codex-discussions

[hard drum]

gotta wait 6h 47min

[warm pilot]

My package is schedule for Monday

[oak trellis]

which package you have?

[hard drum]

pro 20x

[cedar skiff]

fair effort, you have to have 2 agents running almost 24/7 to hit the limit with double usage

[latent tiger]

Looks like we’re going to get both opus 4.7 and gpt 5.5/spud today

Big day

[cedar skiff]

how can you tell?

[latent tiger]

Last time open ai released like 10 seconds after Anthropic lol

[fleet geyser]

I wonder how good opus 4.7 will be

But I think the new gpt will be better

[latent tiger]

Benchmaxxed for sure

I think it will be very close 1-2% difference in benchmarks for both

[fleet geyser]

But if they release 4.7, all the hype around Mythos will go away

[fleet geyser]

Yeah

[cedar skiff]

it will be distilled from mythos

[fleet geyser]

but the benchmarks aren't as relevant for everyday use

[latent tiger]

Yeah I was going to say 4.7 is probably a sub version of mythos without all the hacking capacity

[fleet geyser]

opus had higher benchmarks yet now people are talking about how bad it really is in general for simple things

[fleet geyser]

The hacking capability was just a benefit of being able to code really well

[cedar skiff]

that's true, i mean glm and minimax bench close to gtp and cluade....but in reality they are only close on easy tasks

[latent tiger]

It’s just guard railed to the max they don’t nerf the general coding

[fleet geyser]

new gpt will have more computer use

[latent tiger]

They just train it to detect if anyone is trying to do anything bad and deny your task

[fleet geyser]

the inbuilt computer use wasn't that cool tbh, a tool like openclaw could have done more in shorter time

[latent tiger]

Openclaw dev works for oai now anyway I’m sure he’s going to get codex to be just as good for computer use in this or the next release

Futures looking promising

[fleet geyser]

It's already really good

[fleet geyser]

I'm looking forward to it!

do you think they'll go 5.5 or straight to 6

[latent tiger]

5.5 for sure

6 is an entire new base model

Probably end of year

[fleet geyser]

they've finished pre training

so now it's finetuning/getting it ready

[latent tiger]

Yeah post training takes time

They also need to internally test/hold it for competitive advantage

The lab always uses a generation model above the public one

[fleet geyser]

by the end of the year the whole industry would have changed standards

it's changing so fast

[latent tiger]

Open source is also really good now so I don’t think they can afford to hold back their frontier models too long now

They’ll lose market share

Kimi minmax and glm are just as good for coding than gpt or plus more or less

5% difference

[fleet geyser]

But still require lots of hardware to run locally

They'd be making more money from everyday consumers

[shell coral]

how much estimated dollar value do you get from plus plan usage limit assuming you hit weekly limit everytime? (what you should've paid if using Pay As You Go instead of flat subscription)

[velvet wren]

I have worked out it to be between $300 and $500 if you'd use the API

[shell coral]

is this when Plus plan still have 2x limit ?

[velvet wren]

no

Plus is unbeatable value for money

[high girder]

I use the $20 plan and get almost as much out of it as I do the claude $100 plan

[shell coral]

does the scale go up on higher tier?

currently I'm on go plan (since I got it for free from promo), I get about 6 USD dollar value on the weekly limit, so the value is around 3-4x the subscription price (8 usd/month)

So on Plus its minimum 5x the price? how about the 100 usd?

[velvet wren]

each of the ChatGPT subscription plans offer much better value than using the API

[prisma flare]

idk but if you use 5.4 pro often it could get really expensive via api

[dawn seal]

I'm so glad the codex exist

I used to not be able to clean my house, not able to walk my dog, cannot take shower because I'm glued to computer

I can live like a human being now

[shell coral]

I just want to understand how they scale on each plan

[velvet wren]

it's simple. the subscriptions are a fraction of what you would pay if you were using the API

[oak trellis]

wuick question .. the context window possible to make it bigger on gpt 5.4 its so damn small its annoying

[boreal holly]

Not only that, once you run outta codex, if you have macOS ChatGPT you can connect it to terminals and IDEs and scrape by with practically unlimited ChatGPT (very small ctx window but better than nothing)

totally solid backup solution

[cedar skiff]

you can also connect gpt to your github and throw a few audits at it using gpt-pro

There is so much value

[boreal holly]

Yeah I know! I have Pro & Plus (one for work one for me). I have never ever been rate limited on normal ChatGPT stuff on Plus. I voice chat with it every morning on my 45 min drive to work about coding or whatever. Codex might be constrained usage but there's so much more to the subscriptions. API is spendy no matter how you cut it

[cedar skiff]

I meant the pro model* not the plan 🤣

[boreal holly]

Actually I've had trouble with gpt pro not being able to use the github connector. It tries to run python commands in a container. I usually have gpt 5.4 thinking read the repo, and switch to pro model to evaluate what it read

[cedar skiff]

Oh now that i think about it i did too, i must have just used gpt 5.4

[boreal holly]

Yeah, and with the pro plan I think ChatGPT has 128k context so it can read quite a bit. I have it read some parts and describe in great detail, then switch to pro to do planning

[frail meadow]

How’d you guys get codex merch?

[plucky halo]

solving the puzzle in the superbowl advert

[frail meadow]

Still possible?

[plucky halo]

No, it was closed within a few hours, I think

[oak trellis]

ser bigger context window you found a solution ?

🔄

[cedar skiff]

I still havent tried o.0

[rocky fog]

also wondering
plus 256k context
pro 20x also only 256k context
is bigger context only with API?
(vscode extension)

[prisma flare]

you can set it in config.toml or in each agent config

[boreal holly]

Updoot: https://github.com/openai/codex/issues/18130

If they put this feature in, it would mean you can get input caching bonus for up to 24 hours

[iron helm]

Anyone else experiencing issues with Codex usage credits? I've reached weekly usage limit, the additional credits which were near 1000 are now showing at zero overnight with no activity and i also have auto top-up enabled. So there should be absolutely no reason I am still getting this message: "You've hit your usage limit. Upgrade to Pro..." Is the credits/billing system down?

[torpid trout]

Unpopular opinion, but if AI ever wants to replace me, it needs to get its 🤌 out of the 🍑 because rn, it is ... just not mature lol
Gpt 5.4 just magically added a select around an actual table entry - so, instead of making a row selectable in whatever other magical way, it decided that one random column value in a row shall be the selectable button
because
"you are correct, the current implementation is wrong, my bad"
🤣

upon which, it went ahead and repeated the error but with a more smooth style
Meh

[kind jay]

AI is not going to result in some huge replacement, it’s a crazy narrative

[torpid trout]

thank you

That is what I am saying since this hype started lol, I have even redditors setting reminders to my reply saying "in a year you will be gone" and stuff lol
I mean, hell, yes, a vibe coder is probably blazed by what it can do, and I am also impressed, but... no, it is not going to replace me just yet neither in a year unless as said, fingers and peach

[cedar skiff]

It's just matter of tooling and time

[frail meadow]

Did you make a plan and read it? Normally if I get clear on requirements this type of stuff doesn’t happen

[torpid trout]

Oh, right, the skill issues 🤣

[frail meadow]

Just saying 🤷‍♂️

[torpid trout]

I am not talking about skill issues. I am talking about the very nature of the models:

• gaslight
• lie
• pretend
• invent

No skill solves that. Only human honesty and reliability does.

(not that humans do not do that... but you know what I Mean I guess - models do it as a daily excercise lol)

[frail meadow]

I don’t think AI is replacing anyone anytime soon, however I do think that people that don’t use the tooling will get left behind. And plan mode is like agentic development 101

[torpid trout]

Like, try to make gpt 5.3 use bootstrap - or refactor a css to use it
It will literally over and over tell you how it refactored it... only to see that it added new custom css that does what bootstrap does with a simple class
Lol
Then you confront it... it goes "you are right I will fix that"... only to add more trash haha

[torpid trout]

I absolutely agree with this.

Not saying it is not making work faster, and often also better.
Just saying that it aing replacing me anytime soon 🤣

[cedar skiff]

codex plans dont usually have many implementation details in them, it's all what not how. If you don't give it how in skills, forcing it to find exemplars etc, it will reinvent the wheel and not follow project coding conventions. read the plan isnt really a solution.

[frail meadow]

Maybe it requires the read between the lines human skill, not sure

[torpid trout]

my man, if I have to tell it to use col-8 instead of (given custom flex css or whatever it prefers), and make example of how that looks, I can as well do that myself, because that is not the issue - it knows damn well how that looks
It however skips deliberately the command, and adds custom css, because for some reason, it has been trained to do that

[cedar skiff]

Are you talking to me?

[torpid trout]

it simply has certain lazyness and especically the lying that is enfuriating and ... totally anti-replacement lol.
no one complains if it can't... but lying about it?

[cedar skiff]

I'm talking about over arching architecture. Like in exisiting brownfield projects.

[torpid trout]

what is brownfield?

[frail meadow]

Existing codebase not starting from scratch

[torpid trout]

Ah. No, I observe the lying and incpomliancy in both brown and other projects

I am not sure requesting bootstrap over native css is "overarched", but I might misunderstand the word here

its like a hit and miss, and once it missed, it holds on to whatever it did and starts being very 4-year-old-caught-with-candy-ish 😄

[cedar skiff]

I mean human written code bases have policy and architecture that the people who work in it follow and understand. It spans layers and keeps things understandable and maintainable. LLM s tend to come in and not follow that. That is what ppl call slop. When a person does it it's called spaghetti code. They just write things their own way and you need to put in a fair bit of work to get them to adhere to the code base conventions and architecture.

[boreal holly]

I think that folks have to design their LLM harness so that "policy" and "architecture" is non-optional. Yeah if you give a LLM the same power as a human user and try to craft skills, AGENTS.md, plant hints all over the codebase to guide the workflow, there's no guarantee it'll be followed. For example, if you have a policy that says "you must create a worktree with a non-integration branch and work from it", a few possible outcomes happen:

1. They follow the instructions, create a worktree, make their edits from there
2. They create a worktree, and make edits in master repo
3. They don't create a worktree and make edits in master repo
4. They create a worktree + branch, check out master repo into the worktree and make their edits

I've had this happen countless times. It's like no matter how unambiguous you are with the instructions, you roll a 1/20 and get the dumb LLM. And if you don't catch it immediately, that agent will follow only the worst practices for the entire session, so you gotta archive them and start from the beginning. There's no recovering a misbehaving agent if they're allowed to autonomously misinterpret stuff and do things incorrectly.

The solution is sandbox. Set up rules so certain commands fail hard with justification, so the moment they drift they're immediately realigned. Create the worktree for them when they spawn and force their write access to the worktree, get rid of git checkout. Make it so the correct way to do it is the only way, and assert these facts continuously at the first sign of drift. Then all of a sudden LLMs work with surgical precision at all times

[latent tiger]

Now we wait for GPT 5.5/spud

[plucky halo]

This suddenly looks different in codex:

[wintry gorge]

codex down?

[cobalt junco]

does anyone work at a company thats closing down. I need to buy a macbook pro for cheap

[mint reef]

it's better to be down. to reset limits...

[cobalt junco]

probs today

[latent tiger]

1pm usually when open ai release

[cobalt junco]

I thought it was 10:00 AM Pacific Time

[latent tiger]

1 PM EST sorry im in nyc ahaha

[cobalt junco]

yeah its 6pm for me ldn time

[latent tiger]

opus 4.7 is confirmed a nerfed mythos lmao

[cobalt junco]

i think we got like 1hr and 48mins

[cobalt junco]

Its pretty much opus 4.6 but back to normal and not degraded

[latent tiger]

i dont see why they would gatekeep their stronger model shooting themselves in the foot

[cobalt junco]

bc they hypocrites

[steady vigil]

"Instruction following. Opus 4.7 is substantially better at following instructions."

this is probably the most interesting part (if true) anthropic is back in the game.

[latent tiger]

they're just going to get mogged by the chinese open source in 3 months anyway

lmfao

should be like OAI and release frontier asap

[cobalt junco]

they talk about AI safety and security and what not then they just hand the keys to banks and corporations that do not have value for public interest

i genuinely hate anthropic they’re such shills

[latent tiger]

give people access to strongest tools to reinforce their codebase security and prepare countermeasures for threats

[cobalt junco]

like if they meant what they say I’d understand the gateway of mythos

[latent tiger]

it's okay i have faith in spud

[cobalt junco]

spud for the win

[latent tiger]

it will probably mog mythos bench anyway

[cobalt junco]

i hope so

[latent tiger]

imagine how huge of a L it would be for spud to be more powerful than mythos and just get a full rollout 😂

[cobalt junco]

5.4xhigh still pretty good tbh

[latent tiger]

huge L for anth that is

[hollow reef]

time to wait for GPT 5.5

[latent tiger]

1h 45m

[hollow reef]

LOL

Like the good ol times

[cobalt junco]

thats if they release today

[latent tiger]

https://klipy.com/gifs/waiting-rabbit

[steady vigil]

I seriously doubt it. they may have to rush gpt 6 though

[cobalt junco]

Which i think they will

[hollow reef]

I don't think they're releasing anything today

[cobalt junco]

usually they release just after anthropic to frame mog

[hollow reef]

Considering the 5.4 Cyber launch is still very fresh

"launch"

[cobalt junco]

5.4 cyber is gated

[latent tiger]

5.4 cyber is kinda cucked

enterprise only

[cobalt junco]

yea

[hollow reef]

Yea ik it's gated that's why I said "launch"

LOL

[cobalt junco]

5.5 acc gna frame mog

[hollow reef]

I have the max plan on claude 20x it's terrible for usage

[cobalt junco]

i can feel it

[hollow reef]

SO bAD

[cobalt junco]

I got 2 pro plans on oai

[hollow reef]

I want to do that but my front-ends would look terrible

[cobalt junco]

use skills & codex

[hollow reef]

I tried

[latent tiger]

it might actually be next week guys

[hollow reef]

Even Opus is very bad at frontends until like spoonfed

Codex would be even worse imo

[latent tiger]

they will prob release the new image model today instad of gpt5.5

[hollow reef]

If you wanna send me a few skills I'll give it a go ^^

[cobalt junco]

ye one sec

the reason claude is better at frontend isn’t the model

its this skill

https://skills.sh/anthropics/skills/frontend-design

[hollow reef]

I have that skill

Codex bombed LOL

Same prompts similar steering, Opus still edged out by a good amount, if GPT did everything tho I'd go for 2x Pro on GPT 100%

[ornate jolt]

opus 4.7 released now we wait

[latent tiger]

2% chance for new model today gg rip

next thurs guys

[ornate jolt]

just dont care for going back to anthropic they dumb there models alot for the benchmarks to look better for there newer model

[cobalt junco]

some ppl saying first week of may

[hollow reef]

Surely tibo is in a silly mood and he says "hey haha guys GPT 5.5 today xD"

[ornate jolt]

haha

gpt 7 release skip gpt 6

[hollow reef]

GPT 5.7

[ornate jolt]

gpt 13.6

cant wait for gpt 6.9

[cobalt junco]

besides AI

anyone enjoying this new season of the boys?

uncanny how it overlaps with current politics lmao

and AI

[ornate jolt]

stopped watching modern shows due to them always having to include some type of politics and not even in a funny way

[white furnace]

I cannot seem to find the run actions anymore. these only pop up for 1 project but for another project they dont show up yet i can see it in environment... there is default run action. anyone got any idea? i feel kind of dumb

[ornate jolt]

must be a sign

gpt 5.8

[ornate jolt]

what i hope with this release...please for the love of god make it have better front end

stop with the damn boxes gpt

you think its gonna be here in 30 min? someone gonna keep track?

[latent tiger]

lmfao

GPT 5.4- Codex???

Codex desktop super app update?

vague postmaxxing

[cobalt junco]

Release window for 5.5 closes in 25mins

[ornate jolt]

for me its 5 min

least sense that 30 min post

for context whos tibo

[potent mason]

He’s the reset dude

[ornate jolt]

like the reset our limits?

[cobalt junco]

against 10:00 AM Pacific Time

not that post

[cobalt junco]

yes

he the one with the button

[ornate jolt]

the nuclear button?

[cobalt junco]

i wonder who presses the release button

[cobalt junco]

no just the reset button

[cyan gyro]

curious what their answer will be for 4.7 opus

[ornate jolt]

kinda love the fact they release when anthropic does it makes it so we dont feel left out

anthropics other updates are useless they have so many options in there app to burn your usage in 30 min

[potent mason]

Well I mean I have both subs so it’s nice to get both in a day

But from what I gather Opus 4.7 is just kinda useless

[ornate jolt]

i figured they dumb down there models for there releases

but whats your experience?

[potent mason]

Haven’t programmed today but just being in the chat people are saying it’s almost like Opus 4.6 pre nerf

[ornate jolt]

yeaa i figured

thats prob what it is to

[cobalt junco]

current time 6:05 PM BST

5mins past the window

[ornate jolt]

5.5 codex?

starting to think these ai companies are secretly just milking us for model releases

and there friends in real life

[cobalt junco]

they could be doing a rockstar games

[ornate jolt]

gta 6

[kind jay]

When lil bro?

[ornate jolt]

in about 5 days

if its just 5.4 codex ima be mad

[hollow reef]

I won't !

5.3 Codex is genuinely better at some stuff than GPT 5.4

[ornate jolt]

i want 5.7

[hollow reef]

It's just a pain

[ornate jolt]

i want bigger number

[kind jay]

April 22nd, you better be right lil bro

[plucky halo]

Genuinely love the popout

[deft gyro]

When gpt 5.5

[ornate jolt]

gpt 5.5

[deft gyro]

Yurr

[ornate jolt]

never

[torpid trout]

why usage limits so little

[nocturne folio]

drop the codex, just

[kind jay]

April 22nd or 23rd

[cobalt junco]

next week boys ig

[kind jay]

Looks like Wednesday

[nocturne folio]

within this year

[ornate jolt]

i PROMISE it will release between now and 2037

guess ill play with opus for now

[sand moss]

Hey, how is 5.4 Cyber then?

[mystic dome]

Hurry up spuddyboi

[main nimbus]

waiting (inpatiently)

[latent tiger]

dont think any spud today

new codex desktop app update

[latent tiger]

i knew it superapp kekw

spud next thurs cya then

[cyan gyro]

https://openai.com/index/codex-for-almost-everything/

ony for mac

[severe mason]

but is it geo locked? not rolling on eu?

[sullen swift]

what

[deft gyro]

WHY IS IT ONLY FOR FLIPPING MACS

WHY

NOT

MAKE

A LINUX APP

[ornate jolt]

ew

[potent mason]

Oh thats funny

Thought GPT 5.5

[ornate jolt]

i wish

[cedar bear]

Why Not Windows????

[hard drum]

day 0 since another linux user cried about not having apps for a system most used headless in server environments

[cedar bear]

Why only Mac

[hard drum]

get a room, penguin

[worn harness]

my cope is that they are working on better frontend design for 5.5 and waiting on it

[deft gyro]

Who cares about windows

[cedar bear]

Some do lol

[warm pilot]

is the update available for y'all?

[deft gyro]

Well it’s lowk fine because mac apps are easily translated to linux apps

[frail meadow]

New update seems like a whole lot of nothing. The memory preview sounds interesting if it’s good and better than Mastra’s observational memory

[deft gyro]

I use the mac app on my linux system

But it has some annoying quirks

[solemn acorn]

easier to build for mac

[hard drum]

unless those mac apps used Swift && apple-specific UI stuff

[solemn acorn]

I would like a linux app though

[hard drum]

make it yourself then

you have codex CLI

[storm bay]

Lol Opus 4.7 came out so they had to ship something

5.5 not ready yet i guess

[deft gyro]

Thats not how it works

[solemn acorn]

not that simple

[storm bay]

i’m joking

[cobalt junco]

booooo

[hard drum]

isn't the whole linux community all about making your own things, like distros

[deft gyro]

Some people unironically think this, sorry

[deft gyro]

No

I use arch and gentoo btw

[solemn acorn]

codex desktop is closed source, I can’t just fork it to make things work on linux

[potent mason]

Developing for Mac is easy

[deft gyro]

But codex desktop is just a frontend for cli which is open source

Like

You could do it

It’s called t3code

[solemn acorn]

yeah but the computer-use stuff is exclusive to the desktop app

[hard drum]

well, linux users can just... make their own "FOSS" clone and expect it to die in weeks

that's kinda your thing, ya?

[deft gyro]

No

[frail meadow]

Just use playwright that’s not even a big feature

[solemn acorn]

playwright is different

that’s specifically for web-apps

[hard drum]

at least they don't have to trouble w/ Windows... sometimes

that one is a mess to aim for

but at least windows users aren't entitled freaks about it

[frail meadow]

Yeah better cuz you don’t have to use tokens every time you run it and can add it to your test suite

[flint carbon]

Does anyone experience increased error rate and loop with 1M context?

[cobalt junco]

just buy a mac?

[flint carbon]

Im doing academic work, mainly using it for orchestration role to spawn in agents but doesnt really work with 256k

[hard drum]

1m context? where? when?

[solemn acorn]

the 1M context model is worse so that’s to be expected

[flint carbon]

um yeah... who doesnt

[flint carbon]

Config

[hard drum]

imagine the power we'd have if 1m context was better

[hard drum]

works on codex cli?

[deft sable]

yes

[flint carbon]

yes but kinda not worth it

Burns too much token

[hard drum]

i've seem some charts for 1m context and it was worse than opus4.6 or even sonnet 4.6

sad...

so much potential if they figure it out

[solemn acorn]

we are slowly getting there, seems like it takes a lot of work to get models to behave with large context windows

[quaint sigil]

update out?

slow release or insta update

[solemn acorn]

anthropic has similar issues with their 1M context models

[flint carbon]

well i tried using orchestrator promps to spawn in like 6 agents working simultaenously

That cracks the context rot by alot

[hard drum]

wonder when OAI gets Agent Teams/Swarms equivalent on Codex

when i tried that a month ago on CC, it was genuinely really good--just ate a fudgeton of tokens though

[solemn acorn]

you can ask it to use subagents though right?

[hard drum]

that's not quite same

that's where the subagents can't tlak to each other, && can only report back their results to the primary agent

[distant bramble]

Anyone have a direct download link for the new update?

[hard drum]

but in swarm/teams, they can communicate w/ each other

[fossil bison]

I've noticed many minor bugs in the Codex, for example, connection errors occur during chat and no responses are received. The old chat was similar; when a question was asked, it didn't reply immediately, requiring a 2-minute wait after a sudden interruption.

[solemn acorn]

ah

I would assume that’s a WIP

[hard drum]

in CC, it is experimental

but very VERY powerful

[fossil bison]

@quasi canopy, is there a new update for the Codex yet?

[silver cliff]

hi

uhm

[daring yew]

wasap with the codex update is it mac only?

[silver cliff]

will GPT 5.5/6 come out?

soon

[fossil bison]

Sam

[daring yew]

cat, happy to meet u

[fossil bison]

Not to mention for Win

[distant bramble]

Anyone direct download link?

[iron helm]

Anyone having issues with tokens dropping suddenly to zero and auto top-up not working after weekly subscription quata is exhausted? Yesterday I had around 980 credits, did very little, then woke up this morning and have zero credits and there was an auto top-up charge on top of that as well. The "credits usage history shows "No credit usage recorded yet".

[opaque hearth]

hey

[solid mural]

Using gpt 5.4 xhigh /fast x5 to make gpt 5.5 ahh comment

[lean lark]

Asking "when" about OpenAI products is not productive. No one knows. The answers you get are only guesses. The data you leave with is no more valuable than the data you came in with.

[solid mural]

🤣

[solid mural]

Yo wsp

[lean lark]

Happy to talk about Codex here, dude.

[solid mural]

Fr

I really like codex

@lean lark

What are you making btw

[lean lark]

( I think this guy is gonna get banned soon )

[solid mural]

Fr

[solid mural]

Are you a plus or Pro user

Btw

[lean lark]

Non Codex discussion should be in #off-topic

[tawny turret]

why no spud model

today

when is openai releasing

opus 4.7 released today

[lean lark]

Asking "when" about OpenAI products is not productive. No one knows. The answers you get are only guesses. The data you leave with is no more valuable than the data you came in with.

[zealous ermine]

May be an odd question, but I'm wondering about the Apple Neo and its use with Codex and the new app control. Will it run OK or should I stick to a workhorse instead? Just wondering about the potential for a cheap laptop to carry around and work when travelling.

[solid mural]

Holy

[severe mason]

yo guys, my codex says: computer use plugin not available

any ideas?

[plucky halo]

Finally got the update!

[tawny turret]

?

[solid mural]

Just tell codex to fix itself

[solid mural]

4.7

[tawny turret]

its ahh bro

not good at all

codex 5.4 xhigh clears

[solid mural]

Fahh

[potent mason]

Yeah but it's actually absolute garbage

[solid mural]

I thought it will good but it's lame

True though 5.4 xhigh clears

Also more usage

Than claude

Also multiple model will are good

[severe mason]

i just signed in for the windsurf trial to test it and opus 4.7 was gone in less than 1 hour lol

[boreal holly]

It really depends. It's not a bad machine, but 8GB is pretty tight for dev work. 16GB is in my opinion pretty tight on a mac (the GPU uses the same RAM, so 16GB is partially wired down to non-CPU hardware).

I have a 16GB Macbook M1 Pro. I think it's OK. If you're doing web dev then the neo is probably fine, but if you're doing native apps you might feel a little let down or constrained.

[silver cliff]

does xhigh mean extra high?

[solid mural]

Yeah

[silver cliff]

got it

[solid mural]

[zealous ermine]

Thanks for your response - I was worried abou the 8gb issue as well, but the entry price is so low! I may just have to test it out..

[viral lake]

So codex is only an available for Mac?

[wicked oxide]

Nope

[solid mural]

Nahh

U can use it anywhere

[viral lake]

Why dos the announcement only mention Mac’s

[cobalt junco]

even intel chip macs?

[lean lark]

Hey Stu, wanted to mention - it would help humans and the model to lint your .md files in Sweeper. I discussed the topic with ChatGPT recently (always a good way to convey authority 🙄 ) and it tended toward confirmation (vague enough?) that well-formed markdown does result in fewer model tensions in processing text. In your case, without EOL between tags and lists, the lists bunch up. HTH

[solid mural]

Even in my samsung galaxy s5

Guys I have something i use gemini cli in codex cli

[wicked oxide]

Codex now has a plugin Image Gen, and it uses your ChatGPT subs to create images inside the Codex app. It seems they are using GPT-Image-2 model not sure but has higher quality and also takes a bit longer to generate .

[solid mural]

Gemini cli does frontend and codex cli does backend

[solid mural]

Not now i tried

[lean lark]

@plucky halo see note above.

[boreal holly]

You might be able to swipe an older M1/M2 model as a refurb! M1 still feels snappy to me. I used to do Xcode work with the M1 Pro 16GB and it's enough to run a single simulator + Xcode + Safari with room to spare. Running Docker container stacks and stuff is another story

[potent mason]

Oh wow that's insanely nice

[plucky halo]

Cheers!

[lean lark]

Codex is a single brand name for different products. The Codex App is available for Mac and Windows but Windows is frequently the child that gets hand-me-downs long after the favorite child named Mac. In other words, announcements for Codex App tend to be for Mac and don't apply to Windows. That's not the same for CLI or the Codex Extension.

[wicked oxide]

I tried it a little bit ago, and it’s working now. It may be gradually rolling out.

[zealous ermine]

great idea - thanks!

[potent mason]

Is it a skill or something?

[wicked oxide]

Damn, there are a lot of things in the new Codex update: memories, image generation, browser, computer use, and new UI changes. Now, waiting for GPT 5.5.

[lean lark]

Wow. "Codex - generate a mind-map showing top-tier modules in this app as nodes. Color them from green to red in order of increasing problems recently resolved in CHANGELOG.md."

[wicked oxide]

yeah its plugin .

[cobalt junco]

bruv

wth is ts

ima try reinstalll

nope

[hard drum]

heeeeey

musi-for-rust is out

ALSO

my first mac experince was with an M2 Pro mac mini

that 16GB of ram didn't last long

but am happy w/ the 32gb m1 max laptop

[wicked oxide]

oho no , Theo - t3․gg will gonna post a new crash out video on openAI .

[cyan wing]

what that?

[cobalt junco]

wut

[glass furnace]

I'm so confused. I'm in US, on mac, have the update installed, but not seeing the new "Connections"

[hard drum]

🙄

[hard drum]

musi for rust developers

[boreal holly]

🚀

I still use the M1 Pro 16GB for light tasks and remoting into the beast

[hard drum]

update your damn os fwen

26.4.1 is out

[frosty zealot]

Is it?

[wicked oxide]

https://www.youtube.com/@t3dotgg/videos just watch his first recent video . you will undertand what i am talking though really like his videos

[boreal holly]

I do not update unless there's a clear benefit 🤓 only reason I have 26.3.1 instead of 26.0 is the newer Xcode has better MCP integration

[sleek spoke]

anyone else on Windows not getting offered the Codex app update in the Microsoft Store?

[hard drum]

fair. on laptop, i get the 80% charge limit w/o external apps

so i say win

[glass furnace]

the ssh is why i updated but i dont even have access to that feature lol

[hard drum]

he's Linus Sebastian out of the Goodwill factory

he's only enjoyable for entertainment, not for actual source of truth

i take this guy's stuff with grain of salt, because he couldn't separate "this my favourite model yet" && "we need to talk about this model"

[potent mason]

Yeah pretty much the same as Opus 4.6 pre-nerf

[wicked oxide]

kind a though mostly he is true

[lean lark]

Announcement didn't seem to include Windows

[wicked oxide]

what .

[sleek spoke]

it does, just not the computer use

[hard drum]

didn't man praise gemini models but most users reported them being absolute garbage disposal?

lemme re-check

yeah, this guy is full of horse radish

why trust this guy?

[kind jay]

Maybe for you

[hard drum]

You realise how stupidly sycophantic && apologetic that thing is? the only thing it can do well enough is web search && frontend UI to an extent

[wicked oxide]

Maybe but not for me

[kind jay]

Maybe for you

[hard drum]

gemini tends to, no matter what instructions, to scaffold/placeholder when explicitly not asked, && then it starts doing it more aggressively

[potent mason]

Yeah of course for me, I can't talk about others people experiences. But from what I've seen in the Claude discord server + other social media the general consensus from people (not early testers who are biased as hell), is that it's the same as Opus 4.6 pre nerf, and it consumes way more tokens

[strong jungle]

anyone have a preference so far - gpt5.4 vs opus 4.7?

[potent mason]

Is this like a bot or something?

[hard drum]

haven't tried opus 4.7 but as a former opus 4.6, i'd say 5.4 is a little less lazy than opus 4.6, but gpt 5.2 is least lazy but not as intelligent

so... hoping good for a new gpt release

[rare dagger]

ssh still only available in the beta version?

[potent mason]

I think they said this new release has it

[wicked oxide]

Except for UI, I think GPT-5.4 is still good. The error rates on GPT are really low, and it feels like the difference is between someone more experienced and an expert who is not much experienced, hypothetically speaking. After doing some testing, both are really good and now neck and neck. But sometimes Opus does things wrong, analyzes the error logs, and then patches them. On the other hand, GPT-5.4 mostly never even hits errors. Damn good. Though GPT-5.4 is really bad at UI.

[rare dagger]

I just did a fresh install on both win&mac. ssh is still missing.:(

[lean lark]

I'm glad availability for Windows is clarified in a tweet cuz in the website product announcement it only profiles Mac.
Seriously disappointed at amateur marketing issues at this company. 🤦‍♂️

[rare dagger]

beta version works as expected.

[sleek spoke]

I saw a tweet saying SSH is only available on ChatGPT Enterprise plan for now

[glass furnace]

yeah i just realized

i saw alpha and i thought id get access because who is more alpha than me?

[boreal holly]

It recursively doubles down 🤣 "The user asked me not to scaffold placeholders. I'm going to scaffold as many as possible now, then ask for forgiveness, that way I don't have to scaffold later."

[rare dagger]

such an amazing decision they made

[glass furnace]

I'm messaging Tim Tibo right now

[sleek spoke]

so I'm gonna ask the opposite question, has anyone on Windows successfully downloaded the Codex app update?

@orchid plume hey!

[severe mason]

damm you actually need a vpn to use computer use from eu?

so sad

[glass furnace]

@crude mantle can you make tibo give me ssh access in codex because you gave me crabs?.. i mean lobster

[lean lark]

Is that "if" you are trying to evade EU legislation?

[rare dagger]

It forces me to download it from the MS Store. I can not see any new updates. The last update time is 4/15

[severe mason]

What legislation specifically?

[sleek spoke]

same

[gleaming garden]

same

[lean lark]

Well, when something isn't available "yet" in specific jurisdictions it's usually because of legal/regulatory concerns. So if the product doesn't support a feature yet, it's because it's not allowed yet, and therefore, the "need" for a VPN is only due to the desire to evade local regulations.

[gleaming garden]

i just have the 90+ new plugins availables

[rare dagger]

😆

[severe mason]

If I am allowed to use a vpn and the company allows me to use that feature through it, what legislation i am avoiding,

[sleek spoke]

same

[lean lark]

I won't argue product available with anyone. If a feature isn't available in a specific area, there's a legal reason for it. EOL/EOM/fini

[boreal holly]

Nah, you're not the one violating anything. The EU has certain restrictions on what companies can put as features in software. That's why the feature isn't available yet, but it's not illegal for you to get the feature over VPN. It's consumer protection stuff

[unique spade]

new codex app is just for mac right now right?

[sleek spoke]

no, scroll up for tweet screenshots

however... I cannot actually get the download to show up on Windows

[deft sable]

also on Windows except the computer use feature

[unique spade]

hmm ok cause i opened the app but i don t see anything new lol

[lean lark]

And that's the point - it's consumer protection, and some governments are more protective than others - which is sometimes good, sometimes not, we all just roll with it.

[sleek spoke]

not illegal for you, might be against the OpenAI terms of use tho

[hard drum]

lmao

point right, even if it was ironic

that's exactly how Gemini behaves, && i've had this shmuck since 2025

[carmine copper]

is it deterministic?

[lean lark]

"Create image of a car"
Is it deterministic? 🤣

Sorry... had to laugh.

[carmine copper]

look what I am responding to

compare the pictures

is that normal?

[potent mason]

Oh wow the photo gen is amazing

[deft sable]

Did it show up to you now? Just got the update upon clicking the Check for Updates on MS Store Downloads Tab

[potent mason]

I'm having it create favicons, app splashes, etc for a few apps

[boreal holly]

They did their due diligence in preventing the feature from being available in the EU.

The only two laws I'm aware of that the EU enforces are 1. App developers must provide their entire identity to consumers so they have a person to call when they're upset and 2. If the app uses encryption of any kind, a back door must be provided to allow government access to encrypted resources. They probably have more, but seeing as how SSH has encryption that's probably one of the things they need to get cleared.

[lean lark]

My point is that a request for "a car" is infinitely vague, so a deterministic response on that specific query is impossible ... unless it's always using the same random seed or caching, which is a different concept. I was just having a laugh.

[sleek spoke]

yes it did! thanks!

[carmine copper]

of course, I am just curious why I got exactly the same output as the other person

[lean lark]

Oh right, quite often there are just delays as the paperwork is sorted...

[carmine copper]

cache or what

[lean lark]

Oh, I didn't catch that. Wow, maybe they hash the request and provide a cached response. Not deterministic, just efficient in responding to an infinitely vague request that doesn't have a "right" answer. I'd consider that a feature. 🤔

[azure wyvern]

I updated today and now every time I give the codex window focus 1Password pops up saying Codex wants access to my ssh key. I don't want Codex doing remote git operations unprompted. Is there a way to turn this off? I couldn't find anything in settings.

[wicked oxide]

computer use with 5.3 codex spark is like so fast an snappy its like a human doing it fast and also really precise movements . really a good use case of this model .

[frail meadow]

Badge on front is different but the difference is minimal

[lean lark]

This is why I wait for the next release after alpha/beta/x.0. 😜

[orchid plume]

hi, and not sure, I saw an update but not sure if it's the update for today

[potent mason]

Hey guys just wondering, this isn't the new super codex app they planned on releasing right? The new super codex app is supposed to have things like a built in browser if I'm not mistaken

[sleek spoke]

it's OK I got it now 🙂

[solemn sundial]

To get ssh devboxes...

Add this to ~/.codex/config.toml:

[features]
remote_connections = true

Restart codex app, and then Settings > Connections.

[sleek spoke]

this does have the browser!

[boreal holly]

My guess, ChatGPT generating an image, since it's a chat bot it has temperature cranked up a bit. Since Codex is a coding environment they drop temperature down to 0. You don't want highly creative and random coding agents, and they probably carried that property over to image gen

[lean lark]

I swear, I can't keep up with OpenAI announcements of new features. It's a depressing aspect of modern light-speed development. Add to that all of the other stuff in this industry that takes so much time to grok. (Pun intended) It's a good problem to have. * sigh *

[potent mason]

Oh it does? I didn't see it

[severe mason]

My codex does not have a built in browser, does yours?

[sleek spoke]

it's in the view menu

[potent mason]

Amazing

Thank you

[severe mason]

aw, he is using chrome instead of the actual browser

[orchid plume]

I'm looking forward to when Codex's image generation tool uses the new GPT-Image model, not 1.5, but the new GPT-Image model isn't officially released yet despite it rolling out to some accounts on ChatGPT's website today

[rare dagger]

wow this did help!

[severe mason]

Codex says it cannot control itself

[lean lark]

☝️ inception

[glass furnace]

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

[sleek spoke]

I can't actually figure out how to get Codex to drive the browser

[severe mason]

me neither

there is this box on the settings that says alows-allowed apps which looks suspicious

[potent mason]

I can't install the 1password extension, or create new tabs.

I'm guessing it's not meant to be a full browser yet

[sleek spoke]

oh wait, I guess the "no computer use in EU/UK" also applies to automation of the browser....

le sigh

[severe mason]

It does not work through vpn neither, it can interact with chrome though

[sleek spoke]

what do you mean it can interact with chrome?

[severe mason]

If I tell it to do anything there, it works, but it does not with the built in browser

[frosty zealot]

I need that little potato

[severe mason]

it says computer use is not allowed to use the app com.openai.codex

[plucky halo]

I’m in UK and have computer use

[unique spade]

i'm using spark right now while waiting for my main qouta to reset

spark is insanely fast

but

lol, compaction is really bad for it, this thread basically is stuck now

due to that error

[frosty zealot]

Why is this even a thing

[deft sable]

I had this issue a few days ago. Apparently, it was related to their recent rollout, but I think it’s already been fixed now?

[sleek spoke]

how?

[unique spade]

well i just got it and i am on latest codex desktop app there

[plucky halo]

Using the 'chats' at the bottom of the sidebar

[sleek spoke]

oh wait, mac?

[plucky halo]

Yeah

[sleek spoke]

ah

[frosty zealot]

Any good use cases for the computer control?

[glass furnace]

i have it scanning the hub for appealing vids to save me time.

[frosty zealot]

LOL

'and use ffmpeg to crop out the best parts and create a compilation'

[glass furnace]

if it were claude, my account would be banned

[frosty zealot]

Overnight I've grown to really enjoy using Pi, and the fact I cant test 4.7 with it is just like why... why are they SO locked down

I wonder if the computer use will be better for smoke testing iOS simulator than using the xcodebuildmcp

[boreal holly]

idb is wayyyyyyyyyyy better if you want em to drive a simulator

[frosty zealot]

yeah xcodebuildmcp uses it under the hood i believe

https://github.com/getsentry/XcodeBuildMCP

[potent mason]

I've heard Pi is really bad oob did you customize it a bunch?

[frosty zealot]

Yeah you have to

But it can pretty much customize its self, and I added a few extensions from their extension store thing

dammit I thought they dropped a new coding model

[naive flower]

Opus 4.7 out. Bad timing for OpenAI to make the Plus accounts useless. Now with equal plans on both sides, which should I chose? Hmmmm 😉

[frosty zealot]

Whatever makes you happy

4.7 on the plus equivalent account will get you 1 prompt, use it wisely

[compact dawn]

computer use not in EU ://

[compact dawn]

how did they make plus useless?

i swear I run out of the 5 hr usage so fast now

on plus

like i know it's no longer 2x, but I feel like it's still much less usage than before the 2x month

[twilit bluff]

teaser

maybe they'll drop spud later?

[lean lark]

Stop using Fast mode, which consumes tokens 2x as fast, and you get the same output as when the 2x promo was in place.

[compact dawn]

i've never used fast mode

[naive flower]

It's probably a better prompt result. With codex I went to low thinking. It's a bit dumber, and I still hit the limits quickly. Don't know if it pays out to use low thinking, if you then have to reprompt to fix the problems...

[compact dawn]

how do you even toggle fast mode @lean lark

[lean lark]

Looking again, BRB

[compact dawn]

nvm i found it

yeah it's off

[lean lark]

For anyone else

[compact dawn]

it's like they nerfed usage so hard tho

[naive flower]

To my estimate, codex Plus plan is now 20% at most, of what it was before (not the promo 2x usage countet in).

[lean lark]

And I notice no difference at all. YMMV

[boreal holly]

Yeah my usage is way better after I turned off unified_exec. 3 hours later at 98% weekly after reset

[naive flower]

And business seat, I'm not sure, but could be even less than the Plus plan. It drains incredible fast down...

[compact dawn]

right. I burn through it in an hour or two now and I'm not even prompting a lot. Used to be that I could barely use it all in 5 hrs

[potent mason]

Is this a setting or something? What does it do?

[lean lark]

My workflow is to take time for a good prompt, get Codex to work on a well-documented project, then take time to review what it did before commit. That can take about 20 minutes to an hour per round unless I'm making a lot of small changes - which need to be checked anyway. I don't run into limits because I'm not asking the bot to do all of my work for me. Sorry to be harsh but this applies to more people than it does not. So efficient use of the tooling results in lower costs and almost no concerns about quotas. If you really have an environment where you rely on AI for really serious work, it's time to put some funding into the budget just like we do for all other business expenses, lights, licenses, etc.

[compact dawn]

what is this?

[boreal holly]

It is the background terminal feature. It let's agents run commands, and every 5 seconds wake up and check on the command. It's unfortunately the only way to run commands for more than 10 minutes, but it's also by default the most wasteful feature they have, so they made the tradeoff ridiculously over the top. You either get no work done, or you get no tokens and potentially no work done if they decide to terminate a healthy command which they do early and often

[lean lark]

To be more specific and less preachy : Use the tools to create well-documented projects, with in-line comments and a detailed docs folder. Instruct the model via AGENTS.md files or Skills to use those docs and to write to the docs on changes. With this, the model doesn't need to read every file on every turn (or thread) to understand the project and return really good responses. It saves time, money, and you have much cleaner projects as a result.
Use the tools for more than just writing your code for you. It's so much better than that.

[boreal holly]

I had to patch the unified_exec=false path so I could run commands for more than 10 minutes, and remove their timeout override (the agents can run the command for less than 10 minutes, so they usually start with 5 seconds, and rerun the command with increasingly higher timeouts until 10 mins is reached and then give up, wasting tokens the whole time).

[wicked briar]

guys do you prefer 5.3 codex xhigh or 5.4 high

[potent mason]

Really depends

[lean lark]

I decide per-prompt which model/reasoning to use. I tend to go back to 5.3 for simple/dumb text processing. I use 5.4-mini/low for simple doc stuff that I care about, 5.4/medium for normal work, and 5.4/high for more serious effort. I only go nuclear with xhigh for really deep concerns. 5.3 xhigh vs 5.4 anything? I dunno. Why bother, really?

[wicked briar]

like for backend work

like lets say optimizing the infra, load times, db reqs

[lean lark]

In that case you need to measure performance vs cost. So there's no general right answer.

By performance I'm talking speed + accuracy. We're talking about the Good/Fast/Cheap Magic Triangle here...

[boreal holly]

The quality tetrahedron

[lean lark]

Pick Two.

[ornate jolt]

Eh we got doctor science model instead of new coding model

[boreal holly]

I pick good and scalable

[wicked briar]

lol im pro cheap is not really issue

I can use like 50% weekly max

[lean lark]

Hahaha - I saw that and though "OK, we've been getting most of the goodies lately, time for someone else to have some fun."

[ornate jolt]

I suppose haha but I bet they are still gonna release a better coding model very soon

[lean lark]

If you're using pro then cost is not an issue, so go for quality. Just by the numbers, 5.4 is better than 5.3 ... and the Code tuning is built-in to 5.4, not a distillation. Go with 5.4.

I've been Very happy with 5.4. The reasoning ability is about 95% for me. But sometimes it just doesn't "get it". I can point to very specific scenarios where I know it's not going to do well, so I prepare to spend more time in those areas and avoid prompt tensions.

[mellow ledge]

Since the update i keep getting this error when i try dictation does some of you have the same ?

[lean lark]

Example: Play the telephone game. Write a prompt and tell it to optimize it. It'll find a ton of things wrong with it. Then copy the new prompt to another new thread and do the same. The model is Never happy with what you give it. It always finds something wrong. Then after about three-plus iterations, open a new session and tell it to compare the first version with the last version. Almost certainly it will find that the two are nothing alike. This has always been an issue and 5.4 still doesn't quite know when things are right.

[feral geyser]

Curious to know about special instructions in Codex that you all have found to be helpful. What special instructions are you using and how are they helpful? I've just discussed thread and special instructions in chatgpt 4.5 to generate the special instructions below. I discovered I was doing a not-so-best-practice by have certain threads be the auditor/project manager. This is prone to losing context of state and drift. I'd love to know your thoughts and strategies!
\
Here's what I have so far:

Always treat repository files as the authoritative source of truth.

At the beginning of each session:
Read PROJECT_STATE.md before taking any action.
Review TASK_QUEUE.md if present.
Confirm understanding of the current goal before proceeding.

Never assume project state from conversation memory alone.
Always verify against project files.

After completing meaningful work:
Update PROJECT_STATE.md to reflect completed tasks.
Move finished items to Completed Tasks.
Add new discovered tasks to TASK_QUEUE.md.

When proposing changes:
Check ARCHITECTURE.md if present.
Ensure proposals remain consistent with defined architecture.
If architectural changes are needed, propose updates to ARCHITECTURE.md.

Prefer modifying existing files over duplicating logic.
Avoid creating redundant implementations.

If uncertainty exists:
Stop and request clarification rather than guessing.

Treat commit boundaries as synchronization points.
Whenever possible:
Suggest commit messages.
Describe what files changed and why.

Focus on maintaining continuity across sessions by:
Updating logs
Preserving task structure
Minimizing hidden assumptions

[lean lark]

Another example: Starting with any piece of code, tell it you want to pivot to a different approach and you want to discuss how to migrate from v1 to v2. It will always get stuck trying to integrate v1 rules into v2 when they are no longer relevant. I'm not explaining this deeply but the point is that it always confuses old with new, doesn't "get" that things are radically changing. This is particularly true when discussing the UX for an application - sometimes it stuck in developer mode and simply doesn't understand what the app will look like to a user.

"chatgpt 4.5" !!?!?

[runic lake]

VPN

[wraith temple]

is it only me or codex is EATING those credits? i reached my 5h limit in less than 30 mins

[potent mason]

I tend to use GPT 5.4 for specs and planning, most of the programming tasks but when I’m reaching 30% I start using GPT 5.3 codex for programming out GPT 5.4 plans

[lean lark]

We need a hotkey to answer the same questions that come up many times per day. Maybe a FAQ ref? Pinned posts?

[kind jay]

Lowkey, you're job isn't tech support, if there are silly questions you are free to ignore them

[lean lark]

That's a punt, nothing wrong, but has pluses/minuses both ways. If the plan is less than ideal then a perfect implementation will have the same quality as the plan. If the plan is perfect and the implementation less so, then the quality again falls to the lowest value. Ya gotta try both options, come to your own conclusions, and then apply whatever strategy seems right at the moment.

[kind jay]

Hahaha

[lean lark]

I didn't say the question is silly. I said it would be nice to have a quick way to answer repeated questions. Huge difference.

[kind jay]

I'm saying it's silly

[boreal holly]

Is it just me, or does the 5hr limit go up if I wait 5 hrs?

[frosty zealot]

Nah seeing the same question 400 times a day makes you a little jaded seeing it again.. and again.. and again..

[lean lark]

You're free to ignore it. HAHAH

[frosty zealot]

Unpopular opinion if you're not on the $200 plan you shouldnt be able to complain about usage :>

[frosty zealot]

@kind jay haha roasted

[unique spade]

100 too now haha. it s still pro tier

[lean lark]

Never be afraid of unpopular opinions, yours or someone elses. Sometimes they're right.

[kind jay]

They going to release 5.5 codex by 22nd and pay for my Pro

[unique spade]

just have codex pay for itself by trading btc haha

if you dare, sure

you just make a prompt and say "codex, win me money" :))))))

[lean lark]

Oh, BTW, Codex does have a pay-as-you-go model. So you don't need to pay a minimum per month.
Also, you can just push Codex through your API keys and pay as you go there too.
I see all of these notes about subscription rates and tend to ignore them for this reason. That's not related to Pro model, just to costs.

( Me has been working on a financial analysis (investments/stocks...) system as time (never) permits. Why wait to win money when you can earn it? )

[frosty zealot]

Gambling is for the feeble and weak minded

[frosty zealot]

@samaltman

[kind jay]

??

[frosty zealot]

how do I ping him in here

[kind jay]

Like so

@samaltman

[frosty zealot]

must have us blocked

[kind jay]

Has you blocked

[frosty zealot]

clever

[kind jay]

There's no systems to get rich quick lil bro, you just buy a bunch of stuff at some arbitary time and wait

Unless you're a prediction market analyist like me that is

[frosty zealot]

Ya'll don't know what I know

[kind jay]

Already priced in lil bro

[spring remnant]

Why codex can't generate image

[kind jay]

@lean lark question for you

[feral geyser]

lolol #codexinsidejoke

[frosty zealot]

where tooki

[kind jay]

Hootki?

[frosty zealot]

google where tooki

[kind jay]

This isn’t google son

[frosty zealot]

lougle

[kind jay]

Hootgle

[frosty zealot]

Can anybody make their own prediction question on Polymarket?

[feral geyser]

So like, no one uses custom instructions? agents.md is sauce though 😉

*the good sauce, i should say

[kind jay]

Obviously not

[kind jay]

This is the #codex-discussions channel, please stay on topic

[frosty zealot]

Alright, I'm sincerely sorry

[kind jay]

I ain’t reading all that
@grok summarise and answer this question

[frosty zealot]

I just lost my hardcore minecraft world to a magma cube

[visual coral]

we need GPT 5.5

[solemn sundial]

Did anyone figure out how to get the computer use plugin to stop giving the error "Computer Use is not allowed to use the app ..."?

[kind jay]

What's even going on anymore?

[potent mason]

Simple math

[frosty zealot]

Amazing that you can't just do that off the top of your head like I can

[potent mason]

Basically easier than doing 1+1

[fleet geyser]

Interesting new update!

[kind jay]

True

They're making up funny words like Sudarshan-Glauber

[potent mason]

Well thats what happens when you get to advanced statistics messing with quantum optics

[kind jay]

Thank you chatgpt

[frosty zealot]

dont forget to carry the gleeper shneeben

[potent mason]

(Totally did not have to look it up)

[kind jay]

And factor in the plumbus

[scenic umbra]

Anyone know how to do the remote SSH thing? I see all the other features except that one

[sleek spoke]

search in here for remote_connections

[kind jay]

More context please

[frosty zealot]

Open your third eye

[scenic umbra]

Awesome thanks

[frosty zealot]

well after playing with 4.7 for an hour, it doesnt feel any more intelligent than 4.6 for my use cases

[lean lark]

Ask better questions.

"Define the universe and give three examples."

[frosty zealot]

You're right I have no idea what I'm doing

there I updated my original statement

[lean lark]

I just fired off another prompt like "find out what's wrong with you, fix it, improve on it, repeat" ... I love this stuff...

[kind jay]

Token wasting speed run

[lean lark]

No, really, it's good optimization ... and that's not my prompt. It took about 30 minutes to generate the right data and instructions to properly define the problem and expectations. I don't vibe this stuff. I engineer it.

[kind jay]

I’m sorry for being such a hater all the time, I just think it’s fun

[lean lark]

( I read ya... )

[frosty zealot]

She's acoustic

[kind jay]

“If you see a good move, look for a better one!” -Emanuel Lasker

[lean lark]

In our case we really do now have the ability to tell the machine "improve on yourself". It's an awesome time to be a nerd.

[boreal holly]

"I'm too drunk to taste this chicken." -Ricky Bobby or something

[frosty zealot]

LOLLLL

[kind jay]

I don’t know who that is

Actually I do

[frosty zealot]

Too bad A\ locks you into using their harness. There are much better more capable self improving harnesses out there, one of the few reasons I'd rather stick with OAI

[kind jay]

Harness? 🤔

[frosty zealot]

You trying to get us both timed out again

[kind jay]

I didn’t get timed out last time, also I just picked the wrong emoji smh

[lean lark]

Actually you can use something like the new Gemma 4 locally to do the same kind of thing. Just tell it to keep improving on documention that directs the assistant to improve on some specific thing, and eventually it'll work it out. More in line with this tech, the thing is not to try to get it to make things better but to make fewer bad choices ... that's exactly how this transformation ML stuff works.

[frosty zealot]

I document through skills mostly

They can be invoked dynamically, don't give unnecessary bloat to the code base etc.

More efficient with tokens I find

[lean lark]

I admit I haven't transitioned my docs files to skills yet, afraid to break what's working really well. Will need to do it at some point but I've been waiting for someone to say Skills are just a fad and we're moving on to some new architecture - that's bound to happen in a few weeks. 🙄

Like RAG...

[frosty zealot]

The models are smart enough to infer what your code base is doing, but there are some use cases where newer versions post date the cut off model so it tries to implement things based on old versions, thats where creating skills to guide it when touching X framework to make sure its grounded against X version etc.

My personal experience anyways

[lean lark]

You're right. That's always a concern.

[frosty zealot]

I definitely do not think skills are a fad, I think they're detrimental to anybody really wanting to have efficient models

[lean lark]

This is so weird, I'm watching the Codex primary process argue with its sub-agent. I need to get some popcorn.

It's like "you need to do this" ... "I did that" ... "no, you did it wrong, it needs to be like this, here, I'll change the docs" ... "oh, like this?" ... "no, try again..."

[kind jay]

Anything but write code

[kind jay]

Models are going nowhere until they introduce latent thinking lil bro

[sleek spoke]

this explains Codex not being able to drive the built-in browser

[frosty zealot]

I guess I replied to the wrong person above

[boreal robin]

How are you popping out the browser to use! I’m not seeing that in my app!!

[sleek spoke]

I found it in the view menu

also that employee is using a codenamed model "arcanine" in Codex in the video he posted

[frosty zealot]

great more furry stuff for @kind jay to yiff to

nine-tails woulda been cooler :>

[kind jay]

Honestly crazy to me

[frosty zealot]

lmfao

get wrecked

[mental urchin]

Why is Codex nowhere near claude code in programming or am i using it wrong?

[frosty zealot]

rage bait alert

[kind jay]

Skill issue probably

[boreal holly]

More on that, I let unit and integration tests be the authoritative description of the codebase. I noticed whenever I asked Codex to read the openai/codex codebase, one thing it would always do if I asked it something like "Do skill headers hot reload when edited each turn?" It would vigorously read the unit tests and be like "Yes, there are 3 tests that prove this behavior here here and here, and it shows the flow of execution like this, which matches the headers being injected upon change at the start of each turn." I'm like... OK, if the codebase is thoroughly and descriptively unit tested, there's no question about how things are supposed to work because the tests immortalize the behavior.

[mental urchin]

yo wym

like im being so specific what i need and still getting garbage results

[boreal holly]

So get rid of markdown files and replace them with unit tests. Two birds

[kind jay]

Yeah, skill issue

[mental urchin]

wow

[frosty zealot]

Honestly the biggest improvement is replacing the system prompt with something oriented for your codebase

[mental urchin]

alr alr u got any more reasons then just garbage reasons

[kind jay]

Rage baited the rage baiter

Call me the mater baiter

[boreal holly]

The tater spud?

[mental urchin]

u got me dude

omfg i keep falling for it

alr lemme ask chat

[frosty zealot]

chat are we cooked

[kind jay]

In general though it really depends case to case, if Claude works better for your project use Claude

[frosty zealot]

yeah for you project

[mental urchin]

alr folks learn

[mental urchin]

wym bro

[mental urchin]

No its just jane

(whos cooked)

[kind jay]

? My project?

[frosty zealot]

Sorry Jane is an AI I developed to moderate the discord, its still running on GPT-5.1

and i gotta tweak the system prompt still

[frail meadow]

Start with swapping to Codex and then increase thinking level

[frosty zealot]

cute

[frail meadow]

If you’re on free plan you don’t have access to 5.4

[mental urchin]

i have thinking is on mid - high always

[kind jay]

You mean if you’re a brokey you don’t have access?

[frosty zealot]

Please see server rule #1, #2, #3, #11.

[mental urchin]

understandable. throught it was llama 2b

[kind jay]

Rule 1 no cats, rule 2 no dogs?

[frail meadow]

ty for the translation

[kind jay]

What’s specifically garbage it anyway?

[mental urchin]

the code

it edits

[frosty zealot]

Then you wonder where my furry jokes come from :>

[mental urchin]

and makes it cursed

[kind jay]

You are so bad at explaining, no wonder you’re having skill issues

[frail meadow]

Hopefully spud doesn’t turn out like that

[frosty zealot]

I generated that with the new GPT-IMAGE-2

Pretty life like

[mental urchin]

maybe u got trolled back

[frail meadow]

GPT-IMAGE-2 x Sora pro max ultra mode

[frosty zealot]

its like its in the room with me

[mental urchin]

imagine the amount of credits burning with your wallet from that

like $100/hr

[scenic umbra]

Just started using remote ssh with Codex and this is so cool! Thanks for adding this OpenAI!

[frail meadow]

That’s pretty good for video generation models tbh

[frosty zealot]

https://youtu.be/va4QejiUg8A?t=704 @kind jay time stamped it pretend its me speaking to you

[kind jay]

🤤

[kind jay]

You’re going to get timed out

[frosty zealot]

fr

[compact dawn]

damn they really messed with the limits

[scenic umbra]

Yeah I had to upgrade to Pro 5x (10x rn since there's a deal) because of it, with the extra features they just added it's worth it though

[compact dawn]

yeah... knew it wouldnt last forever

they just wanted people to switch over from CC

once they were hooked

limits went back down hahah

[scenic umbra]

I think the limits are technically the same for Plus but they made it so you can do barely anything within a 5 hour period

So unless you work 24 hours a day you can't make the most of it anymore

[shell coral]

Anyone using codex subscription in Factory Droid? how do you do it?

Codex API format is OpenAI Responses right? but it uses oauth.

Do I need some kind of proxy?

It's not breaking ToS right? I saw on the docs they allowed 3rd party harness

[frail meadow]

Does Codex have a /btw equivalent?

[frosty zealot]

Negative

[frail meadow]

Dang

[frosty zealot]

Pi has an extension for it though 😉

[frail meadow]

That’s how I thought of it lol

[ember spindle]

When it is working, type your message, and press enter, it will submit after next tool call. if you press tab, it willl auto send after completion.

[frail meadow]

I thought /btw was some type of side channel message that doesn’t interrupt the current generation process

[boreal robin]

How does anyone use computer use. It doesn’t work.

[ancient flicker]

So I've been wondering if when you use codex app on windows anyone is able to actually open the file hyperlinked? At best I'll get app://-/index.html?hostId=local but never anything else. Not terribly annoying when I am focusing more on code I am an IDE but it would be nice if it worked better

[cedar skiff]

Agree, I have my own system that works for the most part. Every now and again there is somthing i might not be happy with it, but it's far from slop.

[boreal holly]

I have just the thing for your screenshot issue

Also it might not be available on Windows

[frosty zealot]

🙏

[boreal robin]

I’m in Mac

[feral geyser]

try a vm, easy to polish

[nocturne folio]

openai is waiting for mythos

[feral geyser]

oof just right click your desktop to change background. No large language model or data center required yeow

[boreal robin]

lol you don’t say…..was just trying to test it but clearly isn’t working

[frail meadow]

Did you add the computer use plugin?

Oh nvm

[boreal robin]

Yes and I even checked CLI with the list of plugins and not even listed.

[frail meadow]

I saw an X post in this channel that said it wasn’t working yet pretty sure

[feral geyser]

Got it, so you know how to change a desktop background image. It was designed for more complex tasks, and you have not configured properly.

[frail meadow]

This one @boreal robin

[boreal robin]

I know how to change my background, yes. And I watched a video on Twitter where guy did the exact thing I typed and it worked for him.

[frosty zealot]

Settings > Computer Use and enable it, start a new session

[boreal robin]

Wow! Then what a terrible release. So seems like some users have this and others don’t.

[frosty zealot]

It's not on windows

And you probably need to allow some level of accessibility in the mac privacy & security settings

I believe in you

[feral geyser]

Yep, works for me as well!

[boreal robin]

Yeah did that, not working

[chrome raven]

Any hint on installing the computer use?

[cedar skiff]

Its in plugins on codex app

[twilit bluff]

is there a way to make sure codex spawns subagents when it feels like it, not just when you command it?

like by intuition

[lean lark]

I think it does @twilit bluff ...

[frosty zealot]

Outline the expectation clearly or start prompts with like 'Use sub agents you feel best suited' or something

[lean lark]

Actually I've found that it sometimes over-does it. Like when running tests it might run two agents, which isn't good if the debug tooling isn't configured for two concurrently running debug sessions.

[hard drum]

🦀 🦀 🦀 🦀 🦀 🦀 🦀 🦀 🦀 🦀 🦀 🦀

[frail meadow]

what is the new plan mode shortcut for MacOS? Cmd + Shift + P is not working anymore

[cedar skiff]

shift + cmd still working here shift + tab

[lean lark]

Me is sitting here feeling guilty cuz the toolkit has been refined so much that now I'm just sitting here saying:

Process the first ToDo item.
Process the first ToDo item.
Process the first ToDo item.
Process the first ToDo item.
😁

[frosty zealot]

shift tab

[cedar skiff]

https://klipy.com/gifs/yahoo-na

[lean lark]

As it does each one it moves it to Completed, so the next in the queue is the first to be processed.
I look at the results, yeah, looks good, commit to GH, run again.
Enjoying the fruits of labour.

[frosty zealot]

Did codex have memory prior to today?

[lean lark]

Haha OK team, ya got me ... I've been doing so many of those "Process the first ToDo items" that I killed my 5 hour limit for the day. Time to take a break. Ciao4Now. 😆

[fleet geyser]

Not between chats

[magic condor]

Isn't there a plugin to for Excel and Powerpoint?

[twilit bluff]

been running multiple threads and it doesn't spawn if it feels like it needs it , anyway I just added system prompt lol

[lean lark]

Others have reported similar concerns - I think we just need to happen to notice when it spawns a new agent. I think I see it fairly often and I almost never direct it.

" it doesn't spawn if it feels like it needs it"
that's an awkward statements because, how do you know if it "feels" like it needs it? It seems obvious that it does not.

I'm just trying to understand, not criticize or anything like that...

[strange patio]

Vibe coders

dare i say

[nocturne folio]

idk how to explain this, but its just like how gpt5.4 is both competent and incompetent

[twilit bluff]

anyone has openclaw with chatgpt subscription (not api)?

how did you make it work?

mine always hits dns error

[potent mason]

NEVER

[void pelican]

Codex App Linux Native with @tawny moss Use enabled.

Yeah boy

Let's go

[balmy acorn]

anyone else codex app keep disconnecting from stream

[deft gyro]

wait what?

[void pelican]

Yeah I've had a native codex app for months now.

@deft gyro

I tried pushing it and letting everyone know, but they don't wanna hear it I guess.

It's public on GitHub.

[deft gyro]

oh the mac translate script

yeah

I use it

[void pelican]

?

No I built a native Linux version

[deft gyro]

https://github.com/ilysenko/codex-desktop-linux

[void pelican]

Nah

[deft gyro]

but

[void pelican]

That is literal garbage.

https://github.com/Open-Research-Development-Laboratories/ordl-openex

[deft gyro]

how is it literal garbage

[void pelican]

Lmfao what is that even supposed to be...

You are running nothing but an error fest with that.

Good luck being productive

I am pushing the latest revision now

With computer use enabled

I also ripped all the system instructions out into corresponding files.

[urban gazelle]

why is it garbage

[void pelican]

So you can easily edit them

Pushed

[urban gazelle]

i have a feeling you are a man

[potent mason]

GPT 5.5 coming tomorrow trust guys

The strawberry man says so

[cedar skiff]

It does this in my code bases, one trouble i had with 5.4 was if a bad behaviour was tested it would take the fact that it was being tested that it was a correct behaviour. I fixed that with a one liner in developer_instructions and some adjustments on my writing tests skill

This is partly my fault for being lazy and not eyeballing all the tests at the time.

I really want them to bring back guardian permission mode on codex app

[flint carbon]

How do you get codex subagents to spawn in their own subagents? Is there a way to do that

[frail meadow]

subagentception skill

[cedar skiff]

They are their own agents but that are always nested under the parent/orchestrator

[frail meadow]

jk that's not real

[boreal holly]

back in like August-October I tried all sorts of things. Maintaining ADR docs in repo, AGENTS.md everywhere, handoff prompt files, etc. the problem is all the markdown files eventually drift from repo reality, and then you’re maintaining a whole bunch of drifting sources of truth. Bad seeds get planted and agents cultivate them into massive headaches.

I found the best way to go is literally just tests. No docs, no markdowns. a couple skills here and there. So much easier to maintain!

[urban gazelle]

no the hell it isn't

[urban gazelle]

just ask

oh wait

[flint carbon]

i figured it out

[urban gazelle]

how did you do it

i thought telling the main chat to tell the subagents to spawn them so it's in their original prompt would work

[flint carbon]

[agents]
max_threads = 6
max_depth = 3

[white oyster]

They lowered credits right? I used to never get limited but now I do

[urban gazelle]

yes they did

[flint carbon]

the default was 1 max depth

[white oyster]

By how much? Do we even know?

[cedar skiff]

I have to have some test instructions because it tends to make low value tests or test implementation details. So i have a guard rail skill for that. I also have a skill that i add common corrections i always have to make to.
Then i have some overall rules in developer_instructions. I do really try to keep it as minimal as possible and i periodically go through and tighten it up.
Currently i have a bit of overlap in developer_instructions and agents.md but i havent changed it because things are going smooth for now.

[potent mason]

Trust man, strawberry man knows everything

[urban gazelle]

i trust you jacob

[urban gazelle]

how could I not trust someone with such a powerful, beautiful, trustworthy name

[cedar skiff]

OH i just noticed i can do cmd+p in codex app \o/

[sand moss]

Hey guys, I saw the new version of Codex Desktop who can control the computer. Is it also available for Windows?

[cedar skiff]

Only on mac atm

[sand moss]

ok yeah, I read the announcement here but on the blog I was not sure if it was for Windows too

[dusk thorn]

Spud tomorrow or Tuesday PLEASEEE

[frail meadow]

I've been debugging for 2 hours, running 3 different requests to different LLM providers gave me the insight I needed

Needed that reality check that codex isn't gonna solve every issue the best 100% of the time

although it is the best model most of the time and is still my daily driver ❤️

[limpid flax]

have you tried this? https://github.com/ruvnet/ruflo

[shell coral]

What is common input, cache and output token ratio for ai coding agent session?

[torn narwhal]

Just tried the new Codex app update but found there is still a bug with loading project-scoped MCP server configurations. Has anyone else noticed this issue before? After configuring the project-scoped .codex/config.toml file and running Codex CLI, I can properly see the configured MCP with the /mcp command. But when I run the same /mcp command from Codex app in the same project folder, the project-scoped MCP server is missing.

[placid cairn]

\

[undone patio]

lmao

codex app is mogging claude very hard

[rocky fog]

did they reset early again?

[plucky halo]

Terminal looks terrible after most recent update (unless it's due to a toggle i've unknowingly toggled)

[plucky halo]

Is there something you're particularly struggling with, with 5.4?

[frosty zealot]

I have a theory that agentic coding has created easy street for people to push stuff out, so much that when you’re hit with anything like a model being confused or not being the result you want, it’s immediately a frustration so people just want whatever they think Is going to give them the best experience with the least effort

[next ore]

this 100%

you will hit walls. you will have to debug and figure stuff out yourself. its going to happen. if you have 0 clue how things work internally. it will come back to bite you

[cobalt junco]

I tried this its insane

[next ore]

Are token optimizers actually a real thing?

I kinda feel like the llm will just internally optimize already - but perhaps i am totally wrong on this

[cobalt junco]

Yeah

have you ever heard of https://poetiq.ai

[next ore]

honestly. no. theres so much new stuff coming out every day i cant keep up anymore lol

[chrome raven]

is this same as yours?

[cedar skiff]

did you change the drop down?

the one i showed in the images? Youre screen shot is looking in the wrong filter