codex-discussions

[kind jay]

Skill issue tbh, check hugging face fine tunes for specific use cases

[signal tapir]

Oh I don't doubt it's a skill issue. I've tried to find models that can help, but so far I've found none that do tool use very well.

GPT have been uncommonly little help

[kind jay]

I don’t think you can find such small models that are so general, I think you want to look for models for a specific use case

[signal tapir]

You mean like "Unity programming" or, just like "C# programming"?

[cedar skiff]

qwen just released a dense 27b model that does pretty good. But being dense probably still needs some decent hardware

[kind jay]

DeepSeek V4 is out? and I mean programming specific models

[cedar skiff]

i think running a 1.6 trillion param model might need a little more than a mac mini

[kind jay]

There’s smaller versions 😐

[signal tapir]

I'm currently limited to 16GB VRAM

[cedar skiff]

what is the smallest one?

[signal tapir]

Once not long ago, that was respectable. 😛

[cedar skiff]

flash is 284b o.0

[kind jay]

Still maybe cluster is not ideal, would work but would be slow

[dawn seal]

@bright swift I wonder, how did you automate your task?

did you use scheduled automation-feature of Codex?

[bright swift]

what do you mean? i dont have anything automated right now. i only use the cli, not the app. usually a few parallel sessions right now with lots of hook-based enforcements. so tasks usually run 1-3h, sometimes longer.

[dawn seal]

oh so you keep typing into the console ?

I thought you were using some sort of automated agent

[bright swift]

yea i did dictation for a bit but prefer typing

one of the tools im building is for better planning/spec workflows so i can parallelize more

bottleneck right now is not implementation but detailed specs really

defining edge cases etc

[cedar skiff]

You can use an orchestration loop to push through spec writing a bit quicker

[bright swift]

its more about business logic decisions. sure the models can write a spec fast, but i need to make sure that the details are correct

[cedar skiff]

Yeah that's what im talking about

This tactic only really works on brown field though

[signal tapir]

Funny. I was just thinking about getting my fields plowed. 😂

[bright swift]

basically i'm trying to do something similar to what you are suggesting i think. trying to build a process where the agent makes "better default decisions"

[lost drum]

yo can someone tell what those it do? the auto review thingie? like I dont understand cause he reviews the code and edits it and idk what would it change tho

[cedar skiff]

You make a subagent that reads your rough spec, finds ambiguous details or gaps. Then searches the code base and attempts to fill those gaps, if it finds a definitive answer it fills out the spec to fill the gap. If it doesn't it fills out open questions with suggestions on what it found or what it might suggest.
Then you rinse and repeat, but each time the agent starts it checks and tries to answer open questions first.
What happens is you go from a small spec with a rough idea to a large detailed spec quickly.
Each iteration because the spec got filled a little more some of the open question become clearer and can be answered.
You can also insert your self in the loop and have it present open question to you each iteration.

You go from rough outline to detailed spec much faster.

[bright swift]

Yea similar approach. This is good if you already have (decent) code. For new stuff i'm trying to make it fill in the question marks with research from primary sources etc

[signal tapir]

Glad to see people with decent workflows. 🙂

Throughout my career I've met so so many who have no idea what they're doing

Even if they are excellent coders

[bright swift]

This grill-me skill that was/is trending is surprisingly good for nailing down details. Very simple prompt basically. But it can get so detailed that i often find myself annoyed because codex is like "ok, next, question 59: bla blah 3 paragraphs"

[stable obsidian]

Finally got x20

[bright swift]

Most of the time the suggested answer is good, but every now and then i need to clarify sth and that makes the difference

[unique spade]

Hey everyone. So what are the impressions on gpt 5.5?
I had to sleep and I barely gave a couple of test tasks

[signal tapir]

I'm still on 5.4 on the web. Only 5.5 in mobile. :/

[bright swift]

Its less corporate in communication style, havent noticed big behavior changes yet, token usage relatively similar so far

[dawn seal]

haven't tried 5.5

the image gen however is pretty good

I generated some virtual wifu and husbandos

[unique spade]

Imagegen yea is great

[dawn seal]

it can do increased more realistic romantic scenes

its very good at work stuff too

[signal tapir]

You can get pretty simple models to do that though

[unique spade]

Was curious how did some A/B in codex though for 5.5

[nocturne folio]

how do i get 200$ asap

[stable obsidian]

money

[nocturne folio]

ran out of usage on the plus plan

[dawn seal]

go to your chat gpt account on webpage and upgrade it to 200

[bright swift]

do you have a casino nearby?

[signal tapir]

Getting money is easy. Getting money ethically is harder.

[dawn seal]

I disgree, getting money not-ethically is hard too

[signal tapir]

I'm broke, but can think of a thousand unethical ways to make money without breaking a sweat.

[dawn seal]

if it were easy, I would some not-ethical money

[signal tapir]

lol

[dawn seal]

the arguments such as "you kill someone then get their money" is not applicable because you risk breaking the society's rules and being sent to jail

[cedar skiff]

welcome to the light side

[signal tapir]

that's part of the whole unethical thing. It carries risks. shrugs

[dawn seal]

the financial translation of the consequence is negative then

you lost money in the end

[signal tapir]

I don't stick to ethical stuff due to risks though. I like to help people, not hurt them.

[bright swift]

i used to do that, now i'm flexible

[dawn seal]

I thought we are talking about how easy can you make money unethically

my argument is that, punishment for breaking law is financially negative, therefore it is hard to make money even unethically

[signal tapir]

I guess I didn't see it from that angle. I just avoid it anyway.

[dawn seal]

yeah, what I wanted to say was, making money unethically is hard too

[cedar skiff]

unethical != breaking the law

[dawn seal]

so therefore, if you consider the negative financial outcome of unethical method, your previous statements on making money by breaking law should be reconsidered

[signal tapir]

He's right though: The cost of breaking laws does make unethical moneymaking harder. Not necessarily hard, but harder.

I'm glad most AI don't help with unethical stuff easily.

[nocturne folio]

i think i miss gpt4.1

[dawn seal]

often following a law is a physically beneficial choice

[bright swift]

Thats why prompt engineering was invented hehe

[dawn seal]

breaking a law is not necessarily prevented by ethical feelings but more of calculated outcomes

[signal tapir]

I bet Codex is used mostly for legal unethical stuff actually.

[dawn seal]

this is necessary because we have psychopaths in our society and they simply want to benefit without regard of others, this type of people are pretty common (by that I mean, they are more common than you might think)

[unique spade]

Good vibe ethics chat :)))

[signal tapir]

Like attention-farming on fb

[pastel fog]

Is it possible to use 400k context in Codex app with 5.5? It shows 256k for me

[unique spade]

The definition of what is ethical and not can vary a lot from individual to individual, culture to culture and so on.

Law has a much clear definition in comparison and is in effect for it s jurisdiction

[signal tapir]

Venn diagrams and all that

[unique spade]

You can use up to 1 mln context since 5.4

You have to add it in settings. Cost is also different above 256k.

[unique spade]

Could say that, even though python also starts with venn diagrams and then "all the rest" 🙂

[rustic walrus]

wait u can?? hows the performance?

[signal tapir]

How do you mean?

[dawn seal]

yeah for some, crossing red-light on road is ok

[signal tapir]

That's in the intersection of illegal and unethical though

[bright swift]

laws are suggestions

[stable obsidian]

Guys where do I find the browser use plugin? Cant use it somehow

[dawn seal]

what I was trying to say is, for some crossing red-light is an ethical decision

it is ok to cross redlight for some

[pastel fog]

I don’t see it in Codex app. I was using other harness, and I think it had 400k. But I’m used to 1M, don’t know how to enable it

[pastel fog]

Ask Codex to open a page

[stable obsidian]

👍

[unique spade]

That basic logics, of which venn diagrams is like part of 101, are the base of all advanced cognitive stuff, from programming to ethical modeling

[unique spade]

Ask codex, you can add the setting in the toml, tell it to give you also the link to where is documented on oai web, including price

[signal tapir]

I see. I don't have any official schooling, so I didn't know it was part of that. I just saw it as "ethics and law somewhat overlapping"

[unique spade]

They overlap quite a lot. The purpose of law in decent society is to explicitate and enforce the ethical norms of the community

[bright swift]

thats a very idealistic take

[unique spade]

That s why i said purpose, doesn't mean is 100% like that in practice

[bright swift]

yea ideally they would do that, in practice in the modern world i'd say laws are mostly in the interest of elites, not any community or citizens

[unique spade]

All this goes all the way back to Plato, and older with the questions wether something is right cause law says so, or law is good wether it captures what is ethically right

[signal tapir]

My wife is a prison guard. She tells me it is almost universal that the prisoners feels like society is unfair towards them.

[bright swift]

which is why i say laws are only suggestions

criminal law is an exception

i'm talking mostly about business world

[unique spade]

How do you mean suggestions?
Law have causal consequences. They might be badly made but they still hurt when you fall under their hard lane

[bright swift]

well first of all, just because its a law it doesnt mean its correct. laws get overturned or changed all the time in court

[unique spade]

And that applies also for civil and commercial law, not just criminal, since you can end up having to pay a lot

[bright swift]

many country level laws violate international law

etc

[unique spade]

International law is not full law yet. On that we agree, it s still mostly recommendation

As global society we don't have yet global mechanisms that actually enforce downstream

[bright swift]

why should i as an entrepreneur follow every single tiny commercial law, when the state refuses to do its part of the social contract and not prosecute violent criminals, illegal immegration, corruption etc

[dawn seal]

there is always reason to not follow law X for reason Y

following a law is an option

you simply face a conseuquence or not

[unique spade]

Yup

And obviously you can assess the probability of facing those consequences too

That's what good lawyers do, that's how they find loopholes

Under-specified cases in law etc

[bright swift]

And the really expensive lawyers even use AI to invent cases now haha

[unique spade]

And then if the judge finds out it s risky cause he might get annoyed

[bright swift]

Pretty sure for lawyers it can have more serious consequences to have AI hallucinations in their works then for us vibe coding folks 😄

[unique spade]

Actually that's the usecase from which I started my current project, full formalization using hybrid modern systems of legal corpora

So you have a system that can find contradictions, ambiguities in existing laws

[signal tapir]

That is super interesting! Do you publish details about the process?

[unique spade]

Once I get there. The project is public but I'm still in building the pipeline to actually be able to do that in an actual sound manner

[wicked briar]

codex app again sucks

back to cli

[signal tapir]

Can you PM me a link?

[unique spade]

Repo is linked in my discord profile

[signal tapir]

So basically it finds and enumerates ambiguity?

[unique spade]

Eventually it would map all the inferential dependencies

Legal is a semi-formal NL domain

[signal tapir]

That sounds crazy dificult. I am totally following this one. 🙂

NL? Neurolinguistic?

[unique spade]

Natural language

[signal tapir]

I see.

Before now I considered legal a very strict domain.

But thinking about it, of course it isn't.

[unique spade]

Normally NL is under specified relative to concepts that are express, as in between terms and concepts the relation is loose

[unique spade]

Yes it s most strict of NL

But still semi formal relative to truly formal domains like math

[signal tapir]

One would think it should be as formal as math.

[unique spade]

But there are specialized logics to process normative bodies, the name of the family is Deontic logics

[bright swift]

that would defeat the purpose as you wouldnt have any more loopholes

[unique spade]

Exactly, and that's one of my long pending about 20 years projects

But before current ai systems, it was practically impossible to do it

Because of the scale of it

And like you would had to hire armies of humans to do the translation

And humans are error prone anyway

[signal tapir]

I'm reading up on Deontic logics as we speak.

[unique spade]

Legacy formal systems are interesting in themeseleves, because there s a lot of them and there were always debates and paradoxes. Like the larger community never settled upon a canonical one

But yes legal and scientific methodology are formalizable with modern tools

Take academic artefacts, like papers, ideally they should all follow the methodology so their results are machine checkable

Not just be a bunch of natural text

And rhetoric heavy

[signal tapir]

Deontic logics sounds... messy

Like what if you have opposing obligations?

Don't answer that though. Still reading. 😛

[unique spade]

The legacy ones are like from the 60-70's

They were over optimistic

Then they hit their head into the same questions like the one you just put 🙂

And started to make refined models

😂

[signal tapir]

If/else trees?

[unique spade]

Well all logic is only if/else trees. But for a formal system you have to go upward at 2nd level and prove a model about how you make the if/else trees

Then you hit Godel issue

[signal tapir]

A depends on B depends on A?

[unique spade]

Which proved that above 1st order systems there are always some statements that are true, but can't be proved

[unique spade]

Godel is a result mostly about semantics outrunning syntax of you want

You can have things that are semantically true, yet you can't prove them syntactically at that level

So you end up going one level higher, make a richer syntax to prove that truth

And this is an open ended ladder

[signal tapir]

That feels like it can go on for a bit

[unique spade]

Yes it's open

Sky has no upper ceiling from down here

[signal tapir]

Has there ever been attempts at creating a non ambiguous language?

[unique spade]

Programming languages are probably the most successful attempt to do that 🙂

Compilers don't accept ambiguity well

Lol

[bright swift]

i'm tired boss

[signal tapir]

huh. interesting point. 🙂

[unique spade]

How many questions will it give you? Till you say stop? 😂

[bright swift]

idk 😄 usually it stops when everything is resolved but this is a heavy session 😄

[signal tapir]

I asked it to ask me questions until it had enough information to make a good decision. It didn't stop. I believe the attention system makes it keep on forever.

[unique spade]

Yes because as opposed to math or logic, in programming language you also specify a base ontology

And if the base ontology is clear, then it s easy to map all the combinations among those privitime types

[forest crypt]

So far my experience with 5.5 medium and high seems to indicate that the model handles complex design more aptly. At least on the scope of my project regarding finance and in the 500k lines of code, the design and analysis sessions I carry out with Codex are more fruitful. Before my feeling is that it tended to get lost in small implementation minutia and get lost. I had to undo and redirect in all sort of ways the effort. Now it gives me the impression of being able to handle more complexity and keep architectural details more clearly. It's all subjective at the moment and I have no KPIs to measure it yet

[signal tapir]

Harder and harder, the less machine close the language is

[unique spade]

Right

[exotic cave]

Thanks for the info

[signal tapir]

Thanks for the info.

[rancid plover]

Hi! Everyone?

[unique spade]

Nice assessment!

[unique spade]

Hi

[rancid plover]

How's going on?

[unique spade]

Now think of IR but transpose it between the messy untyped natural language humans use and a typed language under which is mapped to clarify meaning.

You get the concept of Semantic IR

[signal tapir]

IR?

[unique spade]

You want your role bounded agents to work on tasks after they ve been normalized through the Semantic IR layer

[signal tapir]

Ah.

[unique spade]

You want the same thing between the human messy prompt and the final semantic compiler step the agent does

[rancid plover]

Imposter found

Gemini Agent Found

[signal tapir]

I feel inadequate for the intellectual requriements of this subject, but I still find it immensly interesting

[unique spade]

Right now all those steps are fused. The agent does all together. It has to disambiguate your prompt, then implement the task

[bright swift]

5.5 seems to be a lot less patient with subagents. its now spying on them via git logs in subagent worktrees like "what they doing ova there"

[unique spade]

Separations like between plan step and implementation you can think of it as early steps in this direction

[unique spade]

Give him "patient orchestrator" directive

[signal tapir]

Sounds sensible. Keep track on them so they don't spiral.

[bright swift]

orchestrator already has "let them cook" instructions

[unique spade]

It s default orchestrator style is very top-down master-slave vision

[unique spade]

That s quite loose, letting someone cook with wanting it to be done in 5 seconds are still coherent

🙂

To me it feels agents really need a better grasp of time

[signal tapir]

I wonder how useful it would be to have an ai that constantly keeps track of what you are saying, and feeds back to you the ambiguity of it

[unique spade]

Not metaphysical but actual time of how it takes for some processes

[signal tapir]

Do they have ANY?

[bright swift]

i think it would have negative effects, like claude code got worse when it became aware of its context limits and status

[bright swift]

They added that in codex app already, with the guardian mode or whatever its called

Or autoreview or whatever

[unique spade]

Claude has different training. Personally I consider the path they went a dead end for my needs

[bright swift]

Claude has anxiety trained into it as prime objective it seems

[unique spade]

Yup

Which is very unproductive for depth

But very productive to fake you shallow success

[bright swift]

Yea became unusable for me, i liked it for non-code general purpose stuff before

[unique spade]

From my tests if you tell codex a 100 step sequence he must do

He will do it

If you give it to Claude, it browse it, jumps on it s own to the conclusion and writes some forgery of how it followed all 100 steps

Even though the logs show it never did it

And if you push pack, it says oh sorry

😂

And does a bit more on 2nd attempt.

[bright swift]

apparently with the newest versions it says stuff like "this is too much work for one context window" and doesnt even start, but i dont know if those were just ragebait screenshots

[fierce steppe]

Are Chinese here?

[unique spade]

It s plausible, we know they are compute constrained

So they have to limit the actual inference budget

[fierce steppe]

I use third API? How use GPT-5.5? I can't see GPT-5.5 in CodeX model list

[bright swift]

Update app/cli

Its not on API yet

[unique spade]

Yea

[fierce steppe]

OK, I'm trying

[unique spade]

Not on api yet I think

[fierce steppe]

Other have used gpt5.5 in codex with third api

[hard drum]

me with CGPT Pro 20x sub using it natively:

[fierce steppe]

Oh I have updated codex cli, Can't find gpt-5.5 in codex model list

[unique spade]

I use it on codex app/cli

No experience with 3rd party api

What third api do you use?

[fierce steppe]

Cool, I'm a pauper, can't buy pro

[echo knot]

I've seen several references on X and youtube that the intended baseline reasoning with 5.5 is medium. Can anyone verify that coding output with 5.5 medium will still be better than 5.4 at high?

[unique spade]

Just do your own A/B

Give them the same task in your codebase

[fierce steppe]

sensoft.top , this third api

[echo knot]

a question is much more token efficient...

[bright swift]

I only use xhigh, it does dumb stuff on lower ones sometimes

[unique spade]

Yea but it s not about your actual project

[signal tapir]

And less likely to give useful data.

[unique spade]

You will get hallucinations from other users relative to what works best for you

[fierce steppe]

Where are you come from? I'm first in here.

[unique spade]

I mean it s cool to exchange feedback

But you want your own baseline first

The first thing I do when a new model is released is to A/B it

[bright swift]

Start with medium, go up if it makes mistakes or struggles with your project

[unique spade]

And analyze the differences

[unique spade]

Agree with that

[signal tapir]

I mostly use 5.4-mini, and go up when I need to. Which is rare.

[echo knot]

yeah, I have always defaulted to high, just using xhigh for repo refactors or checks. 5.5 has been super solid on high, but I've already burned 20% of my 200 plan since 5.5 dropped, so trying to get the hive mind consensus re: dropping back to medium.

[unique spade]

For writing concrete code I assume?

[bright swift]

Just buy more pro plans

[unique spade]

If you already have used high, do a test drop to medium

And see if you feel any difference

For example I did an open ended test on the broad repo architecture

Between high and xhigh

High defaulted to a more bounded operational interpretation of my task

Xhigh took it as I told it

Like high had the task to analyze 60 modules and it just skipped first 30 cause they were history

Xhigh analyzed all 60

But that s on high level open ended task across a big repo

[bright swift]

It depends on what you're working on too. Some webapp, medium probably fine. I work on a custom ml pipeline in rust, so i go with xhigh

[unique spade]

Yup. It really is dependent on that the most

The model itself is just half of the equation

[echo knot]

thx for feedback!

[frail meadow]

Just give this to Codex and it won't

[unique spade]

After you test medium and compare with high, please let us know of you impressions 🙂

[echo knot]

just fired off a medium prompt! (given how solid 5.4 was for me, if 5.5 medium can compete with 5.4 high, I will be happy!)

[unique spade]

Btw i think that gptpro 5.4 on steroids was really 5.5

Cause yday after they released 5.5 I gave gpt pro 5.4 a review task, and it got back to taking 1 hour, whereas the steroids version finished in max 10 mins

I ll have to test 5.5 but I hope it sticks to the 10 mins I had for the past days on 5.4

[forest crypt]

On the usage: I am only using 5.5 High on analysis, design and architectural discussions. Found discussions very insightful and results useful for me. As far as the cost I don't see huge drops in my count. But for sure I would not use at the moment 5.5 for coding. It seems to me that 5.4 medium was doing just fine as long as I tell it to batch operations properly

[bright swift]

i mean some people still swear by 5.3-codex for implementation

[unique spade]

On specific pre 5.5 debugging tasks, 5.3 was still better than 5.4 for me.
Like 5.4 kept going in circles why a panel it made doesn't actually show up

And 5.3 solved the problem at first attempt.

[echo knot]

My issue with 5.4 medium was that the model would seem to stop sooner than it should have with a multi-step ask even when I had a very specific scaffold for the work.

[bright swift]

Yea i saw that too with lower reasoning mode

[unique spade]

Maybe it s made to protect qouta for 20$ peeps 🙂

[bright swift]

xhigh just works for hours and hours for me

[unique spade]

You want long term stuff use xhigh

[bright swift]

and does everything i ask for

[unique spade]

I feel is how it should be tbh

If you want multi hour long tasks that s obviously an "XLong task"

But would be nice to have a different slider for time horizon separate from reasoning effort

So you can choose between shallow but long horizon
And deep but long horizon

[inland sonnet]

i also had issues with that with 5.5 lol, only thing it had to do is split functions

Other model did it instantly, gpt 5.5 wouldnt

[unique spade]

Interesting

[terse kraken]

When is the limit reset coming?

[unique spade]

5.4 and 5.5 being more of a poly-reasoners they might need a different approach

5.3 codex is specialized on coding

So that might buy it a big advantage in focusing on the proper frame

[vital hill]

5.3 codex medium is the sweet spot

[unique spade]

For actually coding implementation it might as well be

[cedar parcel]

So far I can say 5.5 agents are definitely much better at orchestrating workers during long running tasks. It finally learned what patience is.

[vital hill]

right. it's great at refactoring and one shotting features

[unique spade]

For architecturing how intent should be made in code, I always prefer 5.4 and probably now 5.5

But once the implementation spec is ready, codex 5.3 might be the sweet spot

[signal tapir]

Do you guys get decent architecture from specs? I need to do some manual preparation first to get it working properly.

[echo knot]

I've had good luck delegating sub-agent tasks to leverage that seperate codex 5.3 sparc token bucket

[steady vigil]

I think they lie about the deepseek v4 cost like most chinese models. I just tried it briefly in opencode and its already used $2.95... in less then 1h. yes its a wide task, but this would be less then 1% of my gpt weekly

or their caching sucks

[unique spade]

I do the first high level pass myself with chatgpt thinking coming up with objections. And once I like that. I usually tell gpt to craft a prompt for gptpro to add depth. I do usually a refinement or 2 in the same way. Then I move those to support docs and switch to codex where I use 3 layers of concretization, the last one being the actual implementation spec for a specific feat

[unique spade]

That s a lot 3$ per hour. What s their supposed cost for 1 mln tokens?

Would be interesting to test the same task and how much tokens gpt 5.5 uses on it

[steady vigil]

trying to find it

but it was around 10% of gpt 5.4 or something

[unique spade]

It s ok don't bother. I can search later

I assume it aligns with the 8 million it used for you

[steady vigil]

its weird b/c I found also k2.5 was super expensive. not sure if its a problem using with opencode

maybe its super inefficient

[unique spade]

So it should be like 40 cents for 1 million

[hard drum]

save me...

[unique spade]

Their cot sucks

[unique spade]

I am testing them with simple but tricky logical puzzles and the cot of the Chinese models is still vanilla

And it runs for ages

[steady vigil]

have you used anything other then opencode?

[unique spade]

Token cost in itself does not tell much if most of those tokens are junk

[unique spade]

I did not even bother trying them on api, they fail at the in vitro logical test :=)

[signal tapir]

That is very similar to my approach.

[steady vigil]

they dont have any coding plans do they? its just api right?
here it is: https://api-docs.deepseek.com/quick_start/pricing/

[unique spade]

I m just saying that what you say makes sense to me. If they need 20 times more tokens than gpt for same task, even if on paper they are 10 times less expensive per token, you pay twice for same real task

[unique spade]

Think so

[steady vigil]

yeah that's my point

[unique spade]

Yea but if you payed 3$ for something that would have taken 1% of your weekly qouta with codex....

🙂

[tawny turret]

are u guys complaining abt token cost

[steady vigil]

they must have been comparing api to api not against gpt coding plan also

that would make more sense

[unique spade]

Nah

[tawny turret]

if u dont make use of like normal chat and video generation, u can get 2x the amount of codex (in return for those other features)

[steady vigil]

so its still way way way more epensive then gpt coding plan for less capability... only makes sense if you need api

[unique spade]

Yea that's why I m saying testing the same task in codex via api would show exactly the token ratio used for same fixed task

Or can it be counted via sub plan too?

[boreal holly]

Those models are probably good for tasks like this though

[unique spade]

I think there might be some tools

[tawny turret]

loll

[steady vigil]

quality

[plain crescent]

dont leak the sauce

[steady vigil]

@boreal holly dont forget the agentic qa pass on the result!

[boreal holly]

Adversarial capitalization review 😭

[unique spade]

That s some new offer or what?

[hard drum]

you mean coding in 2021? copilot came to be in 2022 around masses

[steady vigil]

prob better to use a team of agents and use consensus review

[unique spade]

How you get 2x codex if you don't use normal chatgpt chat?

@tawny turret the above is for you

[boreal holly]

Don't forget xhigh /fast --yolo, because you want the capitalization to be as close to 100% accurate as possible but not take too much time

[hard drum]

don't forget "dangerously skip permissions"

[bright swift]

Through shady resellers

[tawny turret]

uh idk if i can share it yet apparently theyre still in the "testing" early access phase, im lucky to be a tester but basically itrs a company that offers twice the amount of codex limits for the same price as codex plans, but no access to regular chatgpt web stuff

[tawny turret]

resell? how are they reselling it ? in a loss? they said they have a partnership

[tawny turret]

wrong* they allow u to buy codex only, 50% off the regular chatgpt plans but u get codex access only*

useful for ppl like me who only use ai for coding and skidding

[boreal holly]

"I need to fan out subagents, one to check for capitalized Kanji, one for Cyrillic, another for Greek"

[bright swift]

I assume they use the web interface for distilling and sell the codex usage to soften the financial cost

[hard drum]

One for runes, as well.

[bright swift]

But idk

[tawny turret]

no its 100% codex ive been using it since 2 days ago and I have gpt pro from my company aswell

[torpid trout]

Sometimes I do not understand LLM
It often infers answers based on almost a decade years old specs
Like, ask it this:
Is it allowed to wrap dt in a dl with div?
It will consistently tell you no - ask it the same with "in 2026" appended, which will trigger a websearch, and of course the answer will be yes - its been allowed since basically 10 years now

Asking it why - it says "lots of training data still refers to the old spec"
😐

[bronze sable]

can you send me the link? would be quite useful for me I only use claude for chat

[torpid trout]

I guess it is a conundrum - do not train in old data, it cant deal with old data... but at the cost of losing new data? Sounds like a bad move to me

[bright swift]

https://openai.com/index/codex-flexible-pricing-for-teams/

I guess there are codex only plans now, didnt know 😄

[tawny turret]

tester only rn, but Ill send u a dm once they launch public, probably tomiorrow

[bright swift]

I guess they are exploiting the signup bonuses with multiple accounts etc "To support your adoption, eligible ChatGPT Business workspaces can receive $100 in credits for each new Codex-only team member that joins and starts using Codex, up to $500 per team, for a limited time. To activate the offer, add Codex-only seats to your workspace or create a new ChatGPT Business workspace⁠(opens in a new window)1."

[tawny turret]

no lol thats something entirely different, thats api usage, like 20x more expensive than what i have XD

[unique spade]

does sound shady to me, and probably against oai terms

i mean sure we all know there is grey digital market

[tawny turret]

yeah lol but they said they haev a partnership, tbh I hope its all legit, would be very nice to not pay for voice mode and roleplay webchat when i dont use it

[bright swift]

My guess its an exploit of the API credits i just posted

[tawny turret]

lol that would be crazy

[bright swift]

But maybe there are some unreleased partnership for this type of stuff, who knows

[unique spade]

or they resell legit codex usage to which they have access to

[bright swift]

Why would they do that?

Codex usage is the most valuable part of the plans

[tawny turret]

nah they wouldnt sell in a loss would they? i think legit partnership

[unique spade]

there are loads of gpt users that don t use codex. many companies that have enterprise plans but trust the admin to manage who uses

lol

[bright swift]

Yea possible, but still that wouldnt be an official partnership to resell pretty sure. And definitely not enterprises handing out login info

[unique spade]

but for me it's 100% clear. there are no official codex offers than those offered directly by OAI

*better

i mean

it s just basic economic logic

[boreal holly]

Just saw some activity on the issue I posted:https://github.com/openai/codex/issues/18130#issue-4275895971
If this gets enough 👍, it would allow conversations to be cached for up to 24 hours instead of 10-15 mins for no additional cost. Just saying this is a pretty good feature to have

[unique spade]

oai won t partner some 3rd party to giver you 2x cheaper codex

[bright swift]

You can also buy "residential VPNs", what they dont tell you is that those are just malware infections 😄

[unique spade]

thumbed up it

[tawny turret]

uh those are called residential IP's and they are as legit as they get lol. many big official companies sell them. theres also car dealerships that scam ppl, doesnt mean its all a scam

[bright swift]

Just because a company is big and official doesnt mean they dont do shady stuff 😄

Nobody will sell you a codex subscription cheaper than OpenAI

[boreal holly]

One thing I noticed, with 5.4 I would frequently see "that giant patch failed because something changed, I'm gonna do multiple small, targeted patches instead." 5.5 does not have this issue. Instead it'll run nl -ba on a file beforehand and craft a perfect patch. Fascinating to see it meticulously avoid issues like that

[fair jasper]

hi, is it possible to config browser use to be able to open other local ips?

For some reason my doesnt allow any other than localhost

[ivory crane]

5.5 Codex arived?

[signal tapir]

Now 5.5 arrived for me. 🙂

[bright swift]

No, go back to sleep

[ivory crane]

ohhhhhhhhhh TT

[haughty plaza]

hello everyone, is there any promotion ongoing for a 1 month trial?

[ivory crane]

bruh

[signal tapir]

5.5 feels about the same, or possibly a bit faster than 5.4.

[toxic glen]

Hey folks, to save on usage should I plan with higher reasoning (xh) and implement with lower (med) or the other way around?
I'm testing this but not sure yet which one is working better

[signal tapir]

I would do the former. Best results are if you already know architecture, and only need to guide the AI to write documents for an architecture you have decided on.

[boreal holly]

Plan with Medium, execute with Medium. Maybe execute with high if it truly struggles with something, but otherwise start with medium.

[orchid plume]

I can't keep up with it 😂. Well, ev3 in the model name is experimental variant number 3. but I don't have a clue what "flx" means. Flexible is the only thing which came to mind, possibly trained for long running agentic tasks or something

looks like we're getting a codex model variant soon though

[ivory crane]

[lean lark]

You can get an architect to design your house and a high school kid to build it.
Or you can get a high school kid to design it and a certified contractor to build it.
Or just ask the kid to figure it all out and let you know when it's ready to move in.
Or you can work with the architect and contractor.
Your choice.
It's actually not that cut-n-dried. Quite often the task doesn't require an architect or a contractor. Pounding a nail doesn't require a contractor - ask the kid to do it. Drawing a picture of the house doesn't require an architect. Ask the kid to do it.
Use the right tool for the job. ™

So JaneBot did this for me yesterday. What else should we add to make it useful?

[unique spade]

to be fair since these upgrades are incremental, much of what was already good should ideally remain the same, whereas the real improvement should come in edge cases

[lean lark]

☝️ That seems to be the consensus in all comments. 5.5 isn't revolutionary, it's evolutionary.

[blissful copper]

‪I found a site that would help you pass your evals faster by giving you a algo design to pass evals, it gives you a significantly higher chance of passing evals and a great way to hedge on cheap prop firms like apex-https://propfirmpinescripts.com/‬

[signal tapir]

Yea, that makes sense

[solemn acorn]

it's more token-efficient than 5.4, so it's probably running at about the same speed but completing tasks faster

pretty cool considering it does come with a price increase per-token

[terse kraken]

Is this the first time we dont get a reset with a model release?

[hard drum]

🙁 Oh well...

[signal tapir]

I don't get the feeling it's completing faster, but looking at my usage stats it drains the usage slower

[inland sonnet]

I just noticed gpt modifies prompts lol

let me give a stupid example

[bright swift]

only 25% of my weekly pro 200 used in about 6h yay

[inland sonnet]

if u ask to make a white man black, it will not add the word "black" in the prompt

[solemn acorn]

you can also more safely lower the reasoning level to low/medium which is neat

[torpid trout]

I can't believe what 5.5 just did
It tilted the freaking image, on its own, no ask for whatever crazy.

[signal tapir]

When I got the "want to try 5.5?" question, it set it up for highest settings possible, including fast.

You know, so it will eat my usage as quickly as possible. 😛

[unique spade]

interesting. based on the leaderboard solely for this benchmark

in terms of performance gpt 5.4 xhigh = gpt 5.5 medium

which combined with that cost table you posted, means what costed 2851$ now costs 1199$ which is quite big

also you get 2 extra points for 960 extra $ with gpt 5.5 high
and to squeeze a 3rd extra point it costs another 1200$ for 5.5 xhigh

[signal tapir]

I wish I had enough business to warrant that sort of extravaganze 😁

[unique spade]

well assumption is the ratios remain similar on lower tasks for peasants like us

[lean lark]

Side note: Last warning from GitHub:

On April 24 we'll start using GitHub Copilot interaction data for AI model training unless you opt out. Review this update and manage your preferences in your GitHub account settings.

[unique spade]

who knows. maybe it will be codex 5.6

5.4 skipped codex
maybe 5.5 skips dedicated codex model too

[solemn acorn]

it's unlikely, openai unified codex with 5.4, they're unlikely to go back on that now

[unique spade]

what s that?

is it on this official discord?

or?

is it just me or would be nice if on this official oai discord channel we could trust info relayed at least about oai models is actually legit?

[lean lark]

This is just another public forum, it's not really "official".

[lost drum]

do you think it will be a thing?

if yes then bro

I cant imagine

[unique spade]

public refers to target audience

who are the creators of this forum?

i mean is created by 2 random dudes in a basement? or is it actually made by oai?

the space itself i mean

[lean lark]

The closest thing to "official" is this: https://community.openai.com/

[boreal holly]

OpenAI is the creator. There's a verified checkmark here indicating its legitimacy

[lean lark]

This server is staffed mostly by volunteers but there are many company staff here on their own time and they do occasionally view and rarely post here.

[unique spade]

ok then that is "official" enough for me

it's a community space about oai products created by oai

[lean lark]

Well, it was created by the company and they help to keep it a well-managed community resource, and of course they do publish announcements, pretty much anything else here is purely public commentary.

[wicked oxide]

Hey, is anyone else seeing high resource usage from the Codex desktop app on Mac?

For me, Codex is using around 3 GB memory which is fine , but CPU is the issue. Activity Monitor shows total Codex CPU around 290-300%, with two Codex Helper Renderer processes each going over 100% CPU. Energy Impact is also around 1,100-1,300.

Memory pressure is normal and swap is 0, so it doesn’t look like a RAM issue.

Can other Mac users share their CPU, memory, and Energy Impact numbers, and whether the app is also making your Mac heat up?
though i am on macbook pro m4 pro .

[unique spade]

yea bro but there are hygiene rules to keep the community healthy, like you can't start using foul language

my point was that it would be nice to have higher hygiene

[plucky halo]

How would that be enforced?

[lean lark]

For that, send a note to the Mod bot.
But they aren't gonna suppress community scuttlebut.

[unique spade]

for example, people coming around and shadily promoting dubious offers of some "3rd parties" that give you 2x codex and other stuff like this

i don t think it should be allowed

[lean lark]

We want a healthy environment, not nannies.

[unique spade]

you are WE? 🙂

[boreal holly]

Commit what work you have, quit the app, reopen and resume that thread. See if it return to normal

[plucky halo]

I'm happy for @lean lark to include me in that we

[lean lark]

"I" am "WE" .... community.

[unique spade]

we are all part of that WE that s the point

we are having a discussion inside that WE

lol

[lean lark]

For this healthy community we have:

• Bots
• Mods
• Guides
• Eachother

Would you like more?

Bots scrub for words and images. Outside of that kind of deterministic processing of data, we can only rely on one another ... this is a public resource, we need to pick up our own trash.

[unique spade]

that's a long discussion probably. i was just expressing something that i think would be nice to have, on a very specific topic, like spam/scam related to Oai products

[plucky halo]

Yes, but to be able to shut down community hype around models / misinformation around models it would require a moderation team whose sole purpose/job is to monitor each channel. It's nearly impossible. Let's not forget that our moderators are community volunteers

[lean lark]

I believe we would all agree that keeping the place clean is important but frankly I don't see anything here that needs cleaning up outside of peeps posting wants/desires and misinformation that the rest of us need to counter.

Yeah, staff here aren't paid.

(Former Guide here, been through the process)

[boreal holly]

I have a solution to the hygiene

It's not automatic but gets the job done

[unique spade]

we are in AI age though... 🙂 there might be other solutions already haha

[unique spade]

yea that's epistemic due diligence

[lean lark]

There are actually very few official channels, just keep an eye on those. Anything that is not coming from the home page or a validated X account is subject to scrutiny.

( Hey JaneBot, keep an eye on official channels and let me know when OpenAI makes product announcements. ) 🙂

[boreal holly]

GPT 55 is probably a decade away anyways, let alone GPT 55 codex. That's reaching wayyyy into the future 🤪

[unique spade]

i saw at least 3 occasions with shady promotions about some 3rd party that gets you cheaper codex and to dm about the link

personally i think this shouldn t be floating around here

but i guess you guys are fine with this kind of scams

or consider there s nothing that can be done about it

[signal tapir]

Scams not being handled by platforms is one of the worst parts of this modern age.

[lean lark]

WTH?!!? I just worked with Codex to modify code, did a commit and push, and in the GH repo it says I and Copilot committed.

[unique spade]

100%

[unique spade]

hmm, that's new and bad haha

[boreal holly]

If they committed and pushed with gh instead of git that might do that

[lean lark]

So you're saying you read something on the internet and it wasn't true? I'm kinda doubting you bro....

[unique spade]

anyone worked long enough with 5.5 so far to get a feel of how it compares in qouta usage for same types of tasks vs 5.4?

[signal tapir]

Slightly lower I would say

Although... I just reached the dreaded "You've hit your usage limit. To get more access now, send a request to your admin or try again at 10:19 PM."

[unique spade]

lol let's get back at chill between us. we always end up at odds with eachother haha. let's just close that topic. 🙂

[lost drum]

imagine 5.5 pro heavy

[lean lark]

Please, I'd suggest that the metric is quality vs cost. We're looking at that Triangle again. If you only look at tokens and quotas you're not focused on the value of the model. Though I understand that this is all a part of that cost/value determination process.

@unique spade that was just a funny comment, absolutely not a jab, please assume funny first and rarely jab, especially amongst friends here. Yeah, in plain text we do not get tone. 🙁

[unique spade]

to build that triangle you still want each edge. was just curious as qouta use for alike tasks vs 5.4 and 5.5

[lean lark]

Oh, funny or stupid, more likely stupid... just sayin....

[bright swift]

oai people still vagueposting on X so probably something else coming today

wen reset

[unique spade]

what else did they post

[bright swift]

just these little vague one liners

https://x.com/ajambrosino/status/2047707785597800579

[lean lark]

that's such a tease

[torpid trout]

I thought I was tripping at first...

[lean lark]

Oh, and because CoPilot shows as co-authoring my repo, now commits are Unverified. Ugh, need to address this!

[unique spade]

what s this

[unborn thunder]

codex usage is a bit out of sync

[torpid trout]

results from testing 5.5 on 5*5 html blueprints
Notice the cards on the graph.. they move.
The prompt for these jots is something as simple as "create 5 different business personas and for each 5 different landing pages using distinct styles and layouts"

Compared to 5.4 and prior, this is truly pleasantly surprising everytime I look at a new draft it made.

It comes up with craze, and I love craze.

[unique spade]

nice

[lean lark]

kewl

[wicked briar]

bruh thats nuts

like I mean not crazy but from what earlier models used to do is good

[torpid trout]

What is really nuts, but I am sure this is a system error, is that these 5*5 html dummies, with business personas, consumed 1 percent of my weekly pro sub (200 bucks one)

[boreal holly]

Oh I know! It's not just rounded rectangle cards inside cards with plain white backgrounds and default typography

[torpid trout]

Exactly, and when it does that (sometimes it still does), you tell it... and it actually changes things lol
And that is even more pleasing. I feel understood. I feel listened to. My wife should take notes.

[solemn acorn]

it does seem to have a much better understanding of 2D space based on what I’ve seen

[torpid trout]

What it still cant do at all is manage subagents
But I gave up on that lol. Who needs it, anyway, it mainly is cool to have but does not add much better results, rather a perhaps bit cooler experience... fine, I will not use them

(these 5*5 html pages are like the best ever use case for agents, and it plain out did not use a single one)

[nocturne folio]

im going to cancel my claude sub for a second codex plus sub

which is 60$ cad sadly

but idc

I havent yet used deepseek yet

seems super promissing

[bright swift]

what problems are you seeing with subagents? i'm currently experimenting with the multiagentv2 feature flag, seems to make things better

[boreal holly]

Someone mentioned earlier this morning setting up the orchestrator to "let em cook." 5.5 is really good at this! I feel like the official subagents feature, the main agent can see what happens between full turns. In any case I think if an orchestrator is set up only to judge the end result it keeps the role simple enough they never drift.

[bright swift]

that was me 😄

i give each subagent a fully set up isolated worktree though

and some wrapper commands for the orchestrator to manage adopting commits

[boreal holly]

Me too! I have a hook that creates the worktree and rewrites their CWD to it, and when they get archived it reaps their worktree & container stack. This is the way 😁

[bright swift]

the issue with the multi agents v1 is that subagents cant respond when the orchestrator sends a question

but with v2 this seems to work better

and orchestrator can get impatient if you dont tell him specifically to chill

[torpid trout]

UH, I had seen that a while ago in releases and they said this is not released (even if it was released, but whatever)
So how to use this?

[boreal holly]

That was my impression looking at the official subagents. When the orchestrator is stateless, they send out messages and end their turn, which I think is a stress reliever in its own rite

[bright swift]

yea exactly thats the issue, it asks for status update and the subagent either doesnt respond which pisses the orchestrator off, or the subagent accidentally ends his turn to respond

[bright swift]

[features]
multi_agent_v2 = true

[torpid trout]

So replace this?
[features]
multi_agent = true
or keep and add?

[bright swift]

with v2 the subagents can send a non-final response back

uhm not sure probably both at same time will give config error

actually both yes

one important thing i added in my fork is to pass cwd to the subagent, so he starts in his worktree instead of the main checkout. important for my hooks setup

thumb up welcome 😄 https://github.com/openai/codex/issues/18969

[simple star]

Wait, im lost. Have agents changes?

[bright swift]

?

[simple star]

multi_agent is not valid anymore?

[bright swift]

multi_agent enables the collaboration/subagent tools at all. multi_agent_v2 selects the v2 version of those tools. If multi_agent is off, v2 will not matter because the collab tools are not exposed.

typical oai naming things

[tiny fulcrum]

anybody else has problems with codex app, it's slow and constantly freezes and I have to wait till it unfreezes....

[simple star]

Strange, I have been using codex for 2 days, and I do not see traces of agents anymore in its usage, even though features.multi_agent is true

[unique spade]

ideally you want the orchestrator to oversee the process, i agree
and to actively interfere little in the actual work of sub-agents

[bright swift]

It only uses subagents if you explicitly tell it to

Its not like claude code where its proactive

[simple star]

@bright swift Before it wasnt so, no?

[bright swift]

I think it was always like this

[unique spade]

i saw it very very rarely to use sub agents by it s own

[simple star]

Not true. I have been seeing agents for weeks, even though I never said to use them

[bright swift]

You probably had some reference to it in a skill or agents.md

[lean lark]

I've seen many people reporting many different experiences with agent spawning, just like we're seeing here.

[boreal holly]

Exactly! I set it up so workers cannot run git commands (except commit and request review), so only the orchestrator can approve code changes into integration branch. That's the gate where orchestrator can oversee things in my setup

[unique spade]

it makes sense, i think the practice of every person matters a lot

[lean lark]

Don't you have an agent for GH ops? 🙂

[unique spade]

i mean if you have it use sub agents often i d imagine it becomes some kind of implicit policy inside the memory of the thread

[unique spade]

nice

i wonder if 5.5 suffers from 5.4 orchestrator bug when it used inherit parent context to 5.4 sub agents

[simple star]

@bright swift Doubtful, I have changed nothing

[unique spade]

it was so funny to see how sub-agents thought themselves they are the orcehstrator and tried to mirror the orchesdtrator instructions and spawn themselves worker agents

[boreal holly]

Yes! I think if an agent's only job is to manage subagents they tend to do so, but in the way default codex is set up, the base instructions are like "you are a general purpose, all encompassing agent designed to do whatever", and most folks are gonna be like "make a plan, refine the plan, write the code". But when you want em to use subagents, they don't see any task as so complex they need to fan out so they default to doing it themselves. But when their entire purpose is to use subagents and they can't write code it works out better

[bright swift]

Well they can.. There is agents.max_depth setting 😉

[simple star]

Are the default agents gone? Do I need to create them manually now?

[unique spade]

it s a specific bug that is documented in issues that affects only 5.4

[bright swift]

Ah ok

[simple star]

The configuration of Codex is such a unholy mess

[unique spade]

5.4 sub agent received prompt to do x, but if it got inherited context , it also saw the prompt for the main agent and somehow in half of instances that prompt overpowered the one it received. so it tried to do the orchestrator task instead

😂

[lean lark]

Codex usage/quota reporting sub-thread: https://discord.com/channels/974519864045756446/1497255535664300173

[bright swift]

Read only permissioned orchestrator is an interesting idea but I use it to resolve possible merge conflicts because i have multiple orchestrators running in parallel

In different checkouts though

[undone patio]

me after switching from opus to codex

[boreal holly]

My orchestrator does have workspace_write! The trick I employed is I made a zsh wrapper, tied it into the role identity of the running agent so when they run any command I know what role is running it and can react accordingly.

So they can't write into worktrees or the main repo folder, but I have certain commands escaping the sandbox depending on the role. Rather than direct access to git, I have a bunch of sanctioned scripts that do non-destructive ops only, and only the orchestrator can use most of em. It's pretty complicated, but it's all posted up in https://github.com/robertmsale/.codex

[bright swift]

Ah interesting. I have hooks that block writes outside the checkout/worktree but not per role

[lean lark]

Except for the animation, that reminds me of old wall-sized ASCII-art. Anyone remember Monalisa or StarTrek posters created on LA36 or similar paper terminals?

[unique spade]

i tweaked my orchestrator so it can peek at intermediate reasoning outputs of his sub agents

so it stops wondering what they're doing mid task 😂

[undone patio]

its made using a gif to ascii pipeline i coded

[bright swift]

Via raw session logs?

[unique spade]

patched codex itself and introduced this capability in my fork lol

[bright swift]

Ah nice

[lean lark]

Gawd, the time it took me to create stuff like that by hand in BASIC....

[rocky fog]

its the age of "forking your own" 😄

[unique spade]

well it's AI age, it's supposed to be like this

[undone patio]

https://latentfidelity.github.io/gif2ascii/
https://github.com/latentfidelity/gif2ascii

feel free

[unique spade]

anyone here still uses claude or other agents too?

and if yes for what type of tasks

[undone patio]

only if my codex runs out

[torpid trout]

Now make it like codex let it accept input while subagents work?

[unique spade]

so like back up emergency junior

[undone patio]

😭

im praying sam finds a reason to reset limits

[lean lark]

☝️ Tobi

[lost drum]

where would you use this type of Ascii animations?

[undone patio]

idk, i just like them personally

[unique spade]

will look into it. right now i got side tracked to make my own UX surface to use codex

[undone patio]

maybe as a game asset or something

like for example you can make cool 3D effects with it

[bright swift]

Seriously all this meta work is so addicting lol

[torpid trout]

Thats the thing I loved on claude/copilot cli

[lost drum]

makes sense I thouthg how could I use it in ecom but no clue haha but ye for games or loading to hacker menus coool

[torpid trout]

And truly OAI should take a look.

[torpid trout]

having fun with image gen?

😄

[bright swift]

They did with the codex app didnt they?

[lost drum]

Tripy

[torpid trout]

Maybe, I am not using that app, it doesnt really work (after a while starts crashing)

[bright swift]

Yea I uninstalled it too, felt like it was draining usage in cli faster somehow

[unique spade]

Meta is the edge. It s where you get to compete straight with the giants. So it makes sense it's addictive

[lean lark]

Decades ago I did a TUI/ASCII animation: A Romulan warbird came down from the top of the screen over business data 😆 ... then the Enterprise came up behind it, shot photon torpedos, blew up the warbird, and then exited off to the side of the screen ... then the business data was restored.
Who said green screen was boring?

Now Codex can do that in a minute.

[undone patio]

last one:

[lean lark]

I'm in love

[undone patio]

its a fun time

[unique spade]

Oki can you put it in one concise paragraph? What you d like the current harness to support and it doesn't?

[lean lark]

Weirdly enough that's about the quality of animation available at the time StarWars first came out.

[unique spade]

This rocks

[undone patio]

feel free to star the repo 😎

open source forever

[lean lark]

( Until Intel decides to kill open source )
(Sorry, just reflecting a bit of news...)

[unique spade]

Missed those news... What s Intel planning

[torpid trout]

So in codex CLI when you have a thread open and it (the model) is working, the main input still lets you submit new messages, but it queues them. Imagine the model would actually use subagents reliably (unfortunately most of the time it does not) and, when the main thread is not busy revieweing something or delegating tasks to the sub agents, the main agent becomes immediately available for interaction.

Basically, you can then:

• kick off task via prompt to main agent
• main agent does its thing (delegates to subagents)
• as it is done and waits for sub agents, you can already interact with main agent again, potentially having a cofeechat or do more serious work, while it waits for subagents to return

When I was working via CopIlot CLI using Claude models exactly that was possible, and claude told me it is an explicit feature it has inbuilt, to free up the main thread as much as possible for parallel interaction

[lean lark]

They've been publishing FOSS for a couple years, seriously deep stuff related to chips, but they just terminated all of that as an undesirable expense. Kinda of a kick in the pants to corporate support of FOSS ... especially for their own products.

[slender plinth]

guys. im sick of this stupid claude. 2 questions on opus 4.7
claude limit reached.
switching to codex

[lean lark]

https://tenor.com/view/die-hard-john-mc-clane-welcome-to-the-party-pal-glass-break-gif-15466663

[unique spade]

Thx

[steady vigil]

the 5h limits are terrible on max x5

I just hit that

gpt/codex is a dream in all ways compare to everything else

[slender plinth]

how many prompts?

i wanna get the plus plan

[lean lark]

ChatGPT Plus plan is a great place to start.

[stable obsidian]

Youll prompt it twice and itll be on limit, its more than claude still but still close to nothing

[torpid trout]

I might say, the plus is the plan to start

[torpid trout]

That is just not true
Unless you prompt it to analyse fqacebook's source code or so, you get hours to work with plus
I know because I have both, and I use these products daily, unless I am not.

[slender plinth]

bruv so it aint good as all people say? that codex has more usage

[bright swift]

it has more usage

but $20 is hobby plan unless you are very careful with your prompts

[torpid trout]

Ignore vague "trust me its the internet" statements and just try it first.
It has significant usage. Whether it is more or less than claude I cannot say, but it is very generous even now that limits are not 2x anymore.

[stable obsidian]

Had been running business plan for a while, has the same limits as plus and I had to switch like every 40mins max

[torpid trout]

trust me, ok? 🤣

[slender plinth]

Okay

option 1, keep claude and suffer
option 2, go to gpt and maybe also suffer i dont know yet

both 20 dollar plan

[torpid trout]

sounds good no?

Unknown suffering is better than sure suffering.

[stable obsidian]

Youll suffer less with gpt

claude is ridiculous

[gentle harbor]

i thought they were gonna reset rate limits

[slender plinth]

okay cuz i broke my keuyboard today of rage

[gentle harbor]

i wasted all my tokens since i thought they would

what will i do now

[stable obsidian]

couldve bought gpt pro with that money 🐧

[torpid trout]

Generate a meme with the last 2% please.

[slender plinth]

nah it was a 2 dolalr keyboard sicne i like smashing things

[lean lark]

Your decision shouldn't be based on "more usage" but what you're doing with the usage. Don't run xhigh or fast, use high-quality prompts, document your project/code. These factors drastically reduce time and token usage, and increase the quality of your product.

[slender plinth]

okay so run it on effor low? or medium

[torpid trout]

For rare I agree with captain

[stable obsidian]

Consider using the caveman skill aswell

For more usage

[slender plinth]

looks real right

its gpt image

lol

[gentle harbor]

no lol just look at the keys and the background, its very close to being real was this with the api or with normal chatgpt ?

[slender plinth]

normal chat

the key in the front is a bit weird

[lean lark]

My suggestions about which settings to use: #codex-discussions message
#codex-discussions message

[gentle harbor]

the reason i say no is since all the switches are dif colours and even the keycaps are dif colors on the under part and have a bunch of weird text

[unique spade]

I consider fast only if reset gonna come too soon and I have too much qouta left

Otherwise I see no reason trade qouta for speed 🙂

[slender plinth]

for everyone

im 14

[gentle harbor]

???

???

[slender plinth]

so if yall think whats the for joke dude

https://tenor.com/view/cat-thumbs-up-thumbs-up-cat-thumbs-up-meme-gif-5676038700371371386

[slender plinth]

yea that

[lean lark]

I understand that being 14 you want to throw as much as you can at the prompt and get as much back as you can. As professionals, we need to use the technology better to get better results. Follow what I said, learn how to get the most out of the tech. Beyond that there's not much to offer.

[bright swift]

just use 5.5 xhigh fast

[slender plinth]

LMAO

[bright swift]

or ngmi

permanent underclass

[unique spade]

😂

[lean lark]

get high fast used to be the way to go...

[slender plinth]

rip my usage then

[gentle harbor]

i use 5.5 high on low speed

[bright swift]

real Gs use 5.4-pro via API anyway

or 5.5 soon

[unique spade]

Just manage your usage. It doesn't matter what age you have. It is a basic resource optimization problem

[lean lark]

( Standard speed, there is no Low )
( Do or do not, there is no try )

[bright swift]

just think about it like this.. pro is 200/mo.. thats what, 10-15 hours of fast food work? and you can have agents running 24/7 on xhigh fast with a couple of fastfood jobs + pro plans

[main nimbus]

still is

[unique spade]

24/7 with xhigh fast i honestly doubt it gets you there

Not based on my usage.

[main nimbus]

you should validate the concept for us. Gl

[unique spade]

Lol

[main nimbus]

it hurts every time i see our API costs after using 5.4-pro

[unique spade]

Well would be nice if you can only do those 10-15 hours so you can chill the rest of the month coding with codex.

But I don't know if it works that way

[bright swift]

well you need a couple of pro plans for that especially i running things in parallel with subagents 🙂

[unique spade]

From June it will be 4 pro plans

Since currently it's promotional 2x 🙂

[bright swift]

they will extend.. believe 😉

[unique spade]

Let s hope they do. They still have best offer by far for power users

[lean lark]

As an example of my common usage, I just generated code on Medium, and I toggled to Low and said "fix lint issues". You don't need high reasoning to fix linting issues. (In this utility I just added ESLint which wasn't present for the original code, otherwise the assistant runs lint before and after code gen and corrects its own mistakes before completion.)

[bright swift]

they should just let us add multiple pro plans in one account instead of credits

switching accounts is annoying

[unique spade]

Yup

And then make a top with who wields most pro plans

😂

[main nimbus]

Why would they make it easier to lose money?

[bright swift]

there is a guy on X who claims he uses 24x claude max plans and uses them up

[gentle harbor]

there is a guy on X who claims he can fly

[bright swift]

in this case i believe him because he puts out crazy amounts of code

[unique spade]

We re still in R&D phase

Much of that usage goes into improving models

Personally i make all my codex available for training, because I want my usage to be part of new training for obvious reasons

[main nimbus]

Actually, yes. But not all users are sharing

[lean lark]

Save on subs: At some point, run local LLMs for little stuff, or use a per-minute/hour VPS with huge GPU for big stuff.

[bright swift]

so youre saying i shouldnt ask gpt 5.5 pro extended thinking for a weather report? blasphemy

[main nimbus]

I have a DGX spark and m5 max both 128GB and can’t find reliable enough local models to use consistently. Probably requires fine tuning currently for any production case.

Definitely not code.

[torpid trout]

you should ask it to use chatpgtp via computer use, to prompt there and give you the answer back 🙂

[bright swift]

did you try gemma 4?

[lean lark]

I'm kinda impressed with Gemma 4...

Heck, I have Gemma 4 on my Android.

[gentle harbor]

its not making me stop :)

[main nimbus]

The data quality is impressive but I’ve seen bad tool call performance

For gemma

[bright swift]

Same as gemini then huh

[main nimbus]

Typical of Google, yes

[gentle harbor]

ive been on 0% for 15 min now and i can still prompt and talk to it

[lean lark]

I'm Not a fan of any Google products but for some purposes I lower my standards.

[bright swift]

Seems like google is making a big push now with their "strike team" announcement couple days ago

[main nimbus]

I’ve read it’s a problem that can be solved- mostly related to their own formats not being consistently supported across current inferencing

[lean lark]

Ignore Google Marketing buzzwords ... they're a daily annoyance.

[unique spade]

quick question i started recently a fresh repo, in codex settings it appears in the review list, but somehow i don t get either auto reviews from codex, nor the manual @codex review works

maybe i need to go through an extra step like updating how codex connects to that specific org/repo?

[main nimbus]

OpenAI has consistently had the strongest models for every use case I have had (real world)

Except design but wbk

[bright swift]

probably need to add it in github connector permission list too

if you set it up with "only some repos"

[unique spade]

yea was thinking about it, have to remember where to do that lol

[undone patio]

dear sam please get drunk and reset limits again

[bright swift]

https://github.com/settings/installations

[lost drum]

I thouhtg its codex 5.5 notificatiin

[unique spade]

thx! you re faster than gpt

[bright swift]

wen reset

[unique spade]

🙂

[quasi summit]

Are we finding 5.5 faster to use up rate limits?

[bright swift]

about the same i feel

[undone patio]

its faster

ive never hit limits so quickly

[lost drum]

It burns faster but 10x better output in my project

[undone patio]

im a sad developer now, gonna have to use inferior opus

[quasi summit]

We have spark for this use case

[potent mason]

Surprisingly not really

Just keep it in medium

[quasi summit]

Scary. I need to earn more within a month so I can afford three pro subs

[undone patio]

never, i need all the juice

[lost drum]

3????

Damn

[undone patio]

i have gpt pro, google pro, and claude max

[quasi summit]

One at 2x is barely enough

3 is $1000 CAD tho

[potent mason]

Bruh Opus uses up limits at least 10x faster

[undone patio]

yeah but

[unique spade]

thx bro! on my main i had all repos set, but in the specific org i started that repo it was set only for individual repos

[undone patio]

its all i got left

lol

besides gemini

and gemini is kind of derpy

[lost drum]

Ye I have 1 and its ideal for me but if u have more things to build then 3 I guess

[bright swift]

the tokens must flow!

[quasi summit]

I’m fulltime indie working 7 days a week

[potent mason]

My problem isn't even with programming, it's with the openclaws they freaking devour it

[lean lark]

I would think that OpenAI would adjust API tiers such that the natural path is to go with the API when quota is exceeded. That eliminates the multi-account hassle. Of course that's where Business and Enterprise pricing come into play as well.

[undone patio]

you dont need a claw with codex automations

[quasi summit]

There are tools for load balancing multiple accounts automatically

API will never be economical

[lean lark]

Speaking of OpenClaw and them ... the 'llm' utility can interface with 5.5 via API, even though 5.5 API hasn't been announced. That gimmick will be obsolete in days but it was a nice feat for 'llm'.

[potent mason]

You do know openclaw is not only automations right?

[mighty lion]

why 258k context max on codex? using gpt 5.5

[quasi summit]

Biz pay insane overages just to have enterprise management of org accounts

[undone patio]

idk, i wrote my own MCP toolkit

[bright swift]

codex autocompact beats bigger context imo

[strong hemlock]

I'm getting Error running remote compact task: timeout waiting for child process to exit errors on my linux vm when i running out of context. Any way to like give it a longer timeout? It works fine on my Windows machine

[undone patio]

https://latentfidelity.github.io/par/

[lean lark]

I think they just haven't updated the FOSS for that. You can set 400000 or above in the .toml.

[mighty lion]

i need for large database

[mighty lion]

even 400000 didnt work

[lean lark]

That's weird.

[mighty lion]

model = "gpt-5.5"
model_reasoning_effort = "high"
model_context_window = 300000
model_auto_compact_token_limit = 8000000
personality = "friendly"
sandbox_mode = "danger-full-access"

[potent mason]

I can't wait until someone comes up with a "token economy" just like the crypto bs

[mighty lion]

see my config

ill fix the compact

[spiral trellis]

anyone know if openai ever said if they r gonna add gpt pro model to codex?

[bright swift]

dont give them ideas

[potent mason]

I don't think so

[lean lark]

That looks like it should work @mighty lion

Are you running codex-cli 0.124.0?

[mighty lion]

codex app

[potent mason]

Besides it would devour usage, it's like 6x more expensive than 5.5, 12x than 5.4, and probably uses like 10x more tokens (basing it off how long it loves to think)

So in reality it's probably like 120x more expensive than 5.4

[lean lark]

I dunno about app details, sorry.

[spiral trellis]

yeah but its so good , i'd love to have that as an option

[bright swift]

https://github.com/steipete/oracle/

works great for reviews or planning, i use it with repomix to build a combined file of additional context

[spiral trellis]

oh I saw one of those like 6m ago or so but if its from pete then I'll def give it a try

thanks!

[lean lark]

HAHAHA - Funny about Oracle, I just did that this morning.

[frail meadow]

Spotted you in the wild on a YouTube comment lol, I was wondering what your linting setup looks like when working with codex?

I was thinking about just starting with the recommended typescript eslint rules

[lean lark]

Which comment?
I have the ESLint extension in all profiles. I npm install eslint and related packages for all projects. I have an extensive eslint.config.mjs. My package.json has a "lint" script. For Codex, AGENTS.md directs the assistant to run lint and save output to a temp file "lint-1.txt", and at the end it runs the same to "lint-2.txt". Then it compares. It must fix whatever is different and it must not address whatever was wrong before it started.
I think that's it.

[frail meadow]

It was from a video from Web Dev Simplified about something called Biome.

[bright swift]

why the two files? i just lint in post tool use hook and tell it to fix asap or it gets replaced

[lean lark]

Oh yeah, Kyle's channel. Awesome stuff!

The two files are a before/after. I don't want the current task to be burdened with fixing unrelated issues. I only want it to fix what it broke. Other issues are fixed and committed separately.

[frail meadow]

Sounds to me that this setup is useful for brownfield projects where existing stuff is known and accepted

[bright swift]

the reason i lint in post tool use is that i dont want linting issues to accumulate until pre-commit which runs the big QA gate

[rocky dragon]

Does anyone know roughly when GPT 5.5 will be available in Codex?

[bright swift]

yesterday

[bronze venture]

I feel like limits run out alot faster than before

Even with GPT 5.4/5.3

[bright swift]

lol rip github copilot.. 7.5x modifier for gpt-5.5 same as opus 4.7 https://github.blog/changelog/2026-04-24-gpt-5-5-is-generally-available-for-github-copilot/

instead of 1x for 5.4

[young locust]

check the fast setting as it changed its default for some people

[bright swift]

they enabled fast mode by default in 0.124.0

sneaky bois

[bronze venture]

its not enabled on my end

at the beginning of April, i could get half of the day with a single 5 hour limit

(right after x2 event ended)

but now the 5 hour limit dont stand 1 hour

[fervent quail]

i signed into codex but now it sign ed off now its asking for number and when i enter i receive no sms
is there a way to solve it,

guys how to sign in so i wont have to verify via sms

[bronze venture]

I was using GPT 5.4/5.3 medium

[lean lark]

Comments on Codex reporting are welcome. 🙏 https://discord.com/channels/974519864045756446/1497255535664300173
Breaking for lunch now.

[bright swift]

if you use codex app or even have it installed but use cli, its sneaks in a bunch of extra stuff into your requests via the app server integration. big mcp tools like github, other apps, skills, etc.

if you use cli and care about usage, uninstall the slapp

[bronze venture]

I believe this has to do with openai switching to token-based calculation?

[fervent quail]

is it asking for sms verification in codex appp for u guys

[kind jay]

Giving a piece of software a persona is crazy work tbh

[kind jay]

To stop multiple free accounts bypassing usage limits

[bronze venture]

Damn a single message used 2% (no tool usage, short answer)

[unique spade]

Only for enterprise accounts

[unique spade]

I use Auth, I have it on since codex cloud time when it was mandatory there I think

[blissful copper]

‪I found a site that would help you pass your evals faster by giving you a algo design to pass evals, it gives you a significantly higher chance of passing evals and a great way to hedge on cheap prop firms like apex-https://propfirmpinescripts.com

[unique spade]

I haven t noticed any change in how fast qouta runs this month

[mighty lion]

holyshit codex do use a lot of usage

to point i feel my pc stutter lot when playing

and using codex

even tho i got a high end pc

whatever it is doing in background testing the game or idfk

[cinder plume]

Am I the only one who thinks token usage is lower than 5.4 when using 5.5?

I am able to crank out significantly more work when using 5.5 than I did with 5.4

[unique spade]

No. I have a similar impression so far

[cinder plume]

I was capping my 5hr limit on 5.4/high in 2-3 prompts all last week and prior, now on 5.,5 Im rocking xhigh and I am at least 8 prompts deep in this 5h session

[mighty lion]

ofc?

[unique spade]

But according to model card 5.5 is supposed to use less tokens I think

[mighty lion]

check benchmark

it use way less tokens

[cinder plume]

5.5 seems infinitely more efficient

[unique spade]

Infinitely is a bit of an overstatement 😂

[pastel pagoda]

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

When will GPT-5.5 be available in the Codex Desktop app?
The new update enabled it with this whole thing saying try GPT-5.5 then I try it and get an error.

[cinder plume]

Perhaps a bit of hyperbole lol but still, it seems much more effecient

[cinder plume]

Make sure you are on latest version from windows store

[unique spade]

It works fine for most. Are you sure you updated your app to latest version?

[pastel pagoda]

I'm on the latest version, yes. Previous versions didn't even have the option of GPT-5.5.
I just updated like 5 minutes ago, after I had just updated an hour before that.

[unique spade]

What version do you have

[cinder plume]

0.124 in CLI?

[pastel pagoda]

I'm on the desktop app.

"Version 26.422.30944 (2080)"

[unique spade]

You don't need cli, desktop app comes bundled with it s own cli binary

[cinder plume]

Ah, didnt realize they were technically separate

or however you would define it

[unique spade]

Yes same for ide extension it uses it s bundled binary

[cinder plume]

Weird Hudson, im on a technically older version of the app and have 5.5

[fervent quail]

thank god bro i fixed it

it was pain

[fervent quail]

i already have an access

[lean lark]

For those who are concerned about token usage, ongoing mantra: Look at your AGENTS.md and other directives. What are you asking the bot to do? What is it doing because you haven't provided guidance? That's a serious biggie! Look at the CoT and see what it's doing, then help it to do better. THAT will save you tokens on every turn!!

[unique spade]

Odd mine is working fine on an earlier version. I just noticed a new update button

I m on 26.422.20832

I ll update and tell you if it s still working, maybe it s an issue with that version who knows

[pastel pagoda]

Hmmm...

Well anyways I hope whatever the problem is gets fixed soon because I'd like to try out 5.5 in Codex.

[pastel pagoda]

There are however some bugs I don't want them to fix.

[unique spade]

Well it s working fine since yday for me

[unique spade]

What bugs

[pastel pagoda]

I'm not going to mention them or they might fix them.

[unique spade]

well maybe your issue comes bundled with those bugs that are features from your perspective

[pastel pagoda]

🤷‍♂️

[pastel pagoda]

Oh hi I saw your tag.
So you're a fellow AI researcher?
What did you think of the new CSA in Deepseek V4?

[unique spade]

i have NDA on ai research discussions

[frail meadow]

would you guys say that gpt 5.5 is generally more token efficient? Besides the cost?

I feel like token efficiency helps with generation speed to completion

[pastel pagoda]

...
I was asking about your thoughts on an attention architecture in a public paper.

[frail meadow]

I haven't looked into it yet, what are your thoughts so far?

[gentle harbor]

when are they reseting codex limits for the new model launch ?

[frosty zealot]

Why did you dump your limits prior assuming there would be a reset?

[pastel pagoda]

Deepseek V4 CSA is pretty cool actually I think. I'm planning to try it for some future runs.

[gentle harbor]

since i thought there would be a reset since sama said there would be

[unique spade]

let us know of your impressions once you do

[unique spade]

when did he say that?

[unique spade]

they already did 2 like 2 days apart.

[bright swift]

feel the agi

[frail meadow]

you design transformers?

[frosty zealot]

This is how I dream

[frail meadow]

the baby should have been eating the burger

[pastel pagoda]

Yes.

Do you?

[bright swift]

try it out yourselves 😄 https://chatgpt.com/share/69ebc6ea-5b6c-83eb-b720-e81cb247eb15

[frail meadow]

no but that's pretty sweet dude

[lean lark]

Concerned about your tokens? Look at my cached input. While the "cost" is less for cache, it's still consuming a huge percentage of the overall transaction volume. Occasional forced compression might help. Anyone know if we can force context compression in the middle of a turn?

[frail meadow]

there was someone that recommended that earlier in the chat history

[cyan gyro]

connection unstable for someone else?

[ripe wasp]

yeah mine wont connect

[frail meadow]

I think someone said that there was a compact command in the codex cli but unavailable on the app?

[bright swift]

did i hear reset?

[pastel pagoda]

If it works out good I may make an smaller open version (without proprietary IP) for release on HF.

[cyan gyro]

well yes if I am not the only one experiencing this hehe

[lean lark]

yeah, we were talking about this yesterday or so, there's no "/compact" in the VSCode extension. Will try CLI.

[bright swift]

They recently added /compact

[lean lark]

oh yeah, missed it. just hit it.

BAM! Yeah, you can see when I hit it. Need to think through the significance... Anyone?

[torpid trout]

literally https://chatgpt.com/share/69ebcb1f-5668-83e9-bf46-c6a01176188d

[turbid axle]

praise god bro

[lean lark]

Between screenshot 1 and 2, input tokens and output tokens dropped, because I compacted between turns. It wasn't running at the moment I compacted.

There's something to this. The trade off is that if you compact context then of course you lose some of the details that might be required to complete a task effectively. From what I see here, I don't think most peeps care. They just post one turn after another and prior context doesn't really matter.

In the turn after compaction, cache hit rate went up because the hits on cache were more relevant.

[frail meadow]

So are /compact and the automatic compaction after hitting the context window doing different things under the hood?

[ripe wasp]

same thing here in CLI

[lean lark]

I don't think so, Trippy

[cyan gyro]

yes same

reset coming

[cobalt junco]

man

i dont want reset

i barely used my weeklly usage

[ripe wasp]

wdym reset?

[cyan gyro]

server is not responding consistently

[frosty zealot]

I think /compact lets you also feed a prompt to guide in the compacting and auto just does its thing

[turbid axle]

kick it

dont forget that the new memory system also eat tokens

[torpid trout]

threats always work https://chatgpt.com/share/69ebcb1f-5668-83e9-bf46-c6a01176188d

I think I shuld have used 5.5 for it, since the kind of "I cannot do this" is typical for 5.4 lol

[ripe wasp]

go back to gpt 5.4 it seems to work for now

[frosty zealot]

This is good, they'll say 'oops' and hit the RESETTI SPAGHETTI BUTTON

[boreal holly]

KNEES WEAK ARMS RESETTI

[kind jay]

How do I get some good codex set up that you guys are on?

Like agents and config?

I don’t want that normie codex

[torpid trout]

reset wen

[torpid trout]

I am currently using the OAI vetted "superpowers"
But honestly? No agent, no skill, just pure gpt and a tiny winey agents.md telling it to use ascii does just fine

Its literally IMO mostly not required to have agents or skills unless you ahve very specific needs (like for example, I jave a skill that is capable to understand 200-pages-client-specs and creates actionable json-to-do lists off it, whether those specs are written in handnotes, images or simple workflow diagrams). Now that requires a skill due to the validation of what it extracts, mostly.

But otherwise? I feel more and more the whole skills/agents stuff just actually makes the model less free.

[frosty zealot]

are you being sincere right now

[kind jay]

Yeah, I haven’t really used codex before

[frosty zealot]

What have you been using?

[torpid trout]

This chat, for satire farming

[kind jay]

My brain

[frosty zealot]

That would not have gotten you this far

[lean lark]

Remember "CLI".

[hard drum]

How are you 24% in reserve???

HOW???

[kind jay]

I made one joke and you're still think it's so funny

[bright swift]

🙁

[kind jay]

😎

[hard drum]

holy deficit

[bright swift]

i bet on reset with 5.5 release fml

[lean lark]

Hoping at some point someone will comment on the Codex usage utility thread. I'm done with that one for the day. Continuing on real projects.

[kind jay]

Here 5.5?

[cobalt junco]

idk

ii dont use it too often