external app. is there any fix besides waiting for update or ditching ipynb?

[unique spade]

Don t think codex ever degraded in reasoning. Not in my experience anyway

[half frigate]

you know what? i should have asked codex first 🙃

[lost drum]

Yes cause my rule is if it works dont tauch

[unique spade]

I am just adapting codex to my own needs. Advantage of it being an oss harness

[turbid axle]

when I type in my prompts for codex, I better make sure only the effective words survive as not to waste my tokens.

[blissful basin]

There used to be some issues last year, mostly before new model release, but past few months it is very reliable all the time

[cedar skiff]

much more reliable than claude

[unique spade]

Yea but nothing of the likes of Claude crowd complaining about months about it till it was acknowledged as such 🙂

[lost drum]

Codex is just better and will be

They progress faster and have better teams I guess

[blissful basin]

yeah defo, for me claude is unusable work wise

[unique spade]

Yea I'm 100% of the same opinion and I have my tests I do on every new other model to check if it's still the case. And so far it always was the case

[blissful basin]

and i gave it a real honest test for 2 months on 200usd plan, to have proper comparison, and it is nowhere close to openai pro plan

right now only other subscription to test different models and have some fun with them is probably gh copilot plan

[unique spade]

Claude is heavily trained like a lazy beuarocrat imo. Finish fast and doctor the narrative to look like perfect work was achieved 😂

[blissful basin]

yeah i call it "claude is amazing unless you know what you do"

[lost drum]

And they make users pay for it

[unique spade]

Yup i thought to test it at some time with a task in my trading project. I was convinced it made what I asked for, and 1 week later I had gaps everywhere in my db

Cause Claude instead of using the actual timescale db, just made some csv scripts of the same level I was doing with gpt4o by copy pasting 1 year before

So I said ok boy, you ll never gonna get actual work tasks from me. That's codex job

[lost drum]

Wow

[blissful basin]

last year i was applying to company (that im working for now). And task was to basically build app leveraging AI up to the fullest and just show them how much i can build in 3 days and explain decissions etc. Last day, claude decided to remove all db migrations and screwed everything 😄

[unique spade]

These days I'm having fun with gptpro & codex designing and implementing from scratch a db specialized for trading ontology data. And it's fun to do experiments with them following scientific method. In 3 days we got better latency than the timescale db which is the vanilla choice for timeseries data

You won't win on quant
Ai changes the game though. Cause you can win on qual. You just need to figure out the qual strategy that works for you and implement that. But that's a long discussion.

[glacial shadow]

Why do you choose to not implement commas and periods in your writings?

[unique spade]

Well that's the thing. Ai won't give you winning strategy, but if you figure a manual method to win. And understand what you are doing and why. Then you can automate that. Which is a way of scaling yourself to play 24/7 on 100 tables

I mean sure you can tell Claude to make you some "winning strategy" but then you'll just join the fish bank using that, since you gonna use some default outdated strategy, that the big guns optimized to eat 1 year ago 🙂

*and I said Claude cauze it won t have a problem with that. Codex is much more careful with giving you risky tools 🙂

[glacial shadow]

Somebody please, get that man punctuation

[frosty zealot]

I’ll donate here take this .

[unique spade]

Lol let the man write how it wants. You re not his English teacher :/)

[kind jay]

This ain’t gonna work lil bro

[frosty zealot]

5.5 today confirmed

[kind jay]

I got updated to 5.5

[frosty zealot]

Excellent the auto update is working

[unique spade]

So basically out of 4 million weekly users there are just like 10 of us geeky enough to daily chat here including a few openclaw bots? 😂

[lost drum]

? Where

[turbid axle]

gpt3 confirmed

[kind jay]

Results of testing?

[kind jay]

Not yet

[turbid axle]

damn

get on it gaben

[half frigate]

are you saying you're modifying your own codex app? shoudl i have set up through the github repo instead of installing from their website if i wanna be able to do stuff like this?

[dense zealot]

nah, its old, today gpt 6.7 confirmed 😎

89% in ARC-AGI-5

[kind jay]

Go all in, $30,000 on the strategy

[frosty zealot]

Me, I confirmed it

[exotic cave]

There are no official information about it...

[frosty zealot]

Well, I’ll be here to say I told you so

[exotic cave]

Well i hope it happens, but many have claimed many things that never happened...

[frosty zealot]

I’m gonna make an AI that watches the road in front of me, if anybody drives slow in the fast lane, it automatically calls the police and reports them for drunk driving and flailing a gun out their window

[glacial ginkgo]

I submitted a feature request for man pages to the GitHub Issues list: https://github.com/openai/codex/issues/18990

[kind jay]

That sounds illegal

[plucky halo]

oops

[frosty zealot]

Looks good, push it

[turbid axle]

gPP confirmed

[frosty zealot]

Is this AGI

[idle sun]

Working for 45m 30s bro makign a nuke

still at step 1 from 4 btw

[turbid axle]

what I am wonder is why its called aguy and not agirl, pretty sexists if you ask me

my process is honestly ridiculous

I ask gpt to come up with issues in linear

then I ask gpt to implement said issues in my codebase

what even am I anymore?

[hasty gale]

everytime codex is analyzing my project it get stuck at "working" i have to stop him and say "continue" to wake up

[idle sun]

yea true

or do it

the app crashes too

so much recently

Error running remote compact task: You've hit your usage limit. Upgrade to Plus to continue using Codex (https://chatgpt.com/explore/plus), or try again at Apr 29th, 2026 3:50 PM.

one prompt

[lost drum]

then pay

[idle sun]

first prompt

how

i aint paying

[plucky halo]

lol - free user using extra high.... of course you're going to hit your limit quickly

Use medium at the most

[glacial shadow]

free plan isnt a plan, if you like the service, spend 100$ be done with it

[bright swift]

remove free plan and increase pro limits kthxbye

plus too

and mini-pro

[frosty zealot]

The ability to act surprised by this is crazy

[glacial shadow]

plus isnt for daily coding, it was developed for the chat website

[bright swift]

still uses my coding compute

[glacial shadow]

if you want to code daily, either spend 800$ per day on a private developer, or spend 100$ gpt plan

by my calculations, the 100$ is cheaper

[bright swift]

i have 2x pro 200 and will need more once my orchestration system kicks in

probably 1-2 pro/day

[glacial shadow]

thats a lot, idk how you fill that much usage

I cannot run out of limit on 200$ plan

youre allowed to buy multiple accts?

[bright swift]

5.4 xhigh only, super strict linting, code coverage, mutants, proptest, constant audits/reviews

[torpid trout]

When installing plugins with /plugins in codex it puts them in a ../.codex/plugins/cache/openai-curated/build-macos-apps/27651a43bf55185d924f7a1fc49043a0a8be65a0/ path, yet both doc and their plugin readme's state they must live in a /plugins/plugin-name folder
Which is correct?

[frosty zealot]

yes

[frosty zealot]

Are your plugins not working, or are you just generally curious

[glacial shadow]

the constant audit/review/test routines is great, while sleeping, or showering, or hiking, or driving

this tech is insane

[torpid trout]

I am asking what is correct, because if you where to create your own, for instance, you most probably would follow the doc, yet, the doc with these guys is often wrong, so you look at how the native approach is handled..
I have not even tried the plugins, at this point, just installed one.

I don't really think the intent matters a lot, the question is which approach is the right one I guess.

[frosty zealot]

I would say for personal creation of plugins, follow the docs, and for installing from the app, don't worry bout where they end up unless they're not working

But to answer your exact question, I don't know with certainty... its like one of those, if it isn't broken, dont fix it things lol

[unique spade]

Codex app and VScode extension are both UX over the underlying codex binary which is opensource

At

https://github.com/openai/codex

Codex engine (the one in the repo) comes with apache 2.0 license

And you can bolt whatever UX you like on top of it, and tinker with harness itself

Sure, you might break it and get worse results than vanilla

But in terms of what you're allowed to do with the harness itself there s not much limitation

Just don't touch oauth and app-server components cause that is how you re supposed to connect your account to the actual remote models

[potent mason]

Damn image 2 is cooking

Finally a decent way to do front end

[unique spade]

It is. It s super strong

And it helps reasoning a lot if you know how to use it

This is a hand written puzzle that the models including 5.4 before image 2.0 couldn t solve

This is de-noise made by image 2.0 and on this variant gpt 5.4 found the solution in 3 mins

[potent mason]

Nice

I'm guessing you can find a solution to 100 times more complicated ones if you go:

Concept puzzle -> Image puzzle with clear boundaries and squares -> Simulated puzzle

[unique spade]

You because when the image was messy and visual engine was poor

What gpt 5.4 ended doing was randomly zooming and wondering about all kind of irrelevant stuff

But once it gets a crisp image it can focus on the actual relevant problem space

Also speaking of front-end check this out

You will notice the 2 variants at the top share identical structure. And only some parameters were changed in a 2nd distinct pass

Also that image 2.0 was able to actually make a quite decent list with the Diff between the 2 (bottom middle image)

Right bottom diff map is hallucinated though, but that can be easily dealt with a governed ux/ui loop where you don t rely just on the visual engine but also use an analytic orchestrator

[glacial shadow]

Are you going to use this front end in a production environment or was this just an experiment?

cant tell if image blurry, or ai gen'd

[potent mason]

It's probably discord compression, image 2 generates at 2k, with 4k experimental

[frosty zealot]

Meeeeeeeeeoooooooowwwwww

Considering I thought 5.5 was dropping last week its been a long two weeks

[dusk thorn]

Finally tomorrow

[unique spade]

Just experiment. Wanted to see how it A/B and if it can also tell the changes visually between them in one shot

Old image engine was useless for anything serious cause it kept regenerating the entire image.

But this new one seems it can finally separate a fixed strcuture from variations on top of it

[lost drum]

or an arrow

[frosty zealot]

I'm able to?

on the iOS app at least

[harsh terrace]

hey guys , which codex modem u use mostly on vscode ? js typescirpt dockerizing backend/frontend apps ?

[gentle harbor]

does anyone know why these blue links do nothing ?

[blissful basin]

are you on windows?

[gentle harbor]

Yes

[blissful basin]

for me it was happening when i wasnt running codex in wsl, so there was some pathing issue

[unique spade]

Try to work around it. Have the image engine using well defined types for structure, so later codex can see those clearly and know to use them as invariants when writing the front-end code

[unique spade]

Powershell is still awful for codex. I had it figure out some paths yday and it took a lot of tries.

Wsl in comparison it s 10 times easier for it

[lost drum]

no hallucination at all for my case

[exotic cave]

Any official report on this?

[blissful basin]

yeah i was working on linux mostly, zero issues, windows however first of all was stuck in a loop very often with powershell. Now using mac and so far no issues

[gentle harbor]

what about the super app ?

[blissful basin]

but windows feels awful for any AI related apps

[gentle harbor]

seems like open ai cares more about mac then windows

you would think they care more about apple but there are still no codex ios apps

would be amazing if i could connect the codex mobile app to the desktop app so i dont have to stay at the pc

[blissful basin]

i think it comes from that a lot of AI apps are not really well optimized, and if you develop for mac, you just build for 1 architecture, you dont have to think about different user configurations

I personally really dont like this approach, but also understand it

[gentle harbor]

at this point i wish nvidia went that route as well and made there own os just like apple, windows is very close to being slop

[blissful basin]

I develop multiplatform apps in work, and most of the issues we have are that it doesnt work on mac, but works everywhere else, but if it is working on mac, it works everywhere else

(web apps)

[solemn acorn]

macOS has a POSIX shell, easier to build codex for

on windows you have to use WSL for that

[gentle harbor]

seems like the next step would be a ios app as well right ?

ive been waiting for that, maybe when the super app comes out

[cyan gyro]

at least we now know for sure, model updates are always on a Thursday

[hard drum]

i moved to macOS because windows+wsl2 was a whole other headache that casued more pain, && not to mention that the same computer hardware lagged on any linux distro thrown at it natively, despite being bog-standard laptop hardware

[blissful basin]

exactly same pain here

[hard drum]

the only time linux WORKED for me was a Lenovo Yoga 7 laptop from 2022

go figure

i mean, if windows+wsl2 fails, then try linux, && if linux fails, macOS it is

never been happier lol

UNIX-based system w/ great ecosystem + POSIX support is enough for me

also the fact that FL Studio is only for Win/Mac but not Linux

of course, i'll migrate my stuff to LPX once i get down to using it because LPX has a few neat things built-in that i need for my sound-sample manipulations

Alchemy being one of them

FL's sampler is kinda bleh

anything that needs FL plugins though... i'll see if i can find macOS versions, && if not, remains to either be a bridge, || stay there

[lean lark]

I haven't had one single issue with WSL2. YMMV

No company should deploy updates on a Friday. If something goes wrong, staff has to work over the weekend.

[hard drum]

Mine was like this:

use for a few hours
oops, i disconnected from wsl2
try reconnect...
whoops, suddelly your debian instance cannot instantiate
reboot pc becuse why not
oh, now it somehow works again
oh god, it stopped after like 40-ish minutes
can you not?
ok sure, reinstalling linux proper native
oh no, linux laggy
back to windows

solution? not windows reinstall, not another linux distro, but... macOS

[lean lark]

I totally understand and I absolutely do not doubt your real experience. I've had issues like that with other tech too.
However, if WSL2 was so bad right from the get-go, it wouldn't be "a thing". I suggest something went wrong with your specific environment, and a change in operating system was a bit of an excessive leap for a solution.
We plainly see people saying the same thing about AI : "I tried it, it said something weird, it's evil, I won't use it, no one else should either."
We can expect that from the common "user" but developers need to look deeper. That said, no one, especially developers, should need to look deeper to solve what was for you a really bad first experience. This MVP crud that we feed one another is horrible the enshi******n of software and services is real. We need to stop producing it, and tell others we don't want it. Because that won't happen, we need to learn to deal with it. 🙁

[frosty zealot]

https://x.com/OpenAI/status/2047008987665809771?s=20

[lean lark]

☝️ Business, Enterprise, Edu only

[hard drum]

and a change in operating system was a bit of an excessive leap for a solution.
Yet I've never been happier than before. Seems like it was an overall good decision. I still have a 2nd older MBP specifically that runs macos/windows10 ltsc thru bootcamp for some legacy stuff && dev testing on x86 mac/win equally, && if i really need to, i got a 2011 dell workhorse of a laptop to pull out that runs w8.1

Something cannot be wrong with env truly if a complete re-install of Windows 10/11 produced the same outcome shortly after.

[turbid axle]

this is essentially a wrapper for codex automations I guess

[hard drum]

I want convenience && "it just works", not "dawai, time to go ask Microslop to look into why WSL2 refuses to work on my machine, except I will likely never see the answer for it because they're too busy dumping their OS with Copilot"

[stable obsidian]

When 5.5

[cyan gyro]

Tomorrow

[stable obsidian]

🤩

[turbid axle]

probably