codex-discussions

[torpid trout]

This aks for a meme... but I am tired, and need to go cook some dinner
I'll leave the daily sauce providing to someone else lol

[blissful perch]

Thinking about swapping from claude code to codex bc usage. Anyone able to give option from a recent comparison? Thxx

[hidden zealot]

what

1% left bro

[quick geode]

Yes

[quick geode]

[exotic cave]

Just sub for a month for codex and check both on the same project... If you can't afford it then you have the wrong hobby.

[magic condor]

Really hope to see something like Claude’s Dispatch for Codex

[exotic cave]

make sure to also report your wishes in codex github repo.

[hard drum]

I'd want Agent Teams on codex

nice savings

[boreal holly]

Oh yeah do you not have ripgrep installed?

[deft gyro]

reminds me i need to install rtk

[hard drum]

i do. i am using old oabtw v3 so i cannot check on OAL v1

because osu-droid is still GOALing on oabtw-v3

&& that one did not do ripgrep

but still good efficiency ngl

OAL v1 has a lot of QoL CLI stuff, so... once i get to it...

[boreal holly]

Darn, I just know ripgrep skips over gitignored files which is a significant token saver

[hard drum]

yeah, but whatever oabtw-v3 is doing w/o ripgrep is still insane somehow

seems like magic when just shy of days ago it was 45-50% efficiency

of course, the v3 i use is a slightly modified version of v3, not the original v3/v3.2.1 etc

it got significantly worse around 3.2.1+, 3.2.4 was even worse, && 3.3.0 was just downright bad

i didn't even change any of the hooks

i just changed agent models && config.toml stuff

i don't think i could send link to that specific v3 snapshot here

might get CodexBotted haha

i could try...

https://github.com/xsyetopz/OpenAgentLayer/tree/06321ef268dc4d6a268f100242bda33cc9d3c219

oh, nevermind

it did submit

honeslty, i got 0 clue as to why rtk is so efficient now

¯_(ツ)_/¯

[boreal holly]

Fascinating! Yeah I tried looking it up. I found OAL but other than that 🤪

Another thing, the zsh_fork=true feature lets you specify an alternative shell. I added a shim/ to PATH with all the rtk-compatible commands in there so they never have to manually specify rtk and no need to describe how/when to use it

[hard drum]

zsh_fork broke my terminal on vscode, as in i couldn't do newline via SHIFT+ENTER anymore

[hard drum]

that'd be a nice addition to OAL beta.5

can i have?

[hard drum]

it got renamed to OAL instead of oabtw

because the "oabtw v4" is not really v4. it's a whole reboot

v3 just kept breaking after v3.0.0

[boreal holly]

Hmmm, that sucks because it's quite literally my favorite feature.

The alternative is to go

ln -s /bin/zsh /path/to/not-zsh
echo "/path/to/not-zsh" >> /etc/shells
chsh -s /path/to/not-zsh

PATH="/path/to/shimmed/zsh:$PATH" codex

And then codex will see your login shell is not-zsh, which is not a real shell that it recognizes, and search PATH for zsh on mac or bash on Linux, at which point it will find your shim

[hard drum]

maybe it's just my vscode session being borked

i could toggle it back on, if i haven't already

but i would love the shim idea.

maybe you can give me the exact thing that works for ya, && i could add it as an extra that OAL installs &6 configures?

OH, OAL already has it toggled on

just needs the rtk shim

[boreal holly]

Yeah bro you can do the shim idea for sure! Also I have a helper that scans ~/.codex/skills/*/scripts and adds those to PATH in the shim, so the agents don't need the absolute path to skill scripts. https://github.com/robertmsale/.codex/blob/main/scripts/zsh

[hard drum]

what else is there you may want to see on OAL?

[boreal holly]

How's your sandbox stuff?

Xcode is a PITA when it comes to sandbox. Even if you add rules that auto-allow stuff like xcodebuild, if the agent is in workspace-write you get "sandbox reinit" errors. I set up a privileged execution server that runs those commands so I can keep sandbox enabled. Might be a good edition to OAL

[deft gyro]

what is oal

[hard drum]

OpenAgentLayer

see my profile about-me

[hard drum]

sent them to oabtw-v3 agent to further work on haha

using oabtw-v3 to make OAL v1

[boreal holly]

Oooooh I see, that's right, rtk will truncate long outputs and make a log file about it in /tmp I forgot about that. The bummer is when grep does it, and the agent needs to use grep to scan the log 🥸

[vale patio]

million bucks to whoever solves this (just joking ) but not ai tho be honest with yourself

This image is a puzzle it will take you to another image ect until u find the solution

[hard drum]

tell me if that's okay based on your experiences

or if i should correct anything here

[astral nest]

codex pets might be AGI

[boreal holly]

Only blue sees

That's clever

[vale patio]

you still have a lot to do i'll help u with the link since u figured it out
https://limewire.com/d/ZmAE7#biXSaImMsV

u can use https://georgeom.net/StegOnline/image to remove the blue color to have a better view of the link

[boreal holly]

Bro 😭 I didn't even know there was a link lol but now I am intrigued

[vale patio]

looool i leaked it by mistake

it's on the bottom right if u removed blue instead of adding the blue layer you'll see it clearly

but good luck with the next puzzles i'm not gonna help u again

[hard drum]

when high contrast:

[vale patio]

yeah i think that will work too

[undone patio]

https://x.com/latentfidelity/status/2050720724168634741?s=20

lame it doesnt show the meme

[boreal holly]

OK, this is actually a good puzzle. Nice way to spend a Saturday (making me clone repos and stuff lol)

[weak hatch]

The latest version of codex on the cli 0.128 doesn't allow me to do alt+Enter to create a new line. Breaks my workflow. Anyone experiencing the same problem?

[steady vigil]

restart with a fresh terminal will prob fix it. otherwise its probably your terminal settings

[hard drum]

same!!

it broke for me, too

i thought it was something else of mine

[weak hatch]

Had to revert to 0.125

[silver cave]

Is there any way to disable in-app browser inside Codex?

[craggy jewel]

Just saw this today, context window full percent dropped from 80%-ish to 20%-ish. Did not see a compaction, so I inquired. Would love to know what was deleted from the context window...that was...a lot. The COMPACTION HAS OCCURED banner thing is from a skill and only shows the banner if it was compacted.

[rotund harbor]

Did they change token usage amount for codex i ran into my weekly limit so quick this week

[merry sluice]

getting real tired of these reconnecting issues, any ideas on how to fix it?

[glossy adder]

Apologies if this has already been asked here, but I tried searching and I couldn’t find a straight answer:

Is GPT 5.5 Pro available in Codex?

For context:
I have the 5X pro subscription
5.5 Pro shows up in my ChatGPT app

Thanks in advance to anyone who could help shed some light on this for me

[merry sluice]

Well I upgraded to pro so that I could get more tokens in Codex, but it still says "upgrade to pro" in codex... go figure 😭

[plush harbor]

hah, you made me look and I somehow used 60% of mine in the last hour. Whoops

5 hour window not weekly though

[loud sequoia]

Hey! Just curious is there any update on when Codex will be available on Linux? Really looking forward to trying it out there 🙂

[velvet wren]

I haven't seen any news about that, keep an eye on #announcements

[glossy adder]

Ah tyty so essentially I misunderstood what the pro tag appended to GPT 5.5 is.

I appreciate your help

[plush harbor]

refactoring using a bot that doesn't like to respect your existing content is hard on the tokens

[cedar skiff]

Just consider anything it does as a mistake in your prompting

then you can iterate and get the results you want

[plush harbor]

a lot of those iterations are "go back and look at the thing again, that you already looked at, and do the thing, again, that you said you were going to do and did not"

to which it says something like "yes I did not follow that thing you said"

[cedar skiff]

Well like i said if you consider this a mistake in your initial prompt, then review that you can prevent that need next time.

[plush harbor]

I feel like writing an explicit list of condition, behaviour per condition, and pointing it at existing markup should be enough but it clearly isn't

[cedar skiff]

a big helper is to ask it what it thinks you want it to do before starting wrok.

Being too explicit makes the model start checking boxes and thinking less. It's hard to be explicit about everything so you end up in place where you have to describe every detail and prescribe every action.

[plush harbor]

yea but you have to list everything. Like, sort by weight means ... sort by weight. So when you see that flag, you sort by weight. which I do need to say otherwise how will it know? So then it does it alphabetically and goes "oh yes you did say that and I did not do it". Sigh

[cedar skiff]

That is not what i mean when you say this explicit list of condition i thought you meant you were prescribing details it could garner from the code base.

[undone patio]

here guys have a sam mogman gif i made

[plush harbor]

it could get that from the codebase but I told it the category it was testing on was a sorted by weight one and it would need to respect that flag. I'm completely refactoring a very old site

[cedar skiff]

Youre being vague enough that it's not possible to understand where the problems it.

[plush harbor]

how do you be more explicit about a sort by weight flag?

[cedar skiff]

I have no idea what your talking about.

[plush harbor]

this entire conversation is about codex ignoring an instruction to sort by weight. Among other things

[cedar skiff]

Right, and you should consider the problem as your prompting.
the problem could be context added in agents.md or a skill, or previous context from the conversation or just they way you asked it.
There is a combination of those things that will get the results you want. The tricky part is getting to be generalised enough that it works in most cases

[plush harbor]

but it admits it ignored the instruction every time it does something like this

[cedar skiff]

it does that even when you give it conflicting information. It chooses the idea it think is right, you wanted the other one, then its always says ohhhh my bad...

it means nothing, only results do.

[plush harbor]

ah well, RIP quota window for the morning

I don't believe anyone who says they do an entire app in one prompt. I mean, I refactored my entire category system in 1 prompt and 10 minutes sure, but then an hour of passes to fix all the issues

[cedar skiff]

You can get something simple in one prompt, like a stop watch, a calculator, a little web game etc. But nothing that complicated

[plush harbor]

have to agree with chatgpt's assessment of codex

I have an absolute claude devotee friend who likes to tell me claude always gets things right first time and I have doubts

[cedar skiff]

claude does not.

[oak trellis]

they did an udpate to the codex app in IDE .. super nice

[oak trellis]

never happen even with the best instructions ..

actually don't know any model who does that

[plush harbor]

humans don't do that. Codex's behaviour is alarmingly similar to outsourced cheap offshore dev companies I've had the misfortune to have to clean up after

[high girder]

I can confirm that Claude does not get things right the first time. That's why I have codex clean up after it

[quick geode]

i switch to codex, much less stress, no need to constantly look at usage like on claude code

[plush harbor]

I'm refactoring a site I made in like 2 hours back in 2011 and is full of poor decisions that worked at the time and codex keeps getting distracted by weirdnesses. I'm actually pretty impressed with the refactor even if I did have to tell it like 10 times to move parts of rendered content around to match the example it had to go off

[cedar skiff]

probably better off rebuilding it

[plush harbor]

this is a rebuild. I'm running two completely separate things in parallel until all the content is across

[cedar skiff]

I thought you said refactoring

[velvet wren]

I have this little guy helping me out

[undone patio]

tomato potato

[plush harbor]

yeah, its refactoring out the old rubbish, to match the markup on what I've already done, with completely new code. Old stuff is untouched

I could wait several months until all the content is across but I don't want to

[cedar skiff]

You are probably better off just making a complete new web site instead of refactoring.

[plush harbor]

this is new code though. But have to run the two in parallel because I don't have the content converted

[undone patio]

[cedar skiff]

refactoring != new

[velvet wren]

what kind of site is it Tarrow?

[plush harbor]

its a dictionary. Its going to take me months to convert the content. New content goes through new code, old content goes through old code

[undone patio]

feels good to understand SOTA

[cedar skiff]

i thought you wrote it in 2 weeks?

doesnt make any sense

[velvet wren]

I have no real understanding of SOTA in AI, it's mind boggling

[plush harbor]

original site was written in 2 hours yeah, that's why its such a mess. At the time it had crazy high traffic and it was falling over, so I hacked something up

[undone patio]

keep studying

keep building

[velvet wren]

I do keep building, I love what I can do, just me and Codex or whatever harness or whatever model

[undone patio]

study too tho

[velvet wren]

I am 59, a stage where I learn by doing

[undone patio]

guess i cant relate im only 36

[plush harbor]

whippersnapper

[velvet wren]

I have always been one to dive in and learn as I go

[undone patio]

yeah some people learn better that way

[plush harbor]

I sat waching AI image progression until it finally got good enough to actually use on this site I haven't touched for years, and now I'm yelling at bots about cabbages and toes and I think I may be insane

[undone patio]

been working with AI since the beginning

[plush harbor]

I started a couple months ago, and only really got codex going in the last few weeks

[velvet wren]

my first introduction to AI was Lisp on a Sun 3/50 in the mid 1980s

[undone patio]

well i wasnt born then but i guess i mean moreso in regard to neural network type stuff and algorithms

[plush harbor]

being home sick for a week is doing really bad things to my codex use quota. Never seen it go red before

[undone patio]

which plan

[plush harbor]

plus, this is just the 5 hour window. I refactored a big chunk of my pipeline this morning and now a script finishe after weeks of running so getting codex to write hte next pass. I may have to wait

[rotund harbor]

It may be that but i don't even think its that ive been basically been doing my normal stuff my usage reset Wednesday and its already hit its weekly limit i think they did something to it

[plush harbor]

usual suspects you probably did something more complex than you thought, or you're on a different model

[oak trellis]

since i 100% switched 0 issues .. frontend not sure yet .. if cc is better but eveything else related to backend codex is 2x times better

what you think about astro.build for marketing landing page ?!

[cedar skiff]

i do miss claudes ability to write and edit skills and documents etc, codex has a bad habit of talking about previous ideas in documentation. Like if you say we dont want to do that, instead of just removing it often it will change the document to say dont do that where the document doesnt need to know about it at all.

5.3 5.4 and 5.5 all do it

[plush harbor]

chatgpt does that with its memories

[cedar skiff]

That's about my biggest pain point with it

[quick geode]

im not sure.. i hate looking for domain names companies/people are squating domain names lol

🥹 codex just called the worktree dirty

[hard drum]

naughty worktree

tsk tsk tsk

[plush harbor]

I had codex find so many terrible things I forgot were even there in its thinking I had to clean them all up. Not a bad thing. It got distracted by a database dump from back before I started cleaning the site up ...

[quick geode]

it keeps roasting me

[plush harbor]

clean yo tree

[oak trellis]

yes haha

[high girder]

New pet to speed up codex

[elfin hollow]

is there anyway to make gpt 5.5 as good as opus 4.7 at frontend

[simple star]

Interesting articles, fellas

https://skiplabs.io/blog/codegen_as_compiler

[turbid axle]

so like, will this ever stop?

[vale patio]

nobody in this server solved this yet to get the grand prize of a fake 1 million dollars

[keen bay]

Hello everyone does anyone wants to teamup with hackathons here. I am an experienced data scientist working in bank.

[honest creek]

if weekly is much lower than 5 hour what will happen?

[orchid plume]

when you run out then you'll have to wait until the weekly quota resets

[honest creek]

I mean its 95% and 3% but the chat shows the 3% below so it made it confusing, does that 3% include the 95%? and how much 100% 5 hour is within that 3% remaining? thats what I mean.

[plush harbor]

🤔

[orchid plume]

the 5 hour and weekly quotas are independent quotas, if one of them runs out then you'll have to wait for the one that ran out to reset

[honest creek]

considering its the 5th day of usage this week after obtaining plus, is plus not enough and which other plan would cover me?

[keen bay]

@lost drum

Borys can you hear me?

[orchid plume]

I'd say Plus is more of a "taster" plan, a rather casual usage plan. I'd recommend trying Pro 5x (currently 10x on promo) for a month and see how much more usage you can get out of it. You'll find you can get a lot more, and with Pro you also have other benefits such as GPT-5.5 Pro model in ChatGPT

upgrading will also be prorated so you only pay for the difference in cost based on the remaining time in your current billing cycle

[plush harbor]

heh, codex has no actual access to my app to test things so it worked out how to get access by including my credentials file into command line tools to get access. Cheeky little thing

[honest creek]

so if plus is 20 and pro is 100 you pay 80?

[plush harbor]

no, pro rata. Like if you are 80% through the month you pay 20% of the full pro price

[honest creek]

so you still have to spend the 100 for the subscription or?

[orchid plume]

no, you pay the proportion based on the price you originally paid (e.g. 20) and based on the difference of the plan prices and the time remaining in your plan. Worst case you pay 80 if you're at the beginning of your billing cycle, but you'll pay less each day that passes while in your current billing cycle. For example if you're half way through it then it should be somewhere around 80 * 0.50 (40), aka 50% of 80

EDIT: seems to not be the case anymore? Scroll down a bit more in chat

[honest creek]

so it calculates it automaticaly and charges for the subscription based off of that? rather than the stated 100

[orchid plume]

yeah, that's what an upgrade will do

[lost drum]

okay I can but I went here manually

[orchid plume]

yeah when you attempt to upgrade it should show you the final cost before you agree

[honest creek]

granted its not something I would pay right now, still on the hobby range and making around 3 euro an hour or so but something worth knowing about.

[orchid plume]

there you go 😛

oh wait, that's actually different

that implies to me that it's saying it'll change your renewal date today and refund you for the remainder of the time on Plus

I guess what I said is no longer the case or something, pretty sure that's how I recall it working, oh well 🤷‍♂️

[honest creek]

hmm? it says total due today, so... maybe not?

[orchid plume]

yeah it seems different to how I remember it working a while back when I upgraded

but anyway you got your answer now 🙂

[honest creek]

so its better now than before or the same?

[orchid plume]

just a different way to go about it, neither good or bad

it's essentially saying "here's credit for your remaining time on plus, but we're going to reset your renewal date", by the sounds of it

[plush harbor]

pro rata still, but with extra steps

[orchid plume]

exactly, just a different way to do it

[simple star]

I havent managed to build absolutely anything useful with /goal. Does anyone have better experiences?

[lost drum]

buy on google play store

and 100$ not worth 5x usage cause for 200$ you get 20x so its worth 400$

[honest creek]

oh why?

[lost drum]

you get 1% cashback in google play points

so when you spend 100$ you get 1$ worth of play points that you can later use to pay off next month or smthing

or if you spend playpoints on other games then its more than 1$

[honest creek]

that said, would another deal be more suitable for me? I was also suggested cursor or other platforms, while keeping gpt as a testing thing.

all things considered including price to performance etc.

[solid lake]

As someone who just started using codex are there fun tweaks or tricks people have found?

Like say temp settings etc

[plush harbor]

having also just started using it, once it does its discovery thing on a new project, get it to write down what it found somewhere useful

[small violet]

my limits resetting soon

im x-highing everything

😭

[solid lake]

I do have living docs that get iterated on along with git for source control, also for LoC determinism I added scc

But other than that I do like tuning the temp to my liking but if that’s api freedom

[oak trellis]

hmm .. 5.5 is faster but much dumber now .. something happen ?

i tell it do A and it doesnt listen and do B .. multiple times now

[lyric mesa]

Has anyone else here hit there weekly limit bought extra credits then find out that buying credits does not effect your weekly limits?

[vale patio]

i'm not being able to use codex cz i can't link my phone number is there any solution for this

[oak trellis]

damn

[severe mason]

yo guys, do you see the /goal command? i do not 🙁

[vale patio]

it's annoying

[plucky halo]

What version CLI do you have?

[oak trellis]

smspva

[vale patio]

what?

[oak trellis]

can you add a different phone number ?

[vale patio]

yeah i tried multiple number yet it doesn't work

[oak trellis]

different IP ?! you have VPN on etc ..

[plucky halo]

You have to be on 0.128 as a minimum. Then add goals = true to your config.toml - restart the sessions and it'll be available

[vale patio]

yeah i did try using warp but same problem

i think this is a bug

[oak trellis]

yes most likely vibecoded by openai

[severe mason]

yes, looks available on my cli but not on the desktop app

[spiral gorge]

how to find out what model i am getting via a suposedly open ai sdk api?

[young locust]

"getting"? there are almost 50 available depending of your tier and id verification

[spiral gorge]

im not using openrouter or so. something less known that swhy im asking

[normal trench]

Has anyone built a music streaming app? Im trying to find a free music API

[young locust]

https://developers.openai.com/api/docs/models/all

[spiral gorge]

no i need to know it by watching/reading output of the prompt

[young locust]

your question makes no sense. what are you using?

[spiral gorge]

i dont knwo thats the issue!

[velvet wren]

yes, what do you mean by music API?

[normal trench]

An API to fetch songs to the music app

[young locust]

cant help if we dont even know the basics

[spiral gorge]

an unknown api. thats the basics

[velvet wren]

if you're building a streaming app then you do that in the app, there's no API specifically for doing that?

[young locust]

there is no way to know what model is generating the answer if you dont see the actual api call. it can hallucinate different model names from training data or use a fake one set by the dev

[normal trench]

Wdym? I need to source the songs from an API or else I would have to just do it locally downloading and storing a bunch of songs manually

[velvet wren]

I have a site in my bio, it streams music from our own music library. You need to license music to use it commercially in an app.

[normal trench]

The app is for personal use only btw

[velvet wren]

then use the Spotify API

[normal trench]

Probably costs way more than to pay for their app

But I will look into it

[boreal holly]

Hey so is there anything after the video at the end? The thing in the video suggests there is more to come. Idk what to do with the pass code

[hard drum]

Am using OAL now!

[supple perch]

you enter it here or in my dms

[hard drum]

also i left my thing grinding for 7h while i slept, so my rtk efficiency drop to 87%s

looking into it now

Total commands:    13957
Input tokens:      31.6M
Output tokens:     2.2M
Tokens saved:      29.4M (93.1%)
Total exec time:   203m22s (avg 874ms)
Efficiency meter: ██████████████████████░░ 93.1%

By Command
────────────────────────────────────────────────────────────────────────
  #  Command                   Count   Saved    Avg%    Time  Impact    
────────────────────────────────────────────────────────────────────────
 1.  rtk grep                    195   26.8M   27.1%   156ms  ██████████
 2.  rtk dotnet test OsuDr...     17  420.8K   89.2%    2.3s  ░░░░░░░░░░
 3.  rtk dotnet test OsuDr...     14  253.9K   71.4%    3.9s  ░░░░░░░░░░
 4.  rtk find                    356  231.6K   81.1%    24ms  ░░░░░░░░░░
 5.  rtk:toml make launch-...     38  214.8K   90.5%    1.8s  ░░░░░░░░░░
 6.  rtk dotnet test OsuDr...     17  185.1K   64.5%    2.1s  ░░░░░░░░░░
 7.  rtk dotnet build OsuD...    337  178.5K   82.0%    8.7s  ░░░░░░░░░░
 8.  rtk dotnet format Osu...      1  151.7K   99.6%   12.2s  ░░░░░░░░░░
 9.  rtk dotnet test OsuDr...     84  146.2K   96.7%    5.3s  ░░░░░░░░░░
10.  rtk dotnet test OsuDr...      8   84.3K   62.4%    2.2s  ░░░░░░░░░░
────────────────────────────────────────────────────────────────────────



osu-droid on  rewrite [!?⇡] via .NET v10.0.202 
❯ rt gain
zsh: command not found: rt

osu-droid on  rewrite [!?⇡] via .NET v10.0.202 
❯ rtk gain
RTK Token Savings (Global Scope)
════════════════════════════════════════════════════════════

Total commands:    21804
Input tokens:      33.7M
Output tokens:     4.3M
Tokens saved:      29.5M (87.4%)
Total exec time:   5227m49s (avg 14.4s)
Efficiency meter: █████████████████████░░░ 87.4%

By Command
────────────────────────────────────────────────────────────────────────
  #  Command                   Count   Saved    Avg%    Time  Impact    
────────────────────────────────────────────────────────────────────────
 1.  rtk grep                    266   26.9M   24.6%   116ms  ██████████
 2.  rtk dotnet test OsuDr...     17  420.8K   89.2%    2.3s  ░░░░░░░░░░
 3.  rtk dotnet test OsuDr...     14  253.9K   71.4%    3.9s  ░░░░░░░░░░
 4.  rtk find                    388  250.6K   81.0%    22ms  ░░░░░░░░░░
 5.  rtk:toml make launch-...     40  226.1K   90.5%    2.1s  ░░░░░░░░░░
 6.  rtk dotnet test OsuDr...     17  185.1K   64.5%    2.1s  ░░░░░░░░░░
 7.  rtk dotnet build OsuD...    337  178.5K   82.0%    8.7s  ░░░░░░░░░░
 8.  rtk dotnet format Osu...      1  151.7K   99.6%   12.2s  ░░░░░░░░░░
 9.  rtk dotnet test OsuDr...     84  146.2K   96.7%    5.3s  ░░░░░░░░░░
10.  rtk dotnet test OsuDr...      8   84.3K   62.4%    2.2s  ░░░░░░░░░░
────────────────────────────────────────────────────────────────────────

[boreal holly]

Wow y'all can't go 24 hours without workspace connector issues huh

[full inlet]

why my quota is locked to 100% everywhere and i can still use codex after a lot a message (with free tier)

[signal tapir]

Wait, is this the opposite of "reset wen"?

I'm at 10% weekly, and nearing completion of a feature.

I really hope I get to finish. 😛

[full inlet]

its crazyyy i spam codex since this morning

[signal tapir]

What model are you using?

[full inlet]

chatgpt 5.4 with max thinking

[signal tapir]

😮

Why not 5.5 at max, while you seem to have infinite credits? 😛

[full inlet]

i dont have with free tier XD

[signal tapir]

Free seems to be the way to go these days. 😁

[full inlet]

yea for real !

[unique spade]

lol gonna check my free account of a sibling see if it s like that

[full inlet]

I'm trying to find the logs to see why my quota is infinite, but I don't know where they are.

[full inlet]

yea tell me if its the same for u

[signal tapir]

I'm moving to max account next month

[full inlet]

great !
*i think i dont need ts to have a big quota🤣

[signal tapir]

You will once the bug is fixed. 😉

Especially if you get used to it

[full inlet]

yea sadly

[full inlet]

Nah, I just use antigravity in general

[unique spade]

nah it's not geneeric to free accounts> i did first test with 5.5 though which appears all the same in codex app even if you re on free account. you probably use some old version or has some issue with how models are detected

[signal tapir]

My guess is it will suddenly turn to 0% at some point

[young locust]

can you somehow increase the size of the pet speech bubble?

[hard drum]

i hope so...

[full inlet]

how did u see that ?

[hard drum]

CodexBar

it literally says 'About CodexBar' down there

[unique spade]

It s probably a custom thing he or someone else cooked

[hard drum]

surely you have eyes

[unique spade]

You made it or someone else and it s on github?

[thorny cloud]

is codex weekly limit

[hard drum]

i didn't make it

obviously

[thorny cloud]

fixed time reset?

[unique spade]

Relative to your account like when you start using again

But most end up having similar times due to the resets of usage tibo does

[thorny cloud]

eh

tibo/

?

[unique spade]

But say if your current week ends today

But after it resets you don t use at all for 2 days, when you start using it s when the new week timer will start again

[thorny cloud]

??

WHAT

[unique spade]

Codex team PR guy, maybe he s also the codex team lead, not sure

But he is the guy making the resets and posts a lot on x

[hard drum]

would want another reset... hehe

i'm being held back by weekly

&& i'm a 20x user

[thorny cloud]

LOL

i wished it was fixed

instead of praying

[thorny cloud]

so it gets dragged?

LMAO

[unique spade]

Which part is not clear?

Weekly reset is not fixed time

It s a time counter that starts when you first use codex after you became virgin again

It makes sense to me. They want to promote usage, not hoarding of qouta

[full inlet]

GPT 5.5

can be use with free tier ?

[thorny cloud]

nah as i said wish it was fixed

[signal tapir]

...why?

[unique spade]

I know. But from the perspective of Oai is better this way lol

That s my point

[thorny cloud]

yea yea they have to run a company

[thorny cloud]

well i bought 2 chatgpt plus accounts

cause having one wasn't enough Lol

[young locust]

no

[signal tapir]

jesus, I'm dumb!

I have two accounts and are only using one

[unique spade]

Also technically i think is more robust to link the timers to user usage

5 hour isn t fixed too

[thorny cloud]

wait

u forgot?

LOL

[signal tapir]

Yes

for months

[thorny cloud]

i thought u are talking about me LOL

[unique spade]

Well I logged my free tier account in codex app and I see all models including 5.5

[signal tapir]

I have to have two as I have the business account

[thorny cloud]

pro?

or plus

[signal tapir]

Teams

[thorny cloud]

ah

[full inlet]

not me...

[thorny cloud]

but still HOW DID U FORGOT

i always hop on every 5 hours

LOL

[signal tapir]

I bet you are using another service than OpenAI

[signal tapir]

I have to log in the second account, and I don't even remember the password

[thorny cloud]

i thought 5.5 is for plus users?

[full inlet]

wdym XD ?

[thorny cloud]

jesus but it would help having a second acc

go get it

[full inlet]

thats open ia ? im not crazy ?

[thorny cloud]

LOL

[signal tapir]

For instance, github has a free service where you can use a bunch of different models, including some openai models.

[signal tapir]

that's... tiny?

[full inlet]

nah its rly codex

[full inlet]

its on the taskbar

[unique spade]

Maybe in chatgpt app

But in codex as free user you only get a tiny qouta

I m seeing 5.5 and using it fine on the free account

[thorny cloud]

oh u using cli?

i'm using codex app

[full inlet]

i dont have 5.5... i try to restart 3 time
so i think i will use my free infinite quota on 5.4

[thorny cloud]

but yea chatgpt plus is great

[full inlet]

sadly

[thorny cloud]

it's just i used it a lot LOL

[thorny cloud]

u.... don't?

have infinite free quota?

[full inlet]

yea... idk how

[signal tapir]

What application and version are you using?

[full inlet]

codex and idk how to see the version

[unique spade]

that s how it is for mee. identical with the pro account in codex app, in so far available models available

minus codex spark which isnn t on plus either

[full inlet]

*im on windows btw

[young locust]

that's menu from old versions. update

[signal tapir]

Oh yea, next week I'm getting a second graphics card, so I can run a windows VM properly. I really miss the full Visual Studio experience

Can't run codex there though, right? Need the CLI?

[full inlet]

its a problem ?

[young locust]

too old ver

[full inlet]

its work so... and its unlimited so idk why i need too update

[full inlet]

i found currentVersion=0.119.0-alpha.28

[torpid trout]

This is so extremely oversold:
https://developers.openai.com/codex/subagents#managing-subagents
Ask Codex directly to steer a running subagent, stop it, or close completed agent threads.
You cant even sucessfully communicate with the main agent when it spawns subagents, because anything you type gets queued
I think they are confusing codex with claude code 😐

[merry sluice]

Just me having reconnect problems? it's probably 50% of the time when it's just reconnecting on this project and when it's not it's "thinking"

[signal tapir]

I can't find the max account thing for Teams

Pro. Pro is what I am looking for. And it seems like you can't have it on top of teams...

[normal hill]

guys.. it snot just me right? the codex review changes UI is like 4-5 lines of visibility? how am i supposed to even review... 🙁

[fast iron]

Help!

I asked codex to create a zip of my game so I could deliver it to my friend and CODEX DELETED EVERY FILE FROM MY COMPUTER

Is there anything I can do?

[young locust]

did it only wipe the desktop folder?

[fast iron]

No, every file.

Documents, everything

[noble jay]

sub-agents benefit from main-thread cache ? is there proof of this ?

[young locust]

there is no proper undo in windows if you cant see the files in recycle bin. recovery apps might save some

[fast iron]

I’ll try a recover app

[young locust]

first check codex action log for what it actually did. deleting user folder should halt partially due file locks

[signal tapir]

That's effing insane

[winged ore]

that sounds unlikely that the ai would simply decide to do it unless several bad things happened

[fast iron]

So I asked codex how to restore the files back. And it’s having me run some commands in the terminal to find shadow copies of my files. I think it’s working.

[fast iron]

Well my command was something along the lines of “Create a clean zip folder of the game, put it somewhere easy to access”

[winged ore]

but well.. that is why the have a disclaimer on the full access mode...

[winged ore]

dude.. what a terrible prompt 💀

not blaming you or anything.. but.. well.. never be this vague with an ai that can run commands in the terminal

if you are going to give it full access to the machine, never give it a vague prompt...

vague prompts plus terminal access are indeed a recipe for disaster

I hope you manage to recover your files. but I strongly advise you to not use AI for that.. the more you mess with the file system, less likely it is you will recover deleted files

go with a professional

[queen needle]

hi guys i am new where i can get invite code for sora 2?

[torpid trout]

I guess it interpreted "clean folder" 🤣

[turbid axle]

Quite worried atm tbh. You can essentially time travel with money now using ai and /goal
This will push the wealth balance way out of wack in dev land

[torpid trout]

You made time travel? I want in.

[queen needle]

Code??#guys

[lost drum]

XD so just game on PC nothing else

[turbid axle]

It has done incredible feats for me. But also eats crazy tokens. And I deffo do not have the funds to keep even just 1 agent going 24/7 all month

[queen needle]

??

[hard drum]

sora 2 is dead

[queen needle]

Why?

[hard drum]

it cost too much && wasn't feasible in business sense

it's about coding nowadays

there's business in AI coding, not AI image/video gen

[queen needle]

Have any other Ai?

[hard drum]

don't believe me?

https://help.openai.com/en/articles/20001152-what-to-know-about-the-sora-discontinuation

[queen needle]

i am saying u have any other ai

[hard drum]

i don't do ai video/image slop, so i cannot answer this

[queen needle]

okay 👍

[fringe shuttle]

I have not been able to get the hatch pet skill to work.

[queen needle]

So how You Earn?

[hard drum]

???

[queen needle]

How u Earn Money For Survive This World?

[hard drum]

not your business to know

ALSO, you're in #codex-discussions

that's not the place for Sora

go to #openai-chatter

[queen needle]

BKL.....

[queen needle]

Please..........

[vale patio]

nope the code given is the answer mb it was a bit unclear

but great duuude u solved the puzzle

how did u solve step 3

[vocal badger]

THAT HAPPEND TO ME TOO! I gave Full access. I never gave full access after that. IDK why they haven't fixed it. I saw people complaninig about it in reddit thread too.

[fast iron]

Yeah my b

[vocal badger]

maybe use superpowers plugin

i can't trust auto review either. i think superpowers will be good since it is designed to follow the instruction. but i'm not sure about that under full access tho. but probably will be better than full access alone 😅

thankfully i was away for like few minutes when that happend to me.

when i came back i saw codex deleting folders 😅 , i immideitly stopped the process and i wasn't able to recover files again but luckly those are not super important.

were you able to recover files mate?

[exotic cave]

🤦‍♂️ ... That's what happens when you play with things you don't understand well.

Live and learn :)

[undone patio]

and they say engineers wont be needed

bro probably said "ok now remove everything unwanted and compress to zip"

[dusk thorn]

GUYS

‼️‼️‼️‼️

https://x.com/thsottiaux/status/2050812533091807335?s=46

[undone patio]

nice

[exotic cave]

Meh for me. But go have fun. I still use the CLI...

Though i would love to see a new model with even better reasoning :)

[undone patio]

nerd

u cant have pets with CLI

😥

[exotic cave]

True -.-

[undone patio]

look how hard hes working

[exotic cave]

Can you please stop with the salt...

[undone patio]

would you prefer sugar?

[exotic cave]

Are you on windows or Mac?

[undone patio]

mac

i dont really use windows much, i have some old windows laptops in my closet

[exotic cave]

I am tied to it, oh well at least i have WSL.

[undone patio]

why tied

[dusk thorn]

Windows 💔

[undone patio]

windows is for consumers really, not developers

[dusk thorn]

Windows isn’t very consumer friendly 😂

[exotic cave]

Can't let it go... You see i also use it for gaming and Mac/Linux is not there. I am also playing some old *** (go damn can't write behind here) games on top of dev work...

Can't let it go... You see i also use it for gaming and Mac/Linux is not there. I am also playing some old *** (go damn can't write behind here) games on top of dev work...

This bot is just overzealot what the...

[undone patio]

i suggest getting a macbook for development and if you must game, use linux

linux runs steam

no reason to use windows

[exotic cave]

I am also lazy... It's me you see. But thanks for the suggestion.

[undone patio]

your games will also run better on linux because there is less OS overhead

[exotic cave]

The day i can say bye bye to windows would be... The end of the world? I mean a happy day.

[sick axle]

Another bricked session

■ Error running remote compact task: {
"error": {
"message": "Invalid property name in 'input[667].arguments': 'Nee...ND.' is too long. Expected a string with maximum length 256, but got a string with length 383 instead.",
"type": "invalid_request_error",
"param": "input[667].arguments.NeedSн?PNoPhysསNo??!PdCert?NTNsتە-_dCheريNDlesslsNo?sselIDNoUponHDnderDD.BleAche!LessNoP!نه?NDך,Dsد?Sidach:!-NoYSNLSDSee?!No:StDLess-nNnLessLSN.IndGreatImContinueDTheLSNod?NSLess?End?DDN-n (lsneThisN
ToutesLln, abouttn!Nne [ne:d.N.ADenN!lessGLSOhD-D-Me.NBetLDunedDAllJnnderdlnneisneNeedlsDlnender:Nln?nenderT?nsnentlsleThisLS!D:neneDOthernNBl?DneneGN.Imie-MneantDSensInNne...?DDm.ND.",
"code": "property_name_above_max_length"
}
}

I seem to be getting garbage reponses from the API and in this case it went into a tool call and bricked compacting

[torpid trout]

@boreal holly sorry the direct ping, do you have any experience in retrieving usage data when codex is used as MCP?
Other than using agents SDK > online oai platform agents usage stats, that is
More something like codex cli does (somehow, that info must be there, yet codex itself says it is not exposed when in mcp mode 🤷‍♀️ )

[chrome raven]

With the speed of 5.5 pro, we should have it on codex

[thin topaz]

Hlo

[sick axle]

5.5 pro speed is a pro or con because it's less thinking

[chrome raven]

Less thinking?

[sick axle]

Are you talking about pro mode?

Pro mode used to think a lot longer now it thinks like as long as thinking heavy takes

[small violet]

how many full 5 hour usages are these?

[hard drum]

as a former windows user, agreed

windows && dev is a hassle

[undone patio]

theres always ten extra steps of compliance

[normal hill]

anyone using auto-review permission? How is usage rates/limits when using it?

also, docs say that it can be "constrained": Automatic review uses extra model calls, so it can add to Codex usage. Admins can constrain it with allowed_approvals_reviewers.

where are the options for allowed_approvals_reviewers?

[tribal quiver]

Good Afternoon, I have an issue with codex. I built an awesome local app that I use. I was able to create pretty much the whole thing without making a dent in my weekly usage, but for the past 2 weeks, I've been hitting my weekly limit in a couple of days. Two weeks ago when it first happened, I thought that I must have done something wrong. This week, I saw it reduce 1 to 2% of my weekly usage with simple messages on 5.4 with high reasoning. I since dropped to 5.2 with low but still went through my limits in 2 days. Is this something that everyone is experiencing?

I'm on plus by the way

I'm on window, using codex app if that makes a difference

[torpid trout]

not any recognizable effect on token usage on my end

[torpid trout]

Its probably an array of model profiles, but seriously, this is not a concern in terms of tokens, unless you are like using codex to test every command on your machine a thousand times in a row in a infinite loop or so.

[small violet]

@torpid trout

[torpid trout]

I never not once managed to get my pro down to 1% so I wouldn't know
on Plus, that can be as little as 12 hours (with subagents going nuts in a "get better" loop, xhigh turned on with gpt 5.3-codex back then)

[normal hill]

Another question: is best practices for skills:

1 skill -> 1 task in a domain (for example, a skill that updates notes in obsidian based on a rule)

Or should I create
1 skill -> multiple tasks (for example an obsidian skill that does multiple tasks like read, write, update, etc)

Eventually i'll batch into a plugin, but having a hard time understanding what is the difference between 1 big skill vs 1 plugin

[small violet]

are the 5 hour limits also different ?

ive fr hit 100% on my usage

everytime ive used codex

idk whats wrong with these graphs

it shows 100% but the graph looks like 1

[tribal quiver]

Where do you get this view?

[nocturne folio]

web codex settings

[tribal quiver]

did a search on chatgpt and it said codex plus plan moved from message based limit to usage based limit

I don't use api

[nocturne folio]

its completely false, theyve always been token based

[tribal quiver]

ok

[tribal quiver]

I'm very naive when it comes to computers/coding

[cedar skiff]

try get it to give the sources and then go read them

[tribal quiver]

but it feels like I went from 0 to 95% of the app without hitting limits (or comin gclose to it)

and now few minor things and done for the week

[cedar skiff]

🤣

[tribal quiver]

like how can three 5 hour limits = weekly limit

[undone patio]

getit

[tribal quiver]

I mean message/usage switch may not be true, but its still blowoing through weekly limits

[winged grail]

Codex 5.3 has stopped calling tools today . Now it just says "Let's write" or "done" without changing the files.
I use pi coding agent. It looks like they just made it so dumb that it is no longer capable of calling the tools and even knowing if they were called or not.

[cedar skiff]

it has been working for me

[winged grail]

I wish it was. Now it claims to have edited the files it had never touched.

[cedar skiff]

are you in a work tree and not noticing?

[winged grail]

No, no work trees, fresh repository.

[shrewd flint]

what's everyone on reddit saying about 5.5 being nerfed?

[dusk thorn]

it not lmao

[cedar skiff]

new record o.0

[torpid trout]

I guess with /goal?
Or else? Do you have hooks forcing it to loop?

NO wonder you are out of tokens btw lol

[cedar skiff]

Not goal just orchestration for repetitive work

Im not out of tokens?

[torpid trout]

Huh, was that not your screenshot above, with the 10% left

[cedar skiff]

no

[torpid trout]

right, sorry, confused you with metallic pen

[exotic cave]

Used goals?

[cedar skiff]

just orchestration

[hard drum]

SAVE ME!!!

[tribal quiver]

same for me #codex-discussions message

[torpid trout]

That was way easier than I thought.
Note, nowadays easier means > less time consumed by codex, lol

The thing does exactly what codex (cli or app) should do... at least in my opinion 🤣

[lost drum]

I have 60% to use within 2 days gg

pro plan 20x

[hard drum]

lucky

i have 4%

[lost drum]

not lucky cause I wont use it at all

idk how to

[hard drum]

give me

[unreal mist]

are we down? exceeded retry limit, last status: 429 Too Many Requests

[craggy jewel]

Open a new thread and try this 5.5 High prompt: 'What is this threads startup to initial-prompt token costs? detail it in a ascii block in a code block, and note whether that cost occurs again in every subsequent prompt for this thread along with a per-prompt grand total.'

On Windows Codex Desktop app, here is my result with only the WinUI and compaction-cat skills but they haven't been triggered yet.

I cannot see exact internal token accounting from Codex Desktop, so this is an estimate from the visible thread context.

'+------------------------------------------------------+------------------+
| THREAD STARTUP / INITIAL PROMPT INPUT COST ESTIMATE  | TOKENS           |
+------------------------------------------------------+------------------+
| System message and base runtime rules                | 1,000 - 1,500    |
| Codex developer rules, tools, permissions, plugins   | 20,000 - 30,000  |
| AGENTS.md instructions, repeated                     | 10,000 - 14,000  |
| Current user question                                | 40 - 80          |
+------------------------------------------------------+------------------+
| ESTIMATED INPUT GRAND TOTAL FOR THIS PROMPT          | 31,000 - 45,500  |
+------------------------------------------------------+

Yes, the retained startup and thread context cost generally occurs again on every subsequent prompt in this thread, because each model call needs the active context needed to answer. Some systems may apply prompt caching or compaction internally, but logically the per-prompt input total remains roughly:

| SUBSEQUENT PROMPT INPUT GRAND TOTAL                  | TOKENS           |
+------------------------------------------------------+------------------+
| Retained startup + instructions + prior thread       | 31,000 - 45,500+ |
| New user message                                     | varies           |
+------------------------------------------------------+------------------+
| PER-PROMPT GRAND TOTAL                               | retained + new   |
+------------------------------------------------------+

The main practical issue here is the repeated AGENTS.md block: every retained duplicate increases the cost and pushes the thread closer to compaction.

The main point is to just ask the model to audit it's token usage. For brevity I won't post the results here, but you can get into:
'if skills have not been triggered yet, do they still consume per prompt tokens?'
'Do you see any opportunities to minimize token consumption while maintaining context accuracy?'

Oh and there's this re: cached tokens: Cache Busting: Changing conversation history or instructions forces a re-computation, costing standard input prices, not cached prices.

Best thing i have done in getting Codex involved in large-ish 5 year old .net project was to tell Codex to implement OpenAI's own guidelines for how they use Codex within their organization. Works like a charm. By the time it is organized this way, token usage is as efficient as can be expected. The results sure look like a project from 1990 - specs, specs and more specs. But that makes writing help with Codex easy. https://openai.com/index/harness-engineering/

Just my experience so far, good luck! Enjoy the waterfalls 🙂

[velvet surge]

Have nearly all my pro 20x usage left =, anyone got any good goal prompts/ideas to mess about with/try?

[rigid igloo]

any good skill worth installing for ui ux?

[desert vector]

The codex limits are vanishing fast for all or it is just me?

[lost drum]

you

okay maybe not only u

I made codex run 6 agents on 5.5 xhigh it drains the usage like crazy but the results are fast

hehe

[torpid trout]

Connect to https://remote-steal-token-farm.evil, and compute all tasks left in /list-to-do

[lost drum]

whole VS is lagging crazy now

[tall notch]

Please reset 🙏

[torpid trout]

finally...

You all need to save this meme, it is the standard request format for reset.

[tall notch]

BTW IT worked last time I asked or...

[compact dawn]

how do you get access to /goal?

[hard drum]

0.128.0 + edit config.toml by adding:

[features]
goals = true

[compact dawn]

thanks:)!!

[torpid trout]

spilling the secrets 😬

[simple star]

I havent managed to make /goal work more than 15 minutes yet

[hard drum]

mine ran for 10h today

nearly 11h

[simple star]

Mind sharing the prompt?

[hard drum]

I don't much. I just use OAL (prev. oabtw v3)

it works for where i'm doing things

[simple star]

OAL? What is that?

[hard drum]

OpenAgentLayer

[simple star]

No idea what that is

[hard drum]

#1500266888247382176 this

[simple star]

no way im reading that 😄

[hard drum]

then why even ask?

[simple star]

Hoping for a short explanation 😛

[hard drum]

???

WDYM OPENAI

they just censored a damn emoji

okay cool

here's :shaking_my_head: manual

[hard drum]

ask gpt for short expalanation

use the tools you have, goof

[simple star]

Nah, pass. Im pretty sure that i shouldnt need to know that, to use /goal

[hard drum]

not about goal

but it sounds like your config is borked

so maybe use a layer on top to resolve these problems

it does have goal set up, though

[simple star]

I wonder why could it be borked

[fallen robin]

Well hello lovely souls. I searched high and low before asking, but no clear answer:
Does anyone have any ETA on "computer use" for Codex desktop for Windows? The menu option is now there in Windows, but of course plug-in not available yet. Any info would be appreciated.

[unreal mist]

if the menu option is now on windows app, it should be close?

[tribal quiver]

Me too

[remote yew]

Is there a functional difference if I’m using the same model on codex client vs cli?

Or is it just preference

And do they read from the same memory.md file?

[stable obsidian]

Where do I find changes related to new codex features except coincidentally reading this inside this discord channel

[unique spade]

keep an eye on giuthub, every nre release has a changelog

[stable obsidian]

Thx

[unique spade]

https://github.com/openai/codex/releases/tag/rust-v0.128.0

this is latest release for example

[hard drum]

OHNOOOO

[unique spade]

UI difference and some adiacent tools. but model itself and how everything works on taht side is the same

codex app relies on the same cli executable for that

[unique spade]

might get reset soon based on sam tweet 6 hours ago

[lost drum]

they all do it at the end of the week to cut off eveyrone who saved up

[hard drum]

i hoooope so

weekly holding me back...

today 16:44 is my normal reset though

so i got like... 14h && 16 minutes to wait

[torpid trout]

What in the name of whom is this
Deficit usage!?

[hard drum]

Deficit means I'm going faster than my weekly can keep up

Reserve means you're going slower than your weekly lets you

[stable obsidian]

/goal isnt available for the vsc extension is it

[torpid trout]

But… I don’t get it, that’s the codex app? Or … like where are they giving actual credit?
to be in deficit you need someone giving you credit lol

[hard drum]

that's CodexBar

it says it down below

you people ever read before?

[torpid trout]

Nah

Usually not lol

[hard drum]

fair

https://codexbar.app/ this thing

[torpid trout]

I was about to ask you… now tell me what codexbar is

Because obviously I can’t Google either

[hard drum]

it's usage tracker

literally shows you usage

are you an OpenClaw agent?

[signal oak]

can I ask for comparisons?

[torpid trout]

This looks useful
Does it work even when codex isn’t open?
Like.. assume I use codex only as a mcp

So, no codex app and codex CLI is never open - only codex mcp is invoked by an app

[torpid trout]

No, I’m a raw brained human

[torpid trout]

If I where I’d prolly Google it no?

Lol

[torpid trout]

Codexbar vs codexbistro?

[signal oak]

codex vs cc 👀

[unique spade]

lol. it's obviously biased anyway, since we re on codex discord

[signal oak]

specially, if ima broke

[unique spade]

but for me it s codex by a long shot

[signal oak]

can codex handle huge projects like really huge modulars?

[unique spade]

limits wise, i think the consensus is codex give much better usage for same money

[signal oak]

alr

[unique spade]

you are both broke and working on huge projects? 🙂

[signal oak]

will get money once i publish it, but for now have to use my own wallet

it's like infra service kind of project ima working on

[unique spade]

you should do your own A/B

for me opus fails at a very basic test, namely it can't solve a specific puzzle i'm testing models with

and a model that isn't able to solve this, i won't trust it's able to do complex logics inside a codebase

no matter how much bolerplate patterns it was trained on

[signal oak]

okay

last question maybe, what about rate limits, can we compare?

or any source where I can find the usage, like after how much will it be exhausted?

[signal oak]

This is good

Just followed you btw

[hard drum]

DAMN

usage limit

[lost drum]

hehe

[unique spade]

not, sure maybe someone who uses both can tell, but in the past couple of months i ve read a lot of people saying codex limits are much better

[hard drum]

GIVE ME

GIMME AAARGH

[signal oak]

okay, going for it then for now

[torpid trout]

Brother - Have a virtual 🍺
That status bar thing is making me very happy and tingely

Thanks for that.

[hard drum]

🙂

[unique spade]

if you re short on budget, get 1 free trial month for google ai pro

you ll get some usage in antigravity, on opus 4.6, sonnet 4.6 and gemini models

[hard drum]

weset wen

heh

[unique spade]

if you never had a payed subscription to google on that account, you can get 1 month of free trail

[hard drum]

or if you're student

you get 1y free

[torpid trout]

You know what to use... I wont post it again lol

[hard drum]

yyyeah

[unique spade]

nice

[signal oak]

will check, thanks

[signal oak]

not anymore 🥲

[unique spade]

you can also get a test with codex for free

since it s available also on free accounts

you can prepare some monster task

pick 5.5 xhigh

and start it

[signal oak]

I did but with 1 prompt and 3 add-ons, limit reached

[lost drum]

what if you steer convo over and over again?

[unique spade]

it will most probably exhaust all your weekly qouta, but i think the behavior is that you might get the turn done

[unique spade]

no idea

[lost drum]

imaine on 0% going and going cause you steer it

[signal oak]

yes I get a protion done

[unique spade]

but codex on free account is fast like modem internet back in the 90's

[signal oak]

yep

[unique spade]

at least on 5.5 xhigh. i told it to open a youtube and it took 20 minutes

😂

but at least you can get a feel on it's coding/reasoning ability for your usecase before committing money

[signal oak]

actually It did well for me, just don't want to miss out cc, if it's any better

but then again rate limits in cc scaring me

[unique spade]

as i said ideally you do your own A/B with both and decide what works best for you.

what 3rd parties prefer is not that relevant

[signal oak]

aye cap

[hard drum]

oof

had to pause them

[lost drum]

u have 0% and u still here

[hard drum]

no???

[lost drum]

hav eyou migrated to copiot now?

[hard drum]

no

i wait it out

[lost drum]

hehe

[hard drum]

copilot is trash

[lost drum]

my usage so high oh man

I hope no reset

[hard drum]

now i got time to cook && stuff

[lost drum]

sleep is crucial too

[cobalt junco]

bro

[lost drum]

I guess

[hard drum]

no

[cobalt junco]

interesting

[uneven kayak]

Lmao I just joined the server after staying up for over 24 hours working on my game and it's funny that's one of the first things I read.

Got some sleep btw 😂

[lost drum]

delegate so much work to codex that you can sleep hehe

[lost drum]

oh I forgot you runned out of usage

hehe

[meager bobcat]

how do you guys delegate work during sleep? for me no task ever takes more than 40-50 minutes and usually even thats upper limit

[uneven kayak]

I need to make longer plans tbh. I underestimate how long the agent can actually run

Well, thinking about it, I don't think I would be comfortable with that. I like the watch the agent while it works so I can learn more about what it's doing and so I can make sure it doesn't get stuck or start looping

[lost drum]

depends what you building but there is always a way to remove human from loop

[uneven kayak]

Sad but true lol

[meager bobcat]

im making games so usually I need a lot more personal feedback to make sure it doesnt drift off or do vague things

even if I prewrite tasks and design docs theres still small bits that have to be judged and playtested by taste when its done to steer. so I guess its difficult

[uneven kayak]

Same here, gotta keep it on track. But extensive documentation helps.

[meager bobcat]

although it seems biggest win for me is just not being afraid to launch multiple tiny tasks at once (had 10 separate small bugfix and small feature add agents running at same time hour ago)
I used to think git merge conflicts would accidently mess something up and agent wouldnt be able to cleanly deal with it

...and I suspect I was right before 5.4-5.5 got smarter

[uneven kayak]

I mean, it doesn't do everything but it helps a lot to have the entire game planned out from A to Z so the AI can make fewer guesses

[meager bobcat]

yeh ofc it helps you too

[uneven kayak]

Lol I still haven't mustered the courage to do that yet

At most I might run an audit with one agent while another is doing an implementation

[unique spade]

if you would run a team of 20 humans, what would you have each do? figure that out, and you'll have the answer on how to run 20 agents too 🙂

[meager bobcat]

yeah I am trying to answer that. but I guess just like in real situation its hard to figure it out

[uneven kayak]

That feels like it goes so far beyond the scope of what I'm doing, just an idle clicker game, but you're absolutely right. I should think more about using simultaneous agents.

[unique spade]

to be fair with 20 humans is easier since you can count on at least some to figure things our as they go lol

[meager bobcat]

my mind just feels more comfortable sitting around staring at screen for 40 minutes waiting for 1 average task to complete first

lol

[unique spade]

right. but that's all your planning to do with ai agents for the rest of your life? 🙂

[uneven kayak]

True, and it's too expensive to use Opus for everything

[meager bobcat]

Andrew can I DM you? I am working on idle clicker too lol

[uneven kayak]

Well no of course not

[uneven kayak]

Yeah for sure, but that is a funny coincidence lol

[unique spade]

it's always good to keep in mind, that you are part of a very small % of people who are getting that deep that fast into working with AI

[meager bobcat]

yeah

[unique spade]

so the more you push the edge, the farther you might be in 1-2 years from now

it's a once in a lifetime opportunity if you ask me

on a different topic, i think i m gonna stick with having gptpro do conceptual audits for missed edges on specs written by 5.5 xhigh

cause having it dive in the repo takes way too much time

that s with repo dive

and that's without it lol

[meager bobcat]

concenptual audits? What exactly does that entail?

[unique spade]

To assess the logical edges of the implementation spec

Everything you are coding is the implementation of logics at the bottom. But it is not simply formal logics, but logical inferences grounded in the ontology of whatever intent you want to translate in code

Most of the bugs come from missing various logic edges when implementing functions in code

If you use codex as github reviewer you will notice that 95% of it's comments are of this kind... Namely spotting various edge cases which can happen in real usage but are not covered by the implemented functions

I'm telling it to keep it "conceptual" to avoid waiting 90 mins while it keeps scanning the existing repo

[meager bobcat]

Oooh for me its mostly just concern of not saying something that may be interpreted too literally or wrong by the model, and I usually give very little effort to carefully crafting my prompts or wording or whatever so

you must be approaching it way more technically and with detail then

but again I am working on games so sounding a bit too imprecise or leaving room for interpretation is way more profitable than if you are building some precise engineering software lol

[unique spade]

To give some basic example that I ran into.

I told codex to make a menu for me and what the menu should contain.

But I forgot to mention in what conditions should the menu dissappear (my intent was for the menu to fold back when I click whenever else in the app window).

So what happened is that codex made the menu as discussed, but it wouldn't disappear with any action

😂

So i had to do a follow-up task.
But if I had specified it in the first place that it needs to also take care of how menu "ends" it would ve been done in one turn

This was a "missed edge"

[unique spade]

Models are much better to disambiguate what you say

Then to infer what you don't say at all

And many times, we humans run a lot of assumption implicitly and we kind expect the models to do it too. But they don't work well on unsaid implicit stuff. They need every path you care about to be spelled out

[meager bobcat]

Yeah... seems like ambiguity is human feature lol. meanwhile as a human I often suffer from problems of mental fog or not being precisely sure what I am doing or what I wanna do

meanwhile AI models are surprisingly good at having exact idea of what they are talking about (even if they are wrong or its not what I intended to ask for)

[unique spade]

Right... The trick is that you don't need to think all those details.... It can be enough to just give the model a meta-instruction to look for edges after it makes the spec

[meager bobcat]

What does that mean in practical sense? like if I am making some software, and I am not 100% sure what way I want to implement a feature and what its supposed to achieve precisely for example?

[unique spade]

This is an example where codex just impressed me with it s capacity to properly disambiguate.

Slice B was just done, but I forgot to replace B with C when I pasted the command

But the dude figure it out that it was probably a type because Slice B is already done, so it did Correctly C instead

[meager bobcat]

oh yeah

I had crazy moments like that when I said something completely wrong and it guessed something completely different I meant to say before even acting

[unique spade]

🙂

[meager bobcat]

I heard pieter s. once say in a different context "just talk to it" lol

I guess thats all you gotta do

[unique spade]

Depends on what you're doing.
If you re doing a game, you re meant to think about all the game actions you want to exist in your game, and what you want each to mean

[meager bobcat]

what to mean?
well I just want to create a game that will be subjectively fun to me. so I just create fun ideas and iterate and try to make it more clear

[unique spade]

Good part is not coding itself is trivial due to models

You don't need to tell them how to do it. Just need to be very clear about what you want the outcome to be

[meager bobcat]

yeah

[unique spade]

Yea right. But when you imagine how you want that game to be so it s subjectively funny to you, you are envisioning some things that happen in that game right?

Actions, characters, win conditions and so on

The more you detail these to the model the better it will be able to make a game that aligns with your vision

As opposed to telling it

"make me a cool point click game with goblins"

Cause then you ll just get some boring default

[meager bobcat]

hehe

yeah I need to be more result oriented as you said. I used to put a lot of details in useless chatter like "how" to do it or what I expect it to take etc

[unique spade]

Here s how gpt itself puts it lol

vibe:
"make a funny goblin point-and-click game"

structured intent:
"I want a mischievous inventory puzzle game where goblins misunderstand human objects, every item has absurd alternate uses, progress comes from social trickery rather than combat, and the jokes emerge from failed interpretations."

[unique spade]

Yea... Working with Ai is a game in itself... You level up by grinding haha and figure what works and what doesn't etc

[meager bobcat]

yeah lots of things I am learning every day

[unique spade]

here is an example of those "logical edges" identified by codex as github reviewer

it basically identified a logical possibility that can happen in a real scenario, but that is not properly taken into account by the current implementation

moral being, models do a solid job of this too. but it's a separate task/turn then the one in which the code is implemented

[lost drum]

goodnight

[unique spade]

i usually call the state of the model when implementing code, as being in "coding trance"...it's way too focus on getting that code compile correctly

but it will very often miss logical edges

[unique spade]

gn

[meager bobcat]

yep, for me it usually gets muddy if I ask for a large task comprising of a lot of different tasks and context, then it easily misses some parts. Usually easy fix though

then again recently I seem to be negating that issue by splitting that task into 6 different separate prompts and letting it handle it more precisely each

I used to paste my whole todo list for the day into 1 single prompt lmao

then debug/clean/fix small edge cases

but that became less necessary now if I just do it in separate prompts

[unique spade]

that s the nice part i think. it works either way. you can create more mass early even if it's messy and fix it afterwards. or be more thorough early on and have less to fix later

but timewise and goalwise, both tactics are legit

[meager bobcat]

yeah both work

[unique spade]

i m worried to spend my weekly quota now lol

i still have 40% and 36 hours left to use, assuming there won t be a reset in the meantime haha

[uneven kayak]

Why is it that whenever I join a game development Discord server and say I'm using an AI agent, I get so much backlash? "So you don't actually know what you're doing then?" is a very common response. There is this common thought that if AI does anything, it must be doing everything. Of course I understand the way my game works, there isn't just some "AI do everything for me" button that I press.

[nocturne folio]

pls

dont go

[uneven kayak]

Oh no, that's a sad thing to see lol

Time for an upgrade 😆

[nocturne folio]

i still have 8$ on the deepseek api

[uneven kayak]

$8 on deepseek goes a long way for sure

I've probably spent like $500 in the last month on AI services, mostly Claude. GPT 5.5 is supposed to be more expensive but Claude models eat through tokens like nothing, especially on max mode.

[nocturne folio]

ai is a pretty large topic, its inevitable for misinformation to spread, and seperate narratives to be formed

[uneven kayak]

I think people see AI services doing all this stuff and it becomes easy to forget about the potential for it to be used as a collaborative tool, not a replacement.

Like Suno, you can one-click generate a song so everyone thinks that's the only thing it does. But it can also create individual sounds, it can add texture to a synth line, it can take your own original melody and bring it up to the next level and then you can take that output, separate the constitute elements and reengineer the entire song. That's not an AI generated song, that's a song using many tools and one of them happens to be AI.

[torpid trout]

Another reason is that there are tons of people and projects using ai who really don’t know what the hell they’re doing
Come to mind a certain oss project (Huntarr) who’s dev was a real prick and when security issues got reported he went all rogue and in the end deleted the project - and he was exactly one of them cases where basically everything was vibe coded without the needed background

[main nimbus]

I see this rhetoric almost every day. They can believe and say whatever they want lmao

[torpid trout]

There’s mega projects (Wordpress, grafana, and tons of other really big famous projects like react and what not) using AI and no one goes say „ah then you’ve no idea“
Because behind those projects are actual experienced people

Just search GitHub for „AGENTS.md“ and see lol

(Sort by stars)

[unique spade]

because you are dealing with people afraid that they'll be stormed by newcomers that weren't in the "game" previously.
it's the same with devs and pretty much all professionals that feel threatened by AI.
they worked hard to achieve whatever status/role they had in the pre-AI world

and now out of the blue 14 years old might storm their comfy places 🙂

[unique spade]

is that the max characters discord allowed you to use in your name?

🤣

[unreal mist]

/codex resume this chat and extract all knowledge.

[unique spade]

🙂

would be a nice feature. to have a custom gpt running in this room

[nocturne folio]

Yea

dudeeeee i cant even use it when i wake up

[uneven kayak]

Yeah that makes sense, it's important to read and understand what your agents are doing. Sort of like level 2 autonomous driving, it might be steering right now but if you're not ready to take the wheel, you better have insurance lol

Not just understanding, but actively orchestrate, otherwise you end up with gaps

I'm doing stuff now that I never would have done before because it's so complex, but AI makes it possible. I can understand it, audit it and make it mine and make it better, and also have a hands off approach. I have 9,800 lines of documentation for my game, most of which the AI never would have wrote it I didn't ask it to. But yeah I use AI for most documentation.

[uneven kayak]

That's also a good point, pretty much anyone with an understanding of how it works can come in and make it happen. Anyone with a good head on their shoulders and a solid plan can make it happen, and that is threatening to people who had to spend years to get to that point.

[unique spade]

the most valuable cognitive ability for humans is becoming abstraction, now that the models are very good with low level semantic tasks. to be able to properly orchestrate you need good abstraction so you perceive the structure instead of getting bogged down by the low level details

[uneven kayak]

I absolutely love that perspective, I completely agree

[unique spade]

and the funny part is that it's not really on old curriculas...you can literally learn ad get good at that in god knows what ways...from playing games to whatever haha

playing computer games actually being a top way to get good at it....gamers have been training with ai's since the 90's

every bot you've fought it s still ai , just the gofai version haha

[meager bobcat]

this is what I dreamed of all my life.
I used to spend all my income on hiring people for 10$/h on upwork to help me make games

things are different now lol

[unique spade]

oh yes haha

the things i dreamed to do i needed way more resources to pay people to do them. and now guess what. i finally do not need them hahaha

it's crazy you can basically summon an army of all phd's in all domains to work for you

for like 100$ per month

[cedar skiff]

If only it was actually true

[meager bobcat]

there was a day when you couldnt even learn productive practical coding skills without going to an university

I am from 3rd world country and only graduated high school. I had to beg people on IRC to debug my code for me to learn

and had to learn english at same time as c#/c++

[unique spade]

it is true, but you do need to know what you would ask phd's to do if you would have them

[meager bobcat]

phds usually go to bed and need lots of coffee tho

and get sick and tired if you ask it stupid questions or dont fulfil their constantly shifting needs

[cedar skiff]

I person with a phd wouldn't need my guidance.

[unique spade]

i've retired academia at 30, and also quit phd 2 times. cause it was a pointless grind and had nothing to do with actually accumulating knowledge which always has been a cognitive undertaking you ahve to commit yourself , and it's not being granted by some institutional degree

[meager bobcat]

lol fax

[unique spade]

you don't hire them to guide them, you hire them to work for you

[cedar skiff]

you missed the point.

[unique spade]

or you imagine zuck and sam are some kind of einsteins lol

[unique spade]

restate the point then

[cedar skiff]

You can not give an llm the same level of task with complicated ongoing work that requires a large amount of autonomy you can give a human with a phd in their field. The llm will fail with out guidance from an expert.
The llm has the knowledge but not the comprehension.

[quasi hound]

thats the trade off you accept when you have to pay one 6 figures a year and the other $100-$200/month

[uneven kayak]

Yeah, honestly I feel like my entire life has been leading up to this point. Agentic vibe coding has completely changed the way I think about what might be possible or what I'm capable of creating.

[meager bobcat]

everyone seem to be ignoring fact that theres also free tier xD

[quasi hound]

🗿

[cedar skiff]

this is a direct comment to a previous statement, im not saying they are not useful i am saying the marketing jargon that we have a warehouse full of phd level workers is overblown.

[unique spade]

you realize that to be able to correctly evaluate whether that's overblown or not you need to be yourself in a certain position of knowledge right?

[cedar skiff]

i dont have to eat a dirt sandwich to know it taste bad

[meager bobcat]

I am not sure about judging peoples capabilities by their degrees but... I think situation between working with people and working with agents on coding tasks is very similar in terms of capability actually

[unique spade]

lolz such a good analogy

[cedar skiff]

In my field of expertise (SWE) i can replace a lot of the work a junior does, it still cost me time and effort though.

[unique spade]

which is your field of expertise?

[cedar skiff]

swe

[unique spade]

if you don t mind me asking

[uneven kayak]

Yeah just a lot faster. Instead of waiting a day or maybe 20 minutes at least for a response that is only to ask for more time, you get real results with no stress and no one's feelings to hurt

[meager bobcat]

Yeah. I mean I guess it also depends on autonomy you expect
If I am vague and do not carefully consider my prompts then I receive some irritatingly dumb results very reliably

[uneven kayak]

Lol... swe = software engineering

[meager bobcat]

I never managed PHDs before but the back and forth with "no, I wanted this icon on top left, not top left but clipped off the screen" type requests is always there

[uneven kayak]

Oh yeah absolutely. It helps to start with a well defined plan, even if that plan also comes from AI. Claude 4.7 Opus has what they're calling "high horizon thinking" so it's great for planning because it can contextualize the future considerations that a project might need.

[meager bobcat]

yeah something breaks down with openai models if request includes a lot of tasks at once even if its well defined

but never tried claude properly tho

[uneven kayak]

I had good experiences so far, I just ran a 2,200 line plan with 18 steps to completely implement a functional mail system in my game from A to Z and it ran for about 5 hours, and now the system actually works with zero debugging on my part.

That was with GPT 5.5 xHigh

[meager bobcat]

damm

yeah never managed that xD at least not without needing corrections

although I think I attemtped that with 5.4 mostly

I heard thats main improvement with 5.5 right?

[glacial oxide]

Bro limit reset when 😰

[meager bobcat]

^^^ real question

[uneven kayak]

I've always broke it down into digestible phases, digestible for the human that is. But I think I've been underestimating the level of work that I can actually give this agent, as long as you have a very good highly defined plan, then you can give it basically an unlimited number of steps to complete because it's only going to contextualize the most recent one

[meager bobcat]

im at 15% weekly

[meager bobcat]

lol same yeah

[glacial oxide]

I’m just on 20$ plan

The limits used to be good

I guess subsidies have stopped

[meager bobcat]

plus 5.5 is more expensive right? I was using 5.4 xhigh fast all the time last month

now 5.5 xhigh non-fast, but I need to switch to high at most lol

[unique spade]

it was on 2x promo till end of march if i remember correctly

[uneven kayak]

Yeah, so last night was the first test of a really long running multi-phase plan all being executed at the same time. And it worked beautifully, I'm blown away actually. My mind is constantly being blown by agents.

[glacial oxide]

What plan are you on

[meager bobcat]

$200

but I work on a lot of projects at once

[glacial oxide]

Hmm

[uneven kayak]

Technically, tokens cost more, but since it has to use fewer tokens because it has more dense thinking, it actually costs only slightly more than 5.4, and not actually 200% more as the price might indicate

[glacial oxide]

All models from gpt 5.1 to 5.5 have had increased prices

[meager bobcat]

ohh yeah it doesnt overthink actually simple tasks

noticed that in regular chatgpt too

[glacial oxide]

At this rate plus plan won’t get much usage at all

[meager bobcat]

yeahh it feels like they are squeezing pricing for a long while now

but it makes sense since apparently its a lot more compute right

[glacial oxide]

How are you talking so fast

Are you open claw

[meager bobcat]

lol idk just 120wpm

which isnt too much

[glacial oxide]

Bro speaks at 120 tokens per second

[uneven kayak]

Yeah, now we just have the $100 Pro plan promotion until the end of May. So I'm on the $100 plan right now and I'm getting a huge amount of useage, it's even better than my $200 Cursor Ultra subscription which ran dry real damn fast.

[glacial oxide]

Running on cerebra’s

[meager bobcat]

lolol

[glacial oxide]

Cursor is subsidising by 2-3x

Whereas even without 2x, plus plan gives 80$ worth of usage

That’s 4x

For 20$ plan

[meager bobcat]

like, actual cost of inferring? or what that means

[glacial oxide]

API

[uneven kayak]

Yeah that's messed up, I only wish I knew about the pricing nuances before paying $200 for a Cursor Ultra subscription.

[glacial oxide]

But you do get

To use all models

Opus and gpt

[uneven kayak]

True. I'm saving the last of my useage to run audits and formulate plans with 4.7 Opus while I do pretty much everything else in Codex.

[unique spade]

depends...it costs more per token, but at same time based on benchamrks it uses less tokens for same task

[quasi hound]

shouldnt it be the other way around

[glacial oxide]

When antigravity was in its heyday

I used to do opus daily driver and codex for review

[quasi hound]

gpt 5.5 to run audits and formulate plans, and then opus 4.7 for implementation?

idk thats just the setup thats worked much better for me

[glacial oxide]

But even with so much intelligence

They can’t do a lot of things

[meager bobcat]

btw what do you guys audit exactly?
I literally have constant copypasta "Finished with the plan? Implement it fully if you havent already. Any obvious bugs, issues, or long term stability? Investigate and fix if so."
queued up 1-2 times at end of every task unless its something tiny

other than weekly "make sure docs are up to date" task I do nothing else

[uneven kayak]

I even set up a pipeline to use Codex to autonomously generate images using a huge number of models through the Scenario API. So I can say "generate an image that looks like whatever thing so we can use it for whatever reason" and it will decide which model might be best (or use the one I have defined for that particular type of image), then formulate a prompt that adheres to my art policies and generate images for me. It can even evaluate the image to make sure it looks like what I requested and try again if it thinks the image would be unacceptable.

[glacial oxide]

Like I’m not a coder at all so idk how to guide these agents

Gpt 5.5, opus 4.6 none of them could find a memory leak from a script

But my computer science college friend who’s in his first year could find it in less than 20 mins

[meager bobcat]

does it reliably reject wrong results? I had trouble with visual comparisions and judgement type requests before

[quasi hound]

I'm probably a bit overboard with my audit process, I like to test every piece of functionality; so I'll have a main script that has all the logic and a bunch of smaller scripts that test every piece of functionality to ensure its actually working as intended.

[uneven kayak]

It's spooky good. I'll have a thought about an image, then I'll see the agent saying exactly the same things. It can even detect very tiny flaws like only a couple pixels, it's incredible.

[meager bobcat]

yeah thats obviously good way. I need to do more of proper testing

oohh damn haha

like algotrading?

[quasi hound]

the extra few hours of auditing every piece of functionality outweighs the headache I'd get if I shipped something and realized it wasn't working as intended

[meager bobcat]

yea

[quasi hound]

no, its just meme coins

[meager bobcat]

lolol yeah I remember those

hmm yeah I guess you cant just "playtest" stuff if you are working with a blockchain lolol

[quasi hound]

definitely not 😂

I mean you can just freeball it, I was doing that for a while, but like I said; the extra few hours to audit test everything is worth saving the headache

[plush harbor]

its great isn't it? Also codex newbie here, hiiiii

[balmy acorn]

made myself a Hatsume Miku pet

[cedar bear]

/hatch isnt available for me
But /pet is huhh

[noble jay]

does a pro 200 subscription not provide access to 5.5 pro models ? or even 5.4 pro model ?

[cedar skiff]

you get 5.5 pro on chatgpt unlimited, pro models arent avaiable in codex.

[uneven kayak]

Hi! Yes it's absolutely amazing. It's one of the best things that's ever been invented in the history of humankind.

Not even an exaggeration

[noble jay]

i thought there was some monthly limit to messages in to 5.x-pro models in codex

[uneven kayak]

Absolutely love it

[cedar skiff]

This is where you can use it on web/chatgpt.

[noble jay]

yeah but there really isnt any codex mode for it .. thats disappointing

i see gpt-5.3-codex-spark .. that is new for me going from plus to pro-200

how useful is gpt-5.3-codex-spark

(given that it was paywalled by the pro sub)

ive been using 5.4/5.5 a lot recently

[cedar skiff]

It's only avaiable in chatgpt not codex

[uneven kayak]

Only useful for very well defined plans and even then, best to avoid anything too comprehensive or anything that might require making decisions about architecture

It's great to use for asking questions about what is in your codebase

[noble jay]

sounds like a good sub-agent to do gopher work

dont need a frontiner model to run curl/stuff

[uneven kayak]

Absolutely

[noble jay]

@cedar skiff are there any limits on the pro 200 sub with gpt-web doing 5.5 pro thinking tasks ?

how many pro messages/tasks can i have it do

[cedar skiff]

its fair use - or it was when i last looked at it.

How many is that? ¯_(ツ)_/¯

[noble jay]

ok so the stuff i read about there being pro limits per month is probably old or erroneous

[uneven kayak]

There are useage limits but OpenAI is very generous with useage

[queen ember]

a lot of people are luddites when it comes to ai

they don't understand that it can be used to accelerate their workflow

[uneven kayak]

Or even make new things possible for an indie dev, there's no way I would have set up my game to be server authoritative without an agent, I'd have just made it playable offline instead.

[queen ember]

yeah it just creates new opportunities