codex-discussions

[simple star]

It has to mean that

[summer iris]

Has the new codex drop Sam mentioned happened yet?

[ember venture]

All I want is 5.3 on the api so I can use it at work

[nimble terrace]

time to upgrade to alpha 11

[honest saffron]

yes

[hushed storm]

where is the mobile app, probably stuck in review

[simple star]

I might risk the upgrade to alpha... might give me access

[plucky halo]

https://github.com/openai/codex/pull/11602/changes/7715cf07470756babaf0cf231588a995ef6be617

Womp womp

[main nimbus]

can't use cursor/codex? both support no-training for businesses.

[nimble terrace]

npm install -g @openai/codex@alpha

join us

[simple star]

Still nothing 🙁

sad.jpg

[cyan wing]

same

[ember venture]

we use an ai gateway to proxy requests so we give everyone flexibility

[simple star]

Sam, you cant do this. Telling us about auto-software, and then not give it to us

[lean lark]

Different topic: I just had an experience where a GH repo that was authorized for access to Codex was suddenly de-authorized. I found out when Codex was reporting a 403 from GH. I don't understand this at all. Either the DB at OpenAI or GH got corrupted, or there is a timer somewhere. GH security log does not show anything related.
Does this resonate with anyone here?

[simple star]

Different topic? Good luck with that 😄

[green field]

I LOVE CODEX

YEAH

[main nimbus]

feel like cursor solves this problem plus some.

[summer iris]

Wait, so this is less accurate than 5.3-codex on low?

[honest saffron]

its like flash models so yes they are built for speed and comes at a tradeoff

[latent tiger]

yeah, fast but less accurate

[main nimbus]

not to mention, direct API users lose out on subsidized credits.

[hushed storm]

they are probobly intended to be sub agents whenever 0.100 drops

[summer iris]

Does anyone ever use low reasoning? Like who is willing to trade it being wrong more often for it being wrong faster?

[honest saffron]

More like how explore in claude cli uses haiku

[summer iris]

okay, for subagents it makes sense

[plucky halo]

Yes - for making requests that don't really require judgement. Returning values from code etc when I'm feeling too lazy to look

[summer iris]

Problem is the time it takes to switch models takes longer than the time saved for those types of requests.

Even in the CLI where I can hotkey a different model, I seldom do.

[astral pine]

5.3 high plans and steering and helping when 5.3 spark hits a wall... until then it hammers until it works.

[summer iris]

Am I the only one using xhigh? I know it is only .8% better in the benchmarks, so I guess it isn't worth it?

[latent tiger]

usually default to high ngl, its perf balance for me performance/speed

also we have infinite rate limits basically during the 2x promo till april 2 so i dont think using xhigh is a issue now

[simple star]

I have closed and re-opened codex 10 times in the last 10 minutes. I need to touch some grass.

[hushed storm]

same the 2x time it takse to work is annoying

[toxic torrent]

is spark, high?

[honest saffron]

ask codex to write code to check codex till it sees the codex spark model and then send you a mail

[astral pine]

I use mostly high for complex stuff. havent tried xhigh yet tbh. High was slow enough for me lol

[neon girder]

https://github.com/openai/codex/releases/tag/rust-v0.100.0 👀

[plucky halo]

I don't know but it really thinks its rug ties the room together

[hushed storm]

now we just need the mobile app

[hushed storm]

spark is its own model?

[boreal holly]

It's a nihilist. It believes in nothing!

[nimble terrace]

alright this is getting out of hand

[plucky halo]

I'm glad you got the reference!

[boreal holly]

I wanted to say something about Donnie but would've got automodded lol

[simple star]

Is Snake game the new "Hello World"?

[plucky halo]

You're out of your depth, Donnie!

[main nimbus]

always. lit always

[simple star]

I also use only xhigh

[warped pine]

only xhigh

[summer iris]

So codex is obviously moving towards something where you'll use xhigh in /plan mode and then spin off a bunch of -spark models to implement.

[obtuse river]

Copying Anthropic's homework as usual?

[latent tiger]

https://github.com/openai/codex/releases

.100 official release

[finite sorrel]

Rollout will be slow and very capacity constrained.

[neon girder]

slopus 4.6 fast is 2.5x faster at 6x the price
this is 15x faster and they are not even making us pay for it yet (separate rate limits)

[finite sorrel]

codex -m gpt-5.3-codex-spark

[latent tiger]

anthropic getting cooked on all fronts, both open source competition and US private

[toxic torrent]

still waiting for the codex update to get pushed :'(

[simple star]

Aha! v0.100. That will do it. Installing. I know it. It has to work

[neon girder]

u can put any model name u want, it will still use 5.3-codex

[finite sorrel]

🤞

[torpid cargo]

doesnt work yet...

[finite sorrel]

ah fair enough, my apologies 🙂

[simple star]

Damn, it didnt work

[simple star]

Blocked

Somebody ban him

until I get it, please

[plucky halo]

@finite sorrel There was a recent commit that meant it doesn't matter what you put as the slug, it'll still work with the base model

[neon girder]

RUST_LOG='codex_api::sse::responses=trace' codex exec --sandbox read-only --model gpt-5.3-codex-spark 'ping' 2>&1 \           
    | grep -m1 'SSE event: {"type":"response.created"' \
    | sed 's/^.*SSE event: //' \
    | jq -r '.response.model'

this command will tell u the real model, u can periodically check to see if u have access :D

[finite sorrel]

tyvm

[lean lark]

Unrelated to current discussion: VSCode output window always getting messages like these:

[CodexMcpConnection] cli:} message= 
ERROR codex_core::rollout::recorder: Falling back on rollout system
WARN codex_core::state_db: state db record_discrepancy: list_threads_with_db_fallback, falling_back

"Request failed conversationId" / "Failed to resume conversation conversationId" : "no rollout found for thread id"

No clue what any of that is talking about yet.

[plucky halo]

Time to set up a terminal on a seperate monitor running:

while true; do RUST_LOG='codex_api::sse::responses=trace' codex exec --sandbox read-only --model gpt-5.3-codex-spark 'ping' 2>&1 | grep -m1 'SSE event: {"type":"response.created"' | sed 's/^.*SSE event: //' | jq -r '.response.model'; sleep 0.5; done

[latent tiger]

gonna burn your rate limits spamming your model checker

[plucky halo]

2x mate :p

[latent tiger]

tru

[neon girder]

codex increase my rate limits. make no mistakes.

[plucky halo]

Imagine - it finally switches and then I've used up the weekly limit

[plucky halo]

"That didn't work. Try again"

[simple star]

✨ is the new 🍓

[chrome raven]

codex -m gpt-5.3-codex-spark

[neon girder]

still uses normal 5.3

[plucky halo]

codex -m gpt-5.3-codex-you-can-put-anything-here-and-itll-still-respond

[simple star]

Lies

[warped pine]

uhh is that expected for "ping" ?

5.3-codex routed to 5.2

[lean lark]

Codex in VSCode suddenly really adamant about not working. Error messages are useless "submit failed" :

[error] [Composer] submit failed cwd=/opt/codex/repo/my_repo error={} followUp=local mode=local

[neon girder]

https://chatgpt.com/cyber

[simple star]

Ok, I think I need to do a Pull Request and fix this, because it is kinda laughable 😄

[warped pine]

crazy

[neon girder]

refer to https://github.com/openai/codex/issues/11189

[latent tiger]

gpt 6

[plucky halo]

He doesn't even need to type a prompt. gpt-6 just knows

You open your computer and everythings there

[neon girder]

gpt-6.7 when

[lean lark]

Cyber verification is very simple, get picture of ID and then of yourself to verify that you're the person on the ID. Invasive, but they think you're working on security, so required.

[warped pine]

yes im working with security stuff but that's weird anyway

i thought it would redirect only specific prompts

[chrome raven]

nop

[neon girder]

once ur account is flagged, its over

(for a while anyway)

[warped pine]

(and also 5.2 isn't really less capable than 5.3-codex, and in some cases even more capable, so this redirect is weird)

[plucky halo]

Until you verify then you're back in business

[warped pine]

i wish i had my passport with me atm 🙂

[feral elk]

anyone see the new model ?

[frosty zealot]

What is this spark thing

[neon girder]

cerebras model

[simple star]

I am afraid of the model they will release on April 1st... They have the potential to enrage the whole internet

That is too much power for one company

[frosty zealot]

Is it just a faster version? Im just wondering what it does different, like what does 'spark' mean

[wide schooner]

crazy

[lean lark]

Look at your token usage for 5.3 compared to 5.2.

[neon girder]

smaller, 15x faster model

less capable tho

[frosty zealot]

Oh ok

So its essentially a Gemini-3-flash versus gemini-3-pro

[neon girder]

pretty much that, except actually good

[lean lark]

Initially the impression I got was that "spark" was somehow tied to games/graphics. 🤷‍♂️

[drifting granite]

Hmmm I’m in on a pro subscription and on v0.100 and I don’t see the model I’m assuming it’s being rolled out?

[simple star]

@drifting granite Get in the queue. I saw it first. Dont cut the line. Or you might get biten.

[frosty zealot]

They seen my crocs and jean shorts and moved me to the front of the line

[drifting granite]

Honestly I have a great use case for my agents with spark. Effectively a middle man agent for summarizing long tool usage (obv not read)

[feral elk]

I found it

[lean lark]

☝️ "10" 🤣

[ivory zodiac]

you gotta wwait for summe 26 for gpt 10 obviously

[feral elk]

this kind of annoy, I am seeing X new and pull the latest binary and nothing happend 🙁

[ivory zodiac]

you can set the model on plus too

codex -m gpt-5.2-codex-spark

but i'm quite sure its not working

[feral elk]

I prefered to use gpt-10

[ivory zodiac]

oh is that how you did it

i thought you bananad it

lmao

wp

[feral elk]

codex -m gpt-10-codex-spark
we can unlock any model we want

[plucky halo]

Hey - I made this for you. It can materially improve ui from Codex. Explain what you want clearly and the style and you'll get great results.
It creates an image first - tries to build the ui to match that image and does another pass to check parity

Helps if I link to it

https://github.com/StuartAJ/image-first-frontend-build-skill

[hushed storm]

I would rather not give my id to openai

[neon girder]

dw it's probably in their training data already anyway

[lean lark]

For better or worse, they're not getting the ID.. It's a third-party that just confirms to their client that we've been verified.

[glass furnace]

My id is the least sensitive info i have given them..

[hushed storm]

that doesnt make it better

[lean lark]

I don't want to go off-topic but that makes no sense. If your beef is with OAI and they don't get info, then there's no issue. If your beef is with security in general in today's world, then we're completely in agreement. But for now we need to do what's required to get what we want.

[ivory zodiac]

my beef is with you actually

[plucky halo]

@versed fjord

[hushed storm]

I don't want anyone to have my id if that makes it clear

[simple star]

@plucky halo in frontend, we are not still there 😛

[neon girder]

i'd rather a company have my id than the government tbh

[plucky halo]

It's a step up from the white on navy default haha.

[hushed storm]

that makes no sense

[neon girder]

governments suck

[plucky halo]

...but they issue the ID...

[hushed storm]

they made your id 😭

[plucky halo]

They must do it with their eyes closed.

Anyway, i think we're veering off topic!

GUYS

It's go time

[neon girder]

lucky, not there for me yet

[plucky halo]

Snake game literally in 7 seconds

[simple star]

One more enemy

[plucky halo]

See you guys later!

[simple star]

you never played snake? 😄

[ivory zodiac]

all the plebs are talking about snake games

but here's the real unlock

(jk about the pleb thing LOL)

[plucky halo]

Woah

[neon girder]

mfw neither of my two pro accounts have it

[ivory zodiac]

totally joking lmao

but hold up

catching strays for no damn reason 😵

[plucky halo]

I hope spark can dry up my tears

[simple star]

You felt the fear of the banhammer the second you pressed SEND

[ivory zodiac]

nah lmao

i just didnt want him to actually get offended lol

this is the unlock.

the smartest models provide the clearly outlined implementation plan broken down into individual tasks

you dont need the smartest model to implement

only to plan.

spark will kick butt at this

[simple star]

We might be getting way too spoiled, you know? I have been trying to fix the bug of ⁨codex -m gpt-7⁩ and it's taking more then 10 seconds, and Im getting annoyed 🤣

[boreal holly]

"make app pls"
"continue"
"continue"
"app no work"
"pls make app work"

That is the future of AI

[neon girder]

missing the most important one

make no mistakes

[simple star]

future? This is me daily

[toxic torrent]

still no codex app update

[plucky halo]

Nope - just shows as 'custom'

[ivory zodiac]

uh that is the today of AI

😂

i think you got into my clipboard somehow

[vestal monolith]

AYYO anyone has access to spark yet?

[ivory zodiac]

everyone on pro has it

[neon girder]

i just got it

[warped pine]

yep

[vestal monolith]

i'm on pro plan, updated codex mac app + codex cli

[warped pine]

got it

[vestal monolith]

WHERE THO

i cant find it

[ivory zodiac]

try this

[neon girder]

rolling out

[simple star]

Everyone except me, then

[vestal monolith]

YESSIR

[ivory zodiac]

codex -m gpt-5.3-codex-spark

might work

[neon girder]

5.3

[ivory zodiac]

yeah my bad

[plucky halo]

Man, that's a dbugger. Hopefully you get it soon

[ivory zodiac]

skill issue 🙁

nah that's weird

[finite sorrel]

oo

[simple star]

GOT IT!

See you next week fellas

[ivory zodiac]

see told u skill issue

here's the other thing no one is talking about.

[simple star]

SKILL.md
⁨⁨```
Add spark

[hushed storm]

is it a seprate limit

[finite sorrel]

yep

[ivory zodiac]

the benchmarks are lower yeah? but you know what we dont see? a mini model

if this is a replacement for mini...

[vestal monolith]

ow i still don't have it

[hushed storm]

intresting

[vestal monolith]

rekt

skill issue

[hushed storm]

codex -m gpt-5.3-codex-spark

[ivory zodiac]

if this is a replacement for mini, at the same price, its literally amazing.

idk if it is tho

[hushed storm]

it wont be the same price lol

[ivory zodiac]

gotta see the price

probably not

[hushed storm]

I am surprised cerebras even has enough capacity for all codex pro users

[simple star]

Dont troll

[hushed storm]

?

[simple star]

@hushed storm That command gives the impression that it is working, even if you dont have access to the model

(im writing a PR to fix it)

[vestal monolith]

lmao

[simple star]

(I am writing == Codex is writing)

[vestal monolith]

how did you get that tho

[simple star]

@vestal monolith skill issue, so they say

[hushed storm]

it triggers a refetch of models

[hushed storm]

they dont accept external prs lol

[simple star]

@hushed storm the welcome box shows that model being used, even if not accessible

@hushed storm Yes, they do. I already got one accepted months ago

[cyan wing]

I've got it 👀

[vestal monolith]

senpai pls help T.T

[hushed storm]

they changed it

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

[chrome raven]

new

seperate rate limit for spark

[ivory zodiac]

how generous tho

[simple star]

Curious

[ivory zodiac]

go burn all ur usage for us

for science

[simple star]

People be opening Github issues to get access

[chrome raven]

even very fast on xhigh damn

[balmy acorn]

So is spark a new model. Is it a dumb model. For the sake of being thst fast? If so why use it?

[hushed storm]

its fast

that why you use it

[balmy acorn]

Why wpuld I use speed if its gonna ruin all my code.

[chrome raven]

it wont of course

[simple star]

HOLY F this is fast! O_o!

[thin thistle]

spark is out!

[cyan wing]

smart, base model to plan
spark to execute the plan

[plucky halo]

Because you might feel the need

The need for speed

[hushed storm]

Sub agents, maybe that’s why it has a separate usage limit

[simple star]

Codex is literally taking more time on reading files, than on inference

My hard drive has become the bottleneck o_O

[chrome raven]

like tooooooo fast damnnn

[balmy acorn]

Idk if yall coders or not but. I dont think you guys realize that if it produces more errors in the codebase because its speed over detail. That just gives me even more work to go back and debugg

[chrome raven]

sub agents will be so damn sweet with spark

[chrome raven]

have you used this model?

[frosty zealot]

Anybody have access to this model?

[cyan wing]

can you set the subagent model tho?
I thought it just uses the current model for its subagents

[simple star]

@frosty zealot All the pro users have access now. It is 100% rolled out

[frosty zealot]

I feel like smoke testing with mcp's with spark will be sweet

[ivory zodiac]

its weird man.

it doesn't always follow

[chrome raven]

i think subagents use parent model

[balmy acorn]

I do not want too until I see proof in other people's large coding tasks

[potent mason]

Wow codex-spark this is going to be insane

Has anyone tried the quality of it?

[thin thistle]

not good, cant make flappy bird w medium in one shot

[potent mason]

What about the pricing?

[balmy acorn]

Why your doing medium level reasoning

[thin thistle]

they used it in the demo vid

[potent mason]

huh I get it in the cli but not the desktop app

[drifting granite]

wow its so fast.

using it in the cli is like holy moly

im running the ol' make pacman test on it.

[astral pine]

Let's go!

[plucky halo]

Update is out for Codex Mac App

[chrome raven]

niceee

[toxic torrent]

Hell yeah lets go

[frosty zealot]

I picked the right time to hop over from Claude

OpenAI be cookin

they're always anime profile pics

No offences @chrome raven you seem pretty normal 😛

[drifting granite]

Ooof the Pacman test was rough lol

Didn't one shot it, and broken gameplay. thats okay def isn't the main use case.

[chrome raven]

xhigh?

[drifting granite]

Yes, on Spark.

I had to tell it not to use the existing codebase from other models lol

[vestal monolith]

which version?

[plucky halo]

@vestal monolith

[vestal monolith]

DAMN got it

time to go BRRR with opencode OmO

[simple star]

v0.101 already out o_O

Do this people ever rest??

[vestal monolith]

switching all my subagents to spark rn

[vestal monolith]

lmao my ATH was few days ago - 800m opus 4.6 and 150m codex 5.3

[frosty zealot]

https://github.com/PanicIsReal/codex-webstrap

I created a wrapper for the MacOS desktop client, it lets you run Codex App in your web browser, I made it cause I'm usually remote from my machine quite a bit, and didnt want to use the CLI cause of the creature comforts of the app, this solves that problem for me.

[balmy acorn]

it not having image support sucks alot

[scenic umbra]

"Did you try GPT-5.3-Codex-Spark yet?" lmao they really think it's that easy to try $200 + chance it actually rolls out to you

[chrome raven]

this is the first version of them on a new stack, let them settle

[scenic umbra]

Actually? Damn now I need to actually decide if it's worth $200 a month lol

[simple star]

😄

do eeet

[scenic umbra]

In theory it pays for itself in a couple hours ig

[stoic zenith]

It has a problem right now... it’s really slow

[scenic umbra]

Really?

I think it's running on special hardware, maybe they need to scale it up more

[blissful basin]

it... is.... fast...

[stoic zenith]

thinking for like 5 min

[scenic umbra]

lol

[stoic zenith]

and do nothing

[cedar bear]

WHat exactly is Spark?

[simple star]

Fellas... I am seeing the "Context used %" dropping... like a rocket losing fuel

[scenic umbra]

What if you give it a real coding prompt?

[simple star]

It can drop 3% in-between eye blink

[stoic zenith]

i gave

[drifting granite]

its only 128k

[ivory zodiac]

its 128k

[simple star]

@cedar bear It is codex going BRRRR

[scenic umbra]

Yeah I think it only has 128k

128k in 2026 is insanely low lol

[cedar bear]

so much better then the "old" version?

[stoic zenith]

after I stop it i want it to run again.

[balmy acorn]

its 122k context window

[simple star]

@cedar bear Depends what you mean with "Better". It is definitely faster. DEFINITELY

[scenic umbra]

Will be interesting to find out if the quality is similar for most common stuff

[buoyant oxide]

what's the difference between Codex and normal GPT?

[stoic zenith]

codex for code

[scenic umbra]

Codex is specifically for coding and terminal usage

[stoic zenith]

i think the main difference is that codex can control your computer.

[buoyant oxide]

I often want GPT to look at a couple source files but not specifically from a repo

[stoic zenith]

by using terminal

yea

[buoyant oxide]

idk how to do that

[scenic umbra]

You can do that with Codex

You can give it access to a folder on your computer

[buoyant oxide]

oh really?

[scenic umbra]

yeah

[buoyant oxide]

can you do that on the free mode? lol

[scenic umbra]

yes

[stoic zenith]

yes

but only this month

[scenic umbra]

I also got Plus for free this month

[cedar bear]

Just with the CLI or with the Codex App too?

[stoic zenith]

both

[scenic umbra]

Codex app too, in the app you basically use the folder as your project

[stoic zenith]

codex app just a GUI, maybe makes you feel better.

[scenic umbra]

yeah

[cedar bear]

ah
So its not bound to repos anymore?

[scenic umbra]

Yeah it lets you run stuff locally

[stoic zenith]

you can, by using git

[scenic umbra]

Codex app is much nicer than the terminal version imo

[cedar bear]

uhh
So bad that im a windows user atm 😭

[scenic umbra]

You can use repos but it's not bound to it like they were asking

[cedar bear]

But the web version is bound to repos right>?

[scenic umbra]

You can run the terminal version on Windows I believe

[stoic zenith]

true

[potent mason]

We need some benchmarks and specs on spark

[buoyant oxide]

Would Codex respect my GPT instructions on how to format the code or you have to re-enter them?

[potent mason]

Freaking 100% context filled up so fast

[stoic zenith]

yes

[scenic umbra]

Hmm never used the cloud version so not sure how it works, I assume it uses repos

[potent mason]

Ah I see why it has a 122k context

[buoyant oxide]

mhh yes to re-enter them or it will respect them?

[hoary turtle]

Is the codex 5.3 spark only in CLI ?

[stoic zenith]

both

[hoary turtle]

CLI and IDE ?

[stoic zenith]

in cli and app

[hoary turtle]

Ah yes, doesnt include IDE i guess

[stoic zenith]

Rolling out today to ChatGPT Pro users in the Codex app, CLI, and IDE extension.

[neon girder]

rate limits are quite low compared to normal 5.3-codex ngl

[stoic zenith]

https://x.com/OpenAI/status/2022009583653576787

[worthy crest]

i tested..... 5.3 works better for me

[worthy crest]

i did some code... not working. fallback to 5.3 and it see what was the bug.

[hoary turtle]

Ohh, im on my Plus personnal account

So if i switch to my pro account, it'll be there?

[stoic zenith]

i guess so

[hoary turtle]

Alright

[stoic zenith]

i dont really use extension so not really sure

[drifting granite]

fwiw this model is not for coding.

Idk why they advertised it. It should be for using/interpretting agent tools.

[gilded galleon]

I am also mildly disappointed so far , really strange , also on xhigh

[buoyant oxide]

oh it's editing my files?

[chrome raven]

it's a small model too

[stoic zenith]

yes it will

[drifting granite]

It's got some valid use cases, but if your expecting workable code then you will be dissapointed.

[stoic zenith]

mainly because the performance is not actually better. if it is just simply faster, I do not know what scenarios it would be suitable for.

[balmy acorn]

thats what i thoguth

[vast geode]

Because I don't have it. ?

[potent mason]

For a lot of repetitive tasks that don't require the model to be brilliant

[hoary turtle]

Seems like a simple lightweight version of Codex 5.3

[wide schooner]

i dont have spark in my cli bruh

[stoic zenith]

pro plan? or update the app

[wide schooner]

its not much dumber and its WAY faster

[frosty zealot]

Is github intermitently lagging out again

[neon girder]

honestly the model is good for what it is: a proof of concept of the speed the openai/cerebras collaboration can achieve. and it would be very useful for a lot of things, but the rate limits are just too low right now.

[potent mason]

As long as it's seperate from the main one I don't really mind

[chrome raven]

exacccctly

[vast geode]

Is this only available for the Pro plan? I have the Plus plan.

[potent mason]

Yeap only pro plan for now

[hoary turtle]

Yes, pro Only for now

[stoic zenith]

yes only pro for now

[balmy acorn]

yea no timpressed with spark...

sure its fast but im getting more errors in my code than with normal codex

[hoary turtle]

Nothing can be perfect. Faster work means more errors and inconsistencies, just like humans to be honest.

[stoic zenith]

actually I quite enjoy the slow speed, because i can spend the time on other things

[balmy acorn]

yea. all of these people that want these models to go super fast shows they have no idea what they are doing

[tame magnet]

is codex spark as good as normal 5.3 codex but just faster ?

[balmy acorn]

it is not better

[plucky halo]

@simple star How about now? I've pushed an update to the skill :p

[gusty basalt]

Darn, from the video in the X post in #codex-updates I was getting the vibe it was better

But I guess not

[simple star]

me? what?

[stoic zenith]

lower performance but faster

[balmy acorn]

^^^^

\

[tame magnet]

ok ahah yeah f it

[last stream]

coooking with spark

[hoary turtle]

That looks promising ! Is this an homemade skill by you ?

[balmy acorn]

bascially its codex 5.3 mini with a new word attached to it

[plucky halo]

Yeah - https://github.com/StuartAJ/image-first-frontend-build-skill

Genuinely works weel

better than my typing anyway

[chrome raven]

so with good frontend skill codex is good at frontend

[balmy acorn]

gpt 5.3 codex x-high built me a pokemon game.... that works

[chrome raven]

this looks cleannn nicee

[glass furnace]

geez ths context window is killing me lol

does it not auto compact?

[hushed storm]

it should

[glass furnace]

it isn't for me. i keep having to compact myself lo

[chrome raven]

cli? or app?

[glass furnace]

cli

[chrome raven]

it auto compacts.

[tranquil hare]

It's worse. But it's much faster than anything you've ever seen

[lean lark]

Apologizing to OpenAI for saying there were no release notes for Codex.
https://developers.openai.com/codex/changelog
I've been having a lot of issues with fast updates from 0.98/99/100/101. (mostly with stale thread entries)
I hope things level off.

[glass furnace]

ahhh i changed my model auto compact in the config in the past, so my threshold was higher than the new context limit lol

[main nimbus]

Yeah and the new model is context hungry. Bet that was frustrating

[glass furnace]

Oh it was lmao

[wide schooner]

making a bootable os with ui from scratch ate 3% of my weekly quota, outrageous

[simple star]

Im compacting context like a madman

[wide schooner]

thats good, although codex is insane at memory so its not as needed as with everything else

[torpid trout]

so this -spark... one only gets it if is a special snowflake? Or how does it work? 😄

Since it is not available in the /model command even on 0.101

[wide schooner]

its compressed 5.3 thats not MUCH dumber while WAAAY faster

[torpid trout]

yeah, i saw that, but... still, its only on twitter, not in the real world for me

[main nimbus]

Pro users atm

And apparently some API customers

[torpid trout]

Ah I see.
Anyway, does not seem very promising anyway lol - fast but more nothing useful I guess.

[plucky halo]

It's being rolled out to users on the pro plan. If you're not on that, you won't get it until it's ready for release.

[eternal spade]

You can just build things—faster... sometimes... sometimes I just fail at doing anything at all 📣

[stoic zenith]

i do believe they have gpt-5.3-codex-max here

[hollow stirrup]

Any word from the OAI team on 5.3 for API users?

[main nimbus]

I’m desperately wanting this, not necessarily faster inference (5.3 was already 25%!!!)

[plucky halo]

Depends how you use it. Create a comprehensive plan using 5.3-codex xhigh. Make sure you work through it properly.

Hand off to Spark for quick execution. It's not a model you'd use without proper instruction

[torpid trout]

Guess what, look at the background... this app is totally unrelated, monitors a btc node, made with guess what. It clearly is not creative lolol
In a few months we will not only be able to tell AI code by reading the code but by looking at the UI

So in other words that skill is meaningless, and GPT 5.3 has a very strong opionion on design ahahah

[eternal spade]

currently taking about 10 times to create a PR... and it's denying letting me add anything to it either. Just fun good times is all.

[ivory zodiac]

https://x.com/LLMJunky/status/2022059906841350357?s=20

[plucky halo]

What was your prompt?
Also how about not talking to people like an utter cockwomble?

[gilded galleon]

the pelican benchmark spark vs regular

https://x.com/simonw/status/2022057896121708735

[torpid trout]

"create an app for monitoring my BTC node, design is up to you, framework too"

[wide schooner]

1 prompt btw

[torpid trout]

No idea what you mean by the utter cockwomble - literally in this channel you cant talk bad, so, no idea what the f you mean 😉

[ivory zodiac]

base 5.3 isn't great a creativity, why would the smaller version be better?

[wide schooner]

bootable os

[ivory zodiac]

use claude for ui

simple

[wide schooner]

use kimi for ui

[ivory zodiac]

kimi as well

[wide schooner]

also opus cant 1-prompt a bootable OS like codex just did

[ivory zodiac]

minimax 2.5 looking solid too

[simple star]

Apparently there will be Codex App for Windows tomorrow (for some)

https://x.com/ajambrosino/status/2022048708331024579

[ivory zodiac]

that news is so 2 hours ago bro

keep up

okay maybe just 1 😂

[simple star]

I. CANT. KEEP. UP.

Jesus, this is madness, really

[ivory zodiac]

honestly

they released all kinds of stuff today and the algo wont even let you see it all

[simple star]

I was gonna say the Codex people must be doing drugs, to be able to release this fast. But I am starting to believe they are all bots.

This speed of development is just incredible

[tranquil hare]

it's neither. They just use Codex to develop (Codex). And thus they're much faster than any other SH

[torpid trout]

When you see how projects like Grafana have AGENTS.md in the git repo, I am not surprised if the entire codex is created with codex/gpt

[plucky halo]

@torpid trout Mine was quite different. This skill isn't just about "creative olololol" - it ensures WCAG 2.2 AA standards are respected. That there is always a mobile-friendly version.
I made it to support someone else. If it doesn't work for you, that's absolutely fine. But you don't need to be so caustic

[elfin summit]

gemini3pro for UI

i think

[ivory zodiac]

lets be real, its all slop

[scenic umbra]

Do you guys think spark will eventually come to the Plus plan?

[ivory zodiac]

doesn't make it bad though

[supple perch]

Until you see devs that make this style on their own as it looks nice

[ivory zodiac]

6 months ago it couldn't do any of this

[elfin summit]

what is?

[torpid trout]

Nah, you do not understand what I mean, go back and read the message, I guess. Or, let it be - no one is attacking you here.

[ivory zodiac]

all of it lol

[simple star]

The next big barrier for models is getting good at UI

[elfin summit]

they already are

[ivory zodiac]

you can always tell. some better than others

[simple star]

@elfin summit they are "meh"

[elfin summit]

but not for vibe coders

[scenic umbra]

Or at least they'll have an API that's cheaper than $200 a month?

[elfin summit]

You can use gemini3pro to plan your UI visually, then turn them into assets, and implement the UI with codex5.3

[simple star]

@scenic umbra Eventually, yes

[supple perch]

These are great, just need to give them the right tools

[elfin summit]

nanobananapro is amazing

[scenic umbra]

nice

[simple star]

Gemini3 is better at UI, but also not that great. A true understanding of aesthetic and UX is not quite there

[plucky halo]

@torpid trout
The entire way you phrased that was meant to belittle and put down. For literally no reason. Someone was trying to help someone that was struggling with Codex's UI skills. And you first thought was "fc this guy - I'll try to make feel as little as possible"

ahahah lolol guess what

[torpid trout]

OK master, if you say so, it must be so!

[plucky halo]

That's completely different to "I don't think your skill.md is doing what you're intending. I got the exact same background"

Hope you can see the difference and that you have a lovely day 🙂

[ivory zodiac]

you were being a bit of a negative nancy

go back, read for yourself

[plucky halo]

Let's not be Debbie Downers @torpid trout - let's be friends

[elfin summit]

it does understand aesthetics well

but you gotta be the judge

[ivory zodiac]

https://tenor.com/view/debbie-downer-nbc-snl-saturday-night-live-saturday-night-live-gifs-gif-10284698

[torpid trout]

Yeah, I see, no one likes to hear how their creative creation is not so creative after all.
But... all good man. No probs, was just a background being 1:1 identical - what a coincidence 😄

[ivory zodiac]

what a 🤡 he never said it was creative. he was trying to help the guy get better outcomes.

which is true

you can guide the model to better outcomes.

does that bother you?

his UI looks WAY better than yours

maybe its a skill issue

[plucky halo]

It's absolutely not that at all. My creative creation was literally just prompting gpt-5.3 a few times.

What's not nice is the fact that anonymity on the internet sometimes causes people to be absolute not-very-nice-people for no reason

[ivory zodiac]

🤷‍♂️

[torpid trout]

jesus effing c
Anonymity? You are the guys hiding behind cryptic names lol

[ivory zodiac]

one of you was trying to help someone

[plucky halo]

Yes, my real name is Montgomery

And Montgomery alone

[torpid trout]

Cool. Do you want to go on on the topic which for me was closed about 20 minutes ago?

[ivory zodiac]

just do better

we're all here to learn

have some personal accountability

[plucky halo]

More than happy to

[ivory zodiac]

DAMN IT MONTGOMERY

anyway its all love

codex released soemthing else today too

new sandbox toys

[main nimbus]

https://tenor.com/view/captain-ocean-pengu-penguin-pudgy-gif-13181729964263281871

they just can’t stop shipping

[astral pine]

Anyone try spark with playwright? is it any good at browser?

[ivory zodiac]

agent browser better imo

its a variant of pw

https://github.com/vercel-labs/agent-browser

[astral pine]

Yeah thats the variant i use… im on windows though so it always gives me trouble tbh even with the strong models

[ivory zodiac]

what trouble

[cyan wing]

so can i use this on Windows? or do I still need to WSL + bubblewrap?

[astral pine]

I gave it its own chrome profile and had it starting in non-headless mode and connecting via CDP, but the daemons always down and the agents like always trying to reinstall chromium. It’s a nightmare.

[ivory zodiac]

i wish i knew bro. i'm on linux.

it should

[proud kraken]

is it just me or does spark end up in an infinite loop of read until it fills context, compact, read until it fills context, compact, ...

[ivory zodiac]

did you ask the agent to help you troubleshoot?

[astral pine]

yes lol. i have like a 2000 line md from it testing things and documenting its learnings on all the quirks... still sucks.

[ivory zodiac]

that sucks. there's another one you can try

okay well 2

first of all, google's computer use model is badass.

and then there's anotehr called browser use https://t.co/qBFmPUsw72

[astral pine]

like the anti gravity browser? i do use that sometimes

[wide schooner]

it runs on 2.5 flash lmao

[ivory zodiac]

and?

[astral pine]

I even sometimes have codex drive claude extension and do it that way i just only have $20 claude plan

[ivory zodiac]

i dont think its even flash. its its own model, and its very good

a bit token hungry and slower

it does everything step by step by step

i use it to automate all kinds of things that dont have apis

[wide schooner]

lol its not selectable but its usable?

[ivory zodiac]

it wont work

[wide schooner]

e

[misty bough]

Good Afternoon I currently Host my Website on my Game Servers Dedicated Machine and I am updating it manually by Sudo Nano into it from my Terminal. I know there has to be an easier way using codex > Vscode > Github to update the website Can anyone advise me on how I can do this?

[astral pine]

through anti gravity or do you have some code going around it?

[wide schooner]

what if spark is a myth

[ivory zodiac]

https://github.com/google-gemini/computer-use-preview

it still uses playwright

i do it through a script

[wide schooner]

LOL GIVING GEMINI ACCESS TO PC

[ivory zodiac]

just ask codex to help you build a script

[wide schooner]

BRO ID RATHER EVEN MISTRAL

🤣

[astral pine]

oh interesting. ill give it a try!

[ivory zodiac]

yeah try it. hopefully playwright behaves a little better this way

[wide schooner]

https://blog.google/innovation-and-ai/models-and-research/google-deepmind/gemini-computer-use-model/ that was year ago btw

[astral pine]

if it works it works... all i know is im not having very much luck with agent-browser... i dont think who ever is building it is very focused on windows quirks.

[ivory zodiac]

its 4 months old

are you 12?

[ivory zodiac]

oh

you can use browserbase too

if playwright doesn't work for you

You wanna prompt it just like you do a normal LLM. I'll show you a sample.

[plucky halo]

He's just changed and manipulated time. He now gets 21 days in a week.

[ivory zodiac]

lol no the model itself is from oct 25 but the tool i shared was released relatively recently

it doesn't matter, its not rocket science

the model doesn't need to be SOTA to click mouse button and scroll

[frosty zealot]

Why people don’t just use chrome devtools mcp is beyond me

[ivory zodiac]

same reason he can click mouse and scroll

you dont need to be that smart 😂

[astral pine]

ill give them all a try lol. i know playwright cli came out a few days ago too i havent tried that yet either... they're likely all the same base code just different ways of doing the commands. i see gemini in that blog post about computer use links to browser base: https://gemini.browserbase.com/ so they're likely all very similar in approach.

[ivory zodiac]

got em.

I have a script that automates ad scheduling for me, it iterates over a list of dates.

[astral pine]

People still use mcp??? jkjk

[ivory zodiac]

I prompt it just like this

Objective: Check for login, find the first 'earn' ad for the current month, and reserve it.\n\nVariables:\n* [CURRENT_MONTH]: The full name of the current month (e.g., "November")\n* Task Steps ---\n\n1. Check Login & Navigate\nOnce the initial page loads, check if it is the login screen.\n* IF it is the login screen: Pause for user email input, then pause for user password input. Once the user submits and the main landing page loads, proceed to Step 2.\n* ELSE (if already logged in): Proceed directly to Step 2.\n\n2. Main Landing Page: Find Placement\nOn the main landing page:\n1. Type 'earn' into the top-most search textbox and submit.

test your workflows to make sure they work right and they are pretty much set and forget from there

[boreal holly]

Heck yeah! I been doing something similar with Flutter. It lets you render widgets without a GUI using "pump & settle", and it's really nice having Codex lock in behaviors with unit tests. Or if it's like "when I load this page, the backend gets spammed with get requests in perpetuity. Use integration tests to troubleshoot page load and controller lifecycle."

[ivory zodiac]

oh yeah dude. i'm experimenting with some stuff now

i just had codex run through an entire checklist of tests

it failed miserably

😂

but claude did a pretty okay job. just doing UI/ UX stuff

login, listing, delete post, etc. Just all the tedious stuff you gotta do before you push to prod

[wide schooner]

glm sucks js tested parsing a pdf into html, glm didnt even make a proper table

codex nailed 1st try

[boreal holly]

Yeah stuff like "when I log out and log back in it doesn't clear the token". Or my favorite is "now that we have this domain locked down, and synchronization for everything here works perfectly and proven with tests, let's sweep through the other domains and match controller and view behavior, confirm with tests" like it's doing right now. And then you fire up the app and it works flawlessly 🤓

[ivory zodiac]

love that

using cursor?

[fallen kettle]

If you’re looking for a good way to control browsers, I’ve spent a bunch of time trying to squeeze all the extra perf and capabilities out of Chrome here: https://github.com/btraut/browser-bridge

[ivory zodiac]

neat

[boreal holly]

Codex app + Zed. Sometimes Sublime instead of Zed because the language servers compete with Codex (locking cargo build dir)

[astral pine]

Any chance your developing it from windows?

[fallen kettle]

I haven’t tried, but I think it should work. I’ll give it a shot tonight and see.

[high girder]

for something like this, you could run this in wsl if you're unsure, but 99% of that has a windows version of it, so the only real issue you need to look out for is powershell vs bash differences. Tell codex to run it on windows and you'll be fine

[lavish dawn]

I even got the raspberry pi

[lavish dawn]

how lucky 🎉

[ivory zodiac]

LUCKY

https://tenor.com/view/key-and-peele-son-of-abitch-upset-gif-4276606

[latent tiger]

DeepSeek definitely about to drop a banger in a few days lol

[high girder]

Holy crap I got the codex merch email.

[ivory zodiac]

i didnt get jack 🙁

💔

[latent tiger]

Me neither rip

[plucky halo]

ditto 🙁

[latent tiger]

Just going to print my own codex logo on a hoodie tho

No one will know

[high girder]

I'm more interested in that raspberry pi 5 codex kit

[lean lark]

Have spent most of my time today trying to recover from Codex / Codex CLI updates. Lots of different errors reported in VSCode output window. Time in GH Issues, config files, logs. Real cluster.... of a time waster. 👿

[frosty zealot]

https://www.npmjs.com/package/codex-webstrapper I made a wrapper for the Codex app that 1:1's it in the browser just fyi put it on npm for quick install & run (requires a mac at the moment because it needs the electron app running for it to work) bun install -g codex-webstrapper@0.1.5

My spark usage like hasnt moved at all, anybody have an idea what the usage is like compared to regular 5.3?

[cyan wing]

grats ! how quickly did you sign up after the ad dropped?

[high girder]

I probably did it in about 15-20 minutes, but I also made it harder for myself because I thought it was specifically for the codex app, so I was trying it from the windows port I did. It still made the skill and got me there, but it was just a lot of effort that didn't need to be spent because I didn't pay enough attention

[thin sorrel]

I got mine too, registered after like 50 minutes

[frosty zealot]

is it too late to sign up?

[ivory zodiac]

i must have just missed it

brb just going to take a long walk off a short pier

[frosty zealot]

Where did you registeR?

[high girder]

https://openai.com/codex/youcanjustbuildthings/
The link you got when you did it through the skill had your account email tied to the link, so you're not getting that link lol.

[mellow echo]

anyone else having issues with 5.3-codex being utterly incompetent compared to 5.2-codex?

[frosty zealot]

Ah I musth ave missed that

[frosty zealot]

https://github.com/openai/codex/issues/11189#issuecomment-3880522742

[mellow echo]

Thanks! If that were the case, I'd probably be less annoyed 😄 5.2 actually worked very reliably compared to this

[main nimbus]

I JUST GOT THE MERCH EMAIL BROOO

[cyan wing]

SAMEEE!! thank you @chrome raven ❤️

[glacial agate]

❤️ u codex

[lean lark]

I wonder if they're going to put up an easter egg and offer merch to Windows/Linux users. 😉

[glacial agate]

me 3

[high girder]

you didn't need to be on the mac version. I'm a windows and wsl user

[glacial agate]

i wonder how distribution of who gets what kit worked

[scenic umbra]

👀 I got the email!

[glacial agate]

i failed to put my codex email in soo i got basic

but shiii cant be mad 😄

[scenic umbra]

I got Core Kit, I don’t think the email you put matters, it’s random

[glacial agate]

if i took a super wild guess, or i designed it id given top gifts to power users

just didnt think of it at the time and didnt want to put in multiple under the same name and get removed XD

[scenic umbra]

Do you think the order number is actually in order of who completed it fastest?

[cyan wing]

😋 😋 😋

[scenic umbra]

What does it actually do? lol

[glacial agate]

so using this in wow

[lean lark]

Say What? I thought they only added the skill into the app, and sincee the app is only available on Mac, then so is the merch.

[scenic umbra]

Does it just launch the Codex app?

[lean lark]

Oh wait - @high girder ported... nvm

I'll revise my statement - I wonder if they'll offer to Windows/Linux users ... who don't do a custom port of the app.

[scenic umbra]

I assume you’ll be able to program the macro key to do whatever

[high girder]

I'm pretty sure you could do it from the CLI or vsc extension. It was a skill setup

[glacial agate]

or ... new power button 😄

[scenic umbra]

👀 That’s so smart

I can’t be bothered reaching down to turn on my PC lol

[glacial agate]

now only if i turned mine off XD

[main nimbus]

I wonder if it was really the first 1k people or there was more

my order number says I'm either one of the first 100 or they let a LOT more people in.

[balmy acorn]

So I was right it is a smaller model. Lolol yea not gonna use this

[glacial agate]

anyone have a screenshot of what all the rare reward options were for the merch drop? & it was first 1k, and def more incase of bots/dupe address/name/etc

[main nimbus]

it's called elite kit but they didn't show the whole thing

[main nimbus]

this one

[wintry crypt]

Lucky

[wintry crypt]

The way they just giving away a single macro that probably is enter lol

[glacial agate]

https://www.amazon.com/Programmable-User-Defined-Button-Customized-Combination/dp/B08SQGWZN4

its programable (i assume its this product just retextured keycap)

[scenic umbra]

Probably something similar to that yeah

[wintry crypt]

lol 💀

[wintry crypt]

Needa hop on the meme make it say lizard

lol

[tall harbor]

wait you're telling me a model that performs much better than the flagship is actually just smaller??? shocked

[stone aspen]

yuy shipment mail :D

[ivory zodiac]

makes no mistakes

[rapid patio]

why is codex so dumb now? two days ago it did everything i asked it to do, and now i told it to fix a problem in my python code, and it broke more things

[ivory zodiac]

they say AI is a mirror

H O L Y

[rapid patio]

🪞

[stone aspen]

I can finally wear clothes again after I receive the kit

[balmy acorn]

Its not betrer?? Who told you it was better??

[ivory zodiac]

oh god i need to comb my hair

[bright swift]

😎

[ivory zodiac]

UGH i'm jelous

i have a pi 5 so whatever but i want a codex version

[bright swift]

I was sure i missed the first 1000 window, pretty surprised to see this now

[ivory zodiac]

so jealous i cant even spell apparently

how many digits is your order number

i suspect they ended up blessing a lot more than 1000 ppl

[bright swift]

29xxx

But probably other orders in there if i'd take a guess

[ivory zodiac]

same

29xxx

[bright swift]

Apparently there was another hidden layer if you won that minigame, but i only gave it a half hearted try 😄

[wintry crypt]

Lucky

[bright swift]

skill

[glacial agate]

lol im just happy i got stickers and a macro XD

[bright swift]

I remember i felt pretty smart for stopping and zooming in. Then after i submitted, i saw all the codex' team not so subtle tweets about it that i hadnt seen and lost my confidence 😄

[wintry crypt]

Def

[glacial agate]

it shall go in the stack with the supreme stickers XD

[tall harbor]

Performance as in speed

[nimble terrace]

woaho

this is sick

[ivory zodiac]

lets freaking go

[nimble terrace]

this is for the first 1k people right?

[main nimbus]

apparently so

[nimble terrace]

im probably going to forget about it because it ships in 4-6 weeks

it'll like fall out of my mind

then when the package comes in i'll be confused asf LOL

[balmy acorn]

watch it never get delivered lost in shipping is usally how these things always go

[nimble terrace]

LMFAO i'll be so upset

im in the UK too and with the royal mail that damn well might happen

you might've just spoke it into existence

gonna pray that doesnt happen lol

[urban falcon]

I want to use codex with Minecraft modding just to see what it can do, can it recognize and use classes by itself or do you have to give it the name?

[main nimbus]

Codex reads and understands your codebase. The more help you give it, the better the result.

[lean lark]

@urban falcon Give it links to the library that you want to use: Paper, Spigot, etc. Also tell it the versions you're using so that it uses the right libs for all the right classes. (as you might be able to tell from my picture here, I'm kinda into Minecraft too.)

Oh, you said modding and I was talking about plugins ... you know what I mean. 😜

[willow brook]

algun codigo

xd

xd

xd

xd

xd

xd

[languid perch]

Does anyone else have problems with codex(browser) with loading and lag? Any suggestions?

[potent mason]

Is there something like a reject list in Codex? I don't want it to ever read my .env

[loud dragon]

Managed to get mine too

[ivory zodiac]

they i suspect they let in way more than 1000 people

[potent mason]

Wait what how do you get that?

[loud dragon]

Easter egg from the Super Bowl commercial. Install a custom skill called Build Things, run it, generate custom url to claim

First 1000 got it

[potent mason]

Ah well it's way too late for that xd

[jaunty talon]

seems like way more than 1000 got it

[lean lark]

hehe - I tried to install to Linux with v0.101.0, it's not in the curated list. 🥹

[lean lark]

No - if it's in the scope of the current workspace then the assistant can/may/will read it. You may need to change permissions to a different user/group and only change permissions back when the app needs to read it.

[stray swift]

I got the hours I didn't spend watching the Superbowl. 🙂

[drifting granite]

Found a fun use case for Spark

I’m having a hook output from my server to a endpoint on my Mac, which then summarizes Codex’s response for me using Spark and a local tts model plays it.

[stray swift]

Actually, I just checked my spam trap, and I'm getting some minor swag: A cap, a keychain, and some stickers.

[drifting granite]

Oo when did they send out the email?

[stray swift]

7:25pm US EST.

[drifting granite]

Dang not seeing it, I filled it out during the Super Bowl. Maybe I’ll get a suprise later lol

[near gorge]

sadge, maybe i redeemed merch too late

[unreal parcel]

is it just me or the TUI is super slow when typing now since 0.101 (if the terminal has sufficiently large chat text)

[potent mason]

Windows or Mac?

[ivory zodiac]

oh i didnt watch either heh. i was coding lmao

i found out in here

i like basketball. football not for me

i did end up getting an email

More easter eggs

[cyan wing]

gpt-5.3-codex-spark is current down
a.k.a. the Cerebras special ™

[unreal parcel]

windows

actually it's slow every with an empty terminal

[potent kestrel]

thought i was going crazy lol

[unreal parcel]

works for me

[hard tulip]

not for me

[hard tulip]

Cerebras cant keep anything online for more then 12 hrs

[potent mason]

Yeah windows is terrible atp with cli's

[hard tulip]

just a few hrs ago

[hard tulip]

i don't think it being on windows has anything to do with how the back end is connecting it.

[potent mason]

No but the front end definetly is and you asked about typing (which should be front end)

[hard tulip]

for the codex CLI ???

[potent mason]

I was talking to Cleroth

[unreal parcel]

my point was it was working blazingly fast before this update

[potent mason]

About his slow typing issue

[hard tulip]

I bet you like 4o

its crashing out

[cyan wing]

#keep4o 💔

[ivory zodiac]

https://tenor.com/view/liam-neeson-i-will-find-you-taken-gif-5593626

[earnest walrus]

Anyone else in an update loop trying to update codex cli?

Just runs over and over on startup of codex

[ivory zodiac]

just install it manually

[frosty zealot]

Anybody find any good edge cases for spark yet?

[unreal parcel]

asking questions

[hard tulip]

good work team

spark is completly removed from the model list lol

[high girder]

wasn't it only available to pro members at the moment? or was that something else?

[frosty zealot]

It's removed from the app as well, and I had it earlier, it was just a preview model maybe it wasnt working as intended

[lone island]

PRO-sub: The 'gpt-5.3-codex-spark' model is not supported when using Codex with a ChatGPT
account

[steel sierra]

I am on a PRO subscription, using CLI on Windows > WSL updated to OpenAI Codex (v0.101.0), Can see 'gpt-5.3-codex-spark' in /model chooser. Tried testing it with a text only task and I get this error ' Error running remote compact task: {"detail":"The 'gpt-5.3-codex-spark' model is not supported when using Codex with a ChatGPT account."}' What am I doing incorrect, please advice.

[unreal parcel]

someone said they got it working after /logout and logging back in

[hard tulip]

i am on the pro tier it's just a generic error because they've removed it from the list and i still had it selected previously.

[steel sierra]

wow, way to go for the O.AI team

[lone island]

yeah, same here

[unreal parcel]

vibe coders trashing others for bugs is quite a sight

[steel sierra]

so codex cannot run agentic workflows like CC, hard to find /Skills, /hooks, agentic harness like CC, and now it is limiting access to latest model via CLI, am I understanding this correctly? Has anyone found a way to perform long running agents (+skills ) very much like CC while on Codex? (genuinely asking, as I am either looking to bring my CC workflow to Codex, or cancel the $200 OAI sub)

[high girder]

Honestly though. I just blew through a 5 hour limit in claude code in 8 minutes flat. One prompt. Granted, I'm working on a huge project, but Codex 5.3 in general, has been solid for a lot of the work. I say keep the codex sub, and ditch the claude sub. You can use Claude Code with a local model back end if you have the hardware for it. Also, you can use Claude Code agents to spawn Codex CLI instances. Go nuts.

[unreal parcel]

unable to resume my large conversations now, just gets stuck in Booting MCP server: codex_apps

:/

[hard tulip]

welcome back spark

[frosty zealot]

• Hooks don’t exist (yet)
• Agents use skills as applicable or you can just type $skill-name in your prompt as much as you want
• Not sure what you mean by hard to fine /skills
• The model is removed from the app as well I’m assuming it wasn’t performing as they intended

I had the same questions as you when I started using Codex I was like really? This is it? But if you can, use the app, especially with their native work tree handling, pr/code review features, mcp and skill manager it’s a way better experience, and the 5.3-codex model performs way better than 4.6 imo

4.6 is like an abusive ex girlfriend i keep going back just cause i think maybe it changed and its what im used to but it just hurts me again

and them charging for /fast is irritating, especially since spark seems way faster and pretty much as capable

[amber scaffold]

Does Codex CLI have an equivalent to Claude Code's /usage command for monitoring rate limits?

[frosty zealot]

yeah

/status

took me a second to remember

Its also in the Codex App if you click settings

[amber scaffold]

Yup, just found it! Thanks. I think I ignored several times. I assumed it was an uptime thing...

[frosty zealot]

Yeah, its definitely some weird verbage threw me off too for a bit

Spark is back

[languid perch]

It wont let me edit my messege, so what can I do to get it to keep working? (for context I was asking it to use a block list and I gave it a blocklist for words and that was too much for it)

It wont let me move on the only thing is the retry button

[timid flicker]

is 5.3 better then 5.1 max?

[crisp gust]

love using Codex for local reviews. Would be nice to have local review sessions count toward the code review quota instead of the main quota. My code review quota is always sitting at 100% and feels a bit wasted, while my main quota gets drained.

[steel sierra]

IRL. I gave 10% of my workflow that I built with Claude to Codex, I gave Codex (heavy thinking) a good 3-4 iterations for codex to build and test (TDD etc etc), at the end of 3-4 hours of wrangling. I showed Codex the exact output Claude Code created with the same instructions . After performing a unbiased review Codex - resigned. IT gave me this message.
---- "You’re right to call this out. I’m recording these failures into AGENTS.md as new learned rules now (noise filtering, header context switching, diagnostics separation, speaker canonicalization), then I’ll
give you a clean handoff note you can give Claude Code."

What am I missing guys?

[steel sierra]

IRL. I gave 10% of my workflow that I built with Claude to Codex, I gave Codex (heavy thinking) a good 3-4 iterations for codex to build and test (TDD etc etc), at the end of 3-4 hours of wrangling. I showed Codex the exact output Claude Code created with the same instructions . After performing a unbiased review Codex - resigned. IT gave me this message.
---- "You’re right to call this out. I’m recording these failures into AGENTS.md as new learned rules now (noise filtering, header context switching, diagnostics separation, speaker canonicalization), then I’ll
give you a clean handoff note you can give Claude Code."

What am I missing please

[vagrant flicker]

Anyone know if you can use different models for sub-agents in Codex? Main = 5.3-codex, sub-agents = Spark? Doesn't look like it's supported but wondering if there's a hack for it.

[frosty zealot]

There's a chance it might be secretely rerouting you to GPT-5.2 https://github.com/openai/codex/issues/11189#issuecomment-3880522742

[ivory zodiac]

its not supported yet

but here's what i recommend.

https://github.com/am-will/swarms

Step 1: RTFM

Step 2: Use $swarm-planner with 5.3 codex high/xhigh

Step 3: at the end of plan mode, esc out and "save the plan to file"

Step 4: Switch to Spark

Step 5: run $parallel-task

What you're describing would be preferrable because ideally you have base 5.3 acting as your orchestrator but we dont have much of a choice.

never post your order number anywhere

even tho this is free

bad habit

🙏

[still trellis]

Sorry sensei

[ivory zodiac]

nah ur good bro

just good general advice 🙏

[cedar skiff]

i was feeling the same thing, in order to tak advantage of it you need to roll out a pr work flow.

[still trellis]

Noted 🤭

[ivory zodiac]

people can divert your packages or get your info

probably the former

[still trellis]

Dang crypto rats 🤣

[ivory zodiac]

hahah yeah man

the best scammers are very smooth

'omg i'm so sorry, i put the wrong address'

'my autofilled used my old address'

you hope that the vendor will actually make you give more info but you never know

[timid flicker]

Ran cat > Fe

why is codex using cat to edit files rather than just editing them

[ivory zodiac]

thats codex for you

you can try to use more explicit guidance in your agents to use apply patch

i noticed it typically prefers apply patch for small edits and cat for larger edits

but yeah that is annoying

[cedar skiff]

skills work much better in codex, hooks are in the works, but honestly most of the stuff you use hooks for in claude code are because of short comings codex doesnt have. codex exec is the solution to agents afaik. But most of my work flow stuff in cc was trying to make sure it was doing the right thing. That just goes away with code.

[ivory zodiac]

oh nah there's great use and need for hooks for sure.

all kinds of fun utility

but i do agree that there's a lot of shortcomings in claude you dont have to worry about though

thems be facts

kinda goes back and forth because a lot of ppl want stop hooks so they can tell codex to continue 💀

where thats actually not really a big issue in claude

so kinda it goes back and forth

might need em for different reasons

[cedar skiff]

yeah for sure , I'm just saying lots of the hooks stuff is trying to make sure claude is doing the right thing in CC - there are also legit cases. So many of the usages that are in CC just go away.

[ivory zodiac]

yeah i mean you can use them for pre commit checks, linting, blocking dangerous commands, pre compaction summaries, logging, and hte list goes on and on and on

i'm very excited for hooks

but re: the stop hook thing, i've gotten around that with a simple statement in agents.

if you use subagents you can try all of this.

the 'execution' part though is what i'm talking about.

ever since adding that, i dont get codex stopping ask me if it should continue

it just does

[cedar skiff]

codex has subagents now or are talking about codex exec wrappers?

[ivory zodiac]

nope

subagents

/experimental

[cedar skiff]

oh nice finally, is it in stable or experimental?

[ivory zodiac]

its stable

/experimental is more or less stable

if it was unstable they wont put it in there

unstable = 'in development'

thats when its just a flag, no /exp entry

like memory is rn

[cedar skiff]

well i when i say stable i mean the stable release 🍌

[ivory zodiac]

yeah i mean its stable

its been out for like a month

[cedar skiff]

nice, that was one thing that really held me away for a while

[ivory zodiac]

i use swarms all day every day

[sacred minnow]

I used claude heavily and codex i love codex now since going back to the cli

[ivory zodiac]

i love both but codex is near parity now

not quite but 85%

prob 80 actually

[sacred minnow]

Claude's to quick to jump in and start a new feature not take time to check what's up if you dont specifically give code

[ivory zodiac]

once hooks 90

the one thing codex will be missing, well two. are teams and extensibility/customization

i dont think teams is THAT important, but will be cool

[sacred minnow]

I havnt used cc in a few months but ive done alot with codex and its turned out real nice with the current updates 🙂

[frosty zealot]

I was switching back and forth, but i found 5.3 is slower, but i usually dont have to itterate on it, it just does it right the first time, claude is quick but then im like "move this, reposition this, fix that" and eventually its good

and the 2x limits are nice too

[chrome raven]

codex does more thorough work too

[cedar skiff]

after using claude for probably close to a 2k hours the current state of it vs codex it isnt even close. In my domain anyway.

[ivory zodiac]

codex generally does better work across the board

ui not as much

writing definitely not

i'd like to see codex get better at convex as well

[still trellis]

Yeah codex is a eng and claude is more of a vibe

[ivory zodiac]

they're both amazing

i was referring to codex the cli tool

the models to me there's no question i prefer codex for most tasks

but claude code is the gold standard for a cli

its so good

codex is really closing that gap tho

i'm building something pretty cool to take advantage of both.

its nothing special really its just a skill but

i built a swarm loop system that you plan with and it builds dependency graphs into your plan.

then you switch to code mode and implement with swarms, it basically launches parallel agents in waves based on unblocked tasks

so if there's 5 unblocked tasks, it'll launch 5 agents. if there's only 1, it'll launch only 1 agent

and it works in a loop until all tasks are done.

but claude is so much better at UI design

https://github.com/am-will/swarms/tree/main/skills/co-design

so i made another variant that calls Claude for all UI stuff

[cedar skiff]

you can create a skill for convex with index and reference files and codex will be an expert in the lastest and greatest everytime.

[ivory zodiac]

yeah it works but claude is still just generally better

but yes, that's how you can make it better

but i'd rather it be good regardless because sometimes it might not call the skill

i think they will be better on the next model.

[cedar skiff]

I have to scream at claude in orchestration just to use skills at the bare min level

codex hasnt missed a skill yet

[ivory zodiac]

claude is funny honestly. i have it calling skills without prompting somewhat regularly but i do hear this complaint ALL the time

so i know people arent just making it up

how many skills do you have?

[cedar skiff]

probably about 18 now

[ivory zodiac]

yeah thats not too bad

watch i can almost guarantee it'll call this skill every time. brb

[cedar skiff]

about 10 of them are topical

i can guarantee claude will call skills the first time everytime. But after it called a skill its 50/50 on the next turn that it calls another one. Then after 120k context, 0% chance

[ivory zodiac]

okay that makes sense

i love this about opus, i didnt have to tell it to launch in parallel

it just does.

i do have HIGHLY complex workflows i do in claude

[cedar skiff]

Yeah its way in front on subagents and tooling in that area

[ivory zodiac]

but it usually doesn't exceed 50% of cw

but 1 workflow might call 10 skills, 4 subagents

nested 3 layers deep

but because they're subagents the orchestration agent isnt seeing creep

occasionally it'll screw up the formatting though

but your suspicions are true. codex doesnt experience drift like claude

opus

you can push the cw to max