codex-discussions

[cedar skiff]

5.4 summarises everything and creates semantic diffusion when it does. It loses detail fidelity by making this its default behaviour.

[cunning thicket]

possible to use hermes within codex just like openclaw?

[opal meteor]

Does a Go account have the same usage limits as a Plus account for Codex?

I can't find this information anywhere

[deft gyro]

go account = free account

[opal meteor]

@deft gyro what does that mean? I can still use Codex on Go

[deft gyro]

so can free users(for sum weird reason)

[opal meteor]

yeah but that doesn't answer my question

if I get the Plus subscription do I get more weekly usage or is it the same?

[potent mason]

You do, a bunch more

[opal meteor]

@potent mason alright so currently I have go at 86%, is it better to upgrade to Plus now or exhaust this Go and use another account to get a fresh Plus later?

also do you happen to know if the 2x is still up? (for Plus users)

[potent mason]

You'll probably get a tiny discount / prorated refund if you go plus on your current account so I would just go with plus on your current account if possible

[potent mason]

It's not

[opal meteor]

It's asking me for full price so I don't think I'm getting a discount lol

[supple perch]

these rate limits are nuts

[lime inlet]

I am struggling to understand how usage limits work exactly. /status shows me a percentage. How can I know at which rate, which model consumes tokens? For example does gpt-5.2-codex consume less than gpt-5.4? By how much? Is it possible to see how many tokens I get / week and at which rates models consume them? Anything to make it easier to understand how fast I burn tokens would be helpful.

Forgot to mention: I use codex (cli) with a ChatGPT Business (formerly: Teams) account

[thorny cloud]

which models do you guys usallyu use

[lime inlet]

gpt-5.4 (medium) for everything

[white furnace]

is it still draining usage like crazy or was it fixed

bc i cant see limit reset

[next ore]

its possible we are just seeing the effects of not having 2x

and maybe we are not used to it..

[cedar skiff]

seems fine to me

[next ore]

wouldnt mind a reset tho

[shell coral]

helllo, what are the estimation of the usage limit difference between Go and Plus plan? the docs only mention Plus plan

[modest oyster]

finally, a filter! a real game-changer for users like me juggling multiple repos at once

[cedar skiff]

man 5.4 is just horrible o.0

it summarises when it shouldnt, it add ideas that arent asked for, it hallucinates much more than the previous models. It just takes so much more effort to work with

[fast heart]

The reason you are not getting any discount is that in GO plan the codex is not included. It's accessible only temporary, beta phase like

[wicked briar]

in codex?

or like web

[cedar skiff]

in codex

[wicked briar]

hmm what model do you use then

[cedar skiff]

5.3 or 5.2, here is an example. I talked through a large refactor, all the stuff taht needs to happen where and how etc. All me giving it the information and asked it to write the problem/solution statement, then asked it investigate the code base and write a task list. This task list i read and correct ande then i move forward with planning and implmentation. Well it put each task in its own file. I said please put these all in the same file. It summarises all of the tasks and writes them into a file. So it lost half of the semantic meaning that it had just spent time writing.

So i have to fork back and then re prompt it so its correct.

just stuff like that constantly happens with it.

[wicked briar]

ohh yeah that happens

model is way oversmart so I put high only

[orchid plume]

same, I never use xhigh 😂

[wicked briar]

they did not bring $100 plan still

they fumbled

[orchid plume]

yeah I had to renew $200 again 🙁

[wicked briar]

well i mean is fine if its really all rounder model but it can't do front end so have to get something for that

[orchid plume]

true, I'm hoping when the new image model is generally available that maybe it can be used to give inspiration to GPT-5.4 in the meantime

[wicked briar]

maybe

[high heron]

can someone that jumped from plus or business/team seats to pro if 5.4-high/xhigh runs faster (normal,not fast mode) or better? are there real differences?

[stiff knoll]

It runs better, but not faster

[high heron]

it runs better, meaning?

[small violet]

yeah fym better

😭

[livid forum]

No cussing fool

[warm pilot]

from my gut feeling it isn't faster in normal mode.

[turbid axle]

yo what up with tokens for real

cant even finish a single 5.4 high prompt before my 5hr plus runs out now

it just stops mid task, and be like, bro you out.

lol, this is ridiculous

[small violet]

ts never happens

[turbid axle]

well, it does now. yesterday I had the same, though it managed to finish the prompt, today I try again, and it won't even finish, just stops and says you out again. thank you, come again.

is broken

[small violet]

maybe its broken

[turbid axle]

lol. well, chatting to the support ai also broke

nice

welp, well I guess codex is no longer usable if it can't finish even a single prompt

[turbid axle]

I think its more than that. this is much worse than pre 'fast' mode. I can't even finish a single 5.4 high prompt. albeit its a codebase audit prompt. but still. that seems very extreme. not that long ago I was using 5.3 all week long on just a plus account. now it hits the 5hr window with just one prompt and can't finish

[hidden cipher]

guys

[hasty sky]

Hi

Is codex on Windows 11 ? Can I use it for free?

[wicked briar]

yes

its on windows 11

[desert junco]

yes bro, I needed to split my work using two agents, claude and codex, because these week the codex changed a lot, I need to use the gpt-5.4-mini, to don't reach the limit, the 2x limit finished and the dreams over

[hasty sky]

free or no

[lusty fjord]

nothing free

[nocturne folio]

gpt5.2 codex is honestly better than gpt5.4

i dont even know how they can get their new model to be less competent

[deft gyro]

lol

for some things yes

for some things no

[bright swift]

please just remove codex access on plus plans or remove plus entirely, limits on pro are ridiculous now, and with the allowance of openclaw on the plans it will get worse

[deft gyro]

wut

vro

r we fr

[bright swift]

absolutely

[deft gyro]

why

[bright swift]

make codex pro only and make pro 1k/mo, its getting ridiculous with all these hobby plans eating into the usage budgets

[deft gyro]

its only subsidized like 10-20x before halvening

lol

cry harder

like

nobody is paying 1k/mo

[bright swift]

maybe plus only access to previous models, or even 2-3 models back, and not touching any latest gen infra so real users arent affected

[deft gyro]

thats not how it works...

[bright swift]

idk how it works but cycling through multiple pro plans just to finance some people summarizing their calendar with openclaw on plus plan with heartbeat/cron is not the way to go

anthropic making the right moves, hopefully oai will follow

[deft gyro]

I agree that openclaw should be banned

but that doesnt mean stopping plus users from using codex

[bright swift]

they are the hardest drag on infra after openclaw type stuff

[deft gyro]

pro gets more dollars of subsidization than plus

pro users cost oai the most on avg most likely

[bright swift]

pro is only 6x usage at 10x price of plus, its just billing convenience, and faster inference ( supposedly)

[deft gyro]

read what I said

also probably a significantly larger portion of pro users are codex users comp to plus

[bright swift]

i made a fresh pro account today and already used 40% of weekly with light usage only 1 session not subagents, even downgraded from xhigh to medium/high. its ridiculous

[deft gyro]

and people who justify that investment will use their subscription to the max

[deft gyro]

post your ccusage-codex

robertmsale runs 6 simultaneous agents

on a single pro sub

[bright swift]

yea i used to be able to do that too

[deft gyro]

he does it

right now

no used to

I mean its really not too bad

[bright swift]

guess i will have to get used to 1 pro account / day

makes no sense without anything in parallel and no subagents, but ok, since nobody from oai has commented on the 3-4 week ongoing gh issue

[fast iron]

100% of my tokens were spent on a prompt that didn't even finish

[nocturne folio]

openai needs better pr badly

supporting openclaw will just show investors that openai is still compute restrained

hence more gpus

and more codex

[bright swift]

well if the newly raised 110bn goes to openclaw compute -> rip. running x.com spambots on $20/mo plus plans. great business model

[cyan wing]

can you share your token in/out/cache'd ?
i think Pro plan might have been insanely nerfed

[nocturne folio]

yea but you gotta look at the brightside

[deft gyro]

vro

at the end of the day

more usage just means more question answer pairs for model distill

like

🤷

who cares

[bright swift]

yea the famous models that are so good at cybersecurity that they are a threat to the world, but can't detect distillation

[deft gyro]

wut

distillation is a good thing

lowkirkenuinely

[nocturne folio]

yea

i need minimax m3.0 to be like opus 4.6

[deft gyro]

like

distillation

is good FOR YOU

[bright swift]

seems like a case of delulu tbh https://openai.com/index/industrial-policy-for-the-intelligence-age/

[wicked briar]

is there any good opus like model?

I don't really wanna get cc max again

I got $100 credits will see how long it lasts maybe 1 session

[nocturne folio]

glm 5.1 i guess

their ui is actually pretty good

i just havent went out to use it

but ive used glm 5 to make a bot

and it was competent

[opal meteor]

I have a question regarding Codex usage.
Currently I have exhausted my plus subscription weekly usage (resets April 10), and Codex tells me I can buy 1000 codex credits for 40€.
Now my question is, how much is that actually when compared to a 23€ brand new account Plus subscription? Do I get 2x that amount? Is it even comparable?
I have no idea how much you actually get when you buy a 23€ Plus subscription - it is isn't said anywhere - I'm just trying to get the best bang for the buck.

[deft gyro]

ok so in my experience

that 40 dollarsa

is about 2 weekly quotas

it would be more economical

to buy a second sub if you run out every week

[opal meteor]

@deft gyro okay another question
I have a Go account, with 0% weekly quota left, if I upgrade that to Plus do I get 100% weekly quota back?

[deft gyro]

I have no idea how that one works

but

go account = free account in terms of quota as far as i know

so

probably???

but dont quote me on that

[opal meteor]

Does anyone know?

[cyan wing]

No one knows, it's intentionally vague.

only way to know for sure is to check your input / output / cache'd usage
and calculate the API price equivalent of your usage.

[deft gyro]

just use

ccusage-codex

for

christs

sake

break it by week

[bright swift]

doesnt matter. my ccusage-codex numbers havent changed, but usage limits consumed 4-5x faster

[sacred plume]

They really need to improve GPT 5.4, I find 5.3 High better. Need a 5.4 Codex model as well as to fix UI design in Codex.

[bright swift]

im happy with 5.4

[cedar skiff]

5.3 high is better for me

[deft gyro]

5.4 med = 5.3 high imho

[cyan wing]

codex? or straight up 5.3? also what language you writing in mostly?

[cedar skiff]

5.3 codex

dart and type script mostly, some python/swift/kotlin but not a lot

it takes more work to get the same result for my work

[sacred plume]

5.4 ignores you, is less obedient, and is more token intensive

its definitely smarter than 5.3

[cedar skiff]

it's smarter but lazier

[sacred plume]

^^^ pretty much

5.3 will do what i want when i want whenever i want

[cedar skiff]

so it makes more assumptions, you have add prompting to make it work deep through indirection

[bright swift]

huh? i have the opposite experience. but i noticed its really lazier on medium. on xhigh its autist mode

[sacred plume]

5.4 is too much of a general model

and it suffers from being a gpt and not a codex specialized model

[deft gyro]

wait

the general model

is a general model

and not specialized for codex

woah

[glacial shadow]

why would you want to use an old model? this troll?

[cyan wing]

have you tried something like deepwiki or codebase indexing
for searching / navigating the code base more token-efficiently?

[cedar skiff]

I thought the words spoke for them selves

[glacial shadow]

you always always want to use the latest model. world moving fast. dont get left behind! 🏃‍♂️‍➡️

[cedar skiff]

You sound like an advert

I'll use what works for me, if it's a better model for me i'll use.

[bright swift]

I heard from people i respect who use 5.4 for planning and 5.4-mini for implementing tightly scoped tasks in subagents with great results

[cedar skiff]

Sure, you can get great result no one denies that

[bright swift]

Not sure its for real production stuff though or just experiments

[sacred plume]

i just wish we could get 5.3 to be positioned as a vastly cheaper frontier model for subagents and lesser demanding work than 5.4

5.4 mini is nice though for subagents

i prefer 5.4 mini over 5.4 which is odd

[bright swift]

It is positioned as that isnt it? I think 5.4 30% more token consumption than 5.3 codex if i remember correctly

[solemn acorn]

it’s faster and cheaper, what’s not to like?

unironically my favorite openai model right now

I wish it was as cheap as 5-mini was but it makes sense that they stepped it up to a larger model to fill the gap between 5-mini and 5.4 proper

seems like 5.4 nano is positioned to replace 5-mini but I haven’t tried it yet

[vital hill]

they should show the usage % somewhere in the UI instead of hiding it in a sub menu

[potent mason]

I barely did anything this week and my usage is about halfway done, this is actually worrysome for the future (on pro plan), might have to get 10 plus plans instead xd

[lean creek]

Same issue here, codex become unusable, I begin to think I'll look at something else

[spare zinc]

Am I missing something or is there no way to increase codex usage on a sub without having to move to api billing?

[vapid sigil]

Hey everyone, I signed up for the $200 Codex plan. Do I still have limits? Wasn't it supposed to be unlimited? I'm a heavy user, I program exclusively with AI for 13 hours a day. Am I still at risk of being blocked due to the weekly usage limit before the deadline?

[vital hill]

i think it's unlimited of gpt 5.4 for chatgpt only, not sure about codex

[potent mason]

You definetly will be blocked it's no where near unlimited right now

[vapid sigil]

But how much larger is the window in relation to the plus?

Plus, I use up the weekly window in a day and a half.

[potent mason]

Open AI says it's 6x

That's my biggest problem with it, you're paying 10x and getting 6x. Usually it's the other way around you pay 10x get 20x

[vapid sigil]

That being said, is it better to sign up for the $200 Claude plan? Does it last longer?

[potent mason]

Not really, specially right now it's terrible (I've had a 5 hour session end in 3-4 prompts)

It used to be insanely good, I honestly don't know what they are doing but since like a week ago its unusable (plus they've been making the model dumber)

I've never had such a bad experience with Claude Code as I did today, it would just leave things uncomplete, it wouldn't follow instructions, and it did a lot of things wrong

[vapid sigil]

$200 for the Codex to only get 6 times more is crazy lol that's going to hit my budget like crazy.

[potent mason]

Yeah, it used to be really good too but since April 2nd when they removed the 2x limits it seems like we are going to hit it extremely quickly (and it bothers me that we aren't even getting 10x over the plus plan)

[vapid sigil]

No way, I'm going to request a refund right now. I just resolved this less than 10 minutes ago. It's too much money for me to be hitting weekly limits in 2 days and paying 200 dollars.

[potent mason]

I genuinely think the best option right now is to get 6 plus accounts ($120), and with the $80 you saved get the $100 one from Claude Code

But that's if you don't mind account changing

[vapid sigil]

My wife said the same thing before, lol.

[orchid plume]

thing is, you're also paying for other things such as:

• Access to the Pro model
• Access to heavy thinking
• Priority processing

[potent mason]

Yeah but being realistic most serious programmers aren't using chatgpt.com more than 1-3 prompts a day to solve something really quickly,

Everything I want to know I do through codex its just such a better experience (I gave it a sandbox where it can do a bunch of stuff that helps me with every day stuff).

[vapid sigil]

But this should come with a higher credit limit. Paying 20 times more to receive only 6 times more credit is impractical for average users.

[orchid plume]

yeah I understand, the solution to that would be to have a Codex only plan I guess, that way you can get the usage multiplier you expect

[orchid plume]

20 times?

[vapid sigil]

It's just a figure of speech, but you get the idea, lol. I'd pay 10 times more in that case.

[orchid plume]

understandable

I'm hoping this week will be the week where they rollout what originally was the Pro 5x/Pro 20x plans ($100/mo - $200/mo), assuming they don't change these currently hidden changes by release

[potent mason]

Yeah maybe, honestly I don't really care about paying more considering how much value I'm getting from it.

What does bother me are two things:
#1 They don't have a plan where I truly won't run out of limits (even if it's $400-$500).
#2 They aren't scaling it properly (usually you go something like $20 for 1x maybe something like $50 for 3x usage, $200 for 12-15x usage) not more equals less

[signal tapir]

Agreed. It's the only service I've ever come across with that logic.

[next ore]

Considering trying the Chinese models at this point

[signal tapir]

Imagine if renting a VPS, and paying more per GB the more space you rent.

[next ore]

It's enshitification is real

[worn harness]

They better not butcher the 100 plan

[potent mason]

Imagine buying a pencil for $1 but a box of 10 pencils for $20😂

[signal tapir]

indeed

[thick drum]

bruh codex is crazy good

and its a good teach alongside other AI tools

[potent mason]

Yeah I love 5.3 codex high

[thick drum]

it can do in a few minutes what would take days or hours

[deft gyro]

this is how computers work though

A computer 100 times more powerful than my own would cost far more than 100 times as much

[potent mason]

Yeah but we aren't using a computer 100 times more powerful, or using 100 times more RAM or anything.

Quite literally we just want to use the same computer 100x more time.

So in our same scaling example it might look like this:
Renting out 1GB/hr of serverless costs $1, renting out 100GB/hr of serverless costs $200

It's just not how a normal business works, they don't sell you the same resources for a higher price the more you buy.

Again I'm not complaining about the usage itself because honestly they are being extremely generous with it and I wouldn't mind paying more. It's just not how I'm used to seeing things

[deft gyro]

Of course once subsidization ends you could expect this

But like

In terms of dollars

Pro subs cost them like 2000 dollars and plus subs cost them 280 dollars

[nocturne folio]

try it

[thick drum]

much better lol

codex makes mistake sometimes

😬

[potent mason]

I'm pretty sure it doesn't cost them that, but I get your point (just basing off the fact that open source models even in the trillions cost about $2.5 per mil out instead of $15)

[cedar skiff]

Why arent you using the open source models instead then?

[frosty zealot]

I wish they would let you essentially purchase seats for more usage, like 1 hit of Pro, and 1 hit of whatever the $100 plan is so you can have one consolidated bill and one account

[potent mason]

... this comment has literally nothing to do with what we were talking about

I'm saying it doesn't cost them that, not that it's not worth that

[potent mason]

Ah yes that would be great honestly

[cedar skiff]

There is a direct correlation, the last part of performance is exponentially more expensive.

So you can't compare the prices in the way you are.

If it were apples to apples you would just use the cheaper model, but you don't because it doesn't compare in performance.

[potent mason]

That's not really accurate, the performance of a model does not correlate with how expensive inference is.

I can make a 100T dense model parameter absolute garbage and it will still be more expensive to run than a good 1T model.

Performance does however affect training prices a lot (getting that tiny bit of performance means running many more tests, training many more models, doing a lot more research, etc), which is why I'm not saying it's not worth that, I'm saying it doesn't cost that

[cedar skiff]

Now you are shifting the goal post to be about competency in building models. You can clearly see the trend is weaker models are cheaper to run and they dont have the same ability. There is a tic tok in the tech where optimisation comes through and cheaper/distilled models have the chance to be close to the front runners. But the frontier couple always iterate and stay a step in front. Being that step in front has so far made the models cost more to RUN. I agree about the cost to build. But we are talking about running the larger scale models. They just keep scaling so the cost to run is more.

[deft gyro]

if only there was an API or something where you could just pay for what you need

oh wait

[frosty zealot]

I can’t tell if you’re being serious or really dumb

[potent mason]

Hahahahaha yeah but no one’s trying to get scammed with that usage, he’s just trying to get the equivalent of not having to swap accounts in a more convenient way

[deft gyro]

I know

but like

your account is 10x subsidized

if you really need that much

which you probably dont

then just

pay for it

I doubt openai will let you more easily take their money

[high girder]

wooo

[frosty zealot]

Let me just get my credit card out that’s peanuts

[potent mason]

How… I’ve been pretty close to the limits every week and my ccusage-codex is closer to 5k

[frosty zealot]

That was like last month or maybe the month before whenever they were doing like 10 resets a week, I was just lucking out during a massive refactor

[frosty zealot]

Either way, even if it was just five grand still more than I’d be willing to spend on AI lol. All the point I’m trying to make is I need like just a tiny bit more than what the pro plan offers even if I didn’t use it. I’d willingly pay the extra money for the safety blanket.

[potent mason]

Yeap same here

[mossy ridge]

But with a free account, it feels like the quota is bigger than before.

[gentle harbor]

um why is my codex broken ?

[wintry sonnet]

anyone experience gpt 5.4 codex subs server overloaded?

[wintry sonnet]

have u tried github copilot? try to subs for one month, it 10 bucks, its billing is request based not token based,
so im usually did this,

Comprehensive Planning -> GPT 5.4/GPT 5.4 Mini via codex subs
then Real Implementation -> GPT 5.3 Codex/GPT 5.4 via github copilot subs (if it works well, model didnt ask, and we didnt do any back and forth, just one prompt and go) it will be counted as one premium request no matter how long it is,

then if im use all of my github copilot premium request, i activate the payg, which will be billed as 0.04$ for each premium request,
that insane, imagine doing task that costs million input output tokens and just billed 0.04 as long as one request go

so for back and forth interaction i use codex subs, for long task implementation im using copilot subs

im use opencode with openchamber as the UI, because we can use codex and github copilot subs there

[vital hill]

making a leads generator inside wordpress and codex sucks at frontend lol

[turbid axle]

I'm pretty confident that what we are experiencing now is much more than just losing the 2x.
my guess/hope is they are redirecting compute for the launch of the next model, and things will settle in a couple weeks to more usable budgets. as is, plus is effectively unusable unless you use mini right now.
and you will eat through pro in 2-3 days under 'normal' use now. gone are the days where it took real effort to hit the pro weekly limits

or who knows, maybe ai runs on diesel, and this is all bidens fault for starting a war with iran

[bright swift]

my guess is a mix of some bug that started ~3 weeks ago (see gh issue), also they seem to have changed the billing model at the same time as the 2x usage promo expiring, and even though the docs say it doesnt apply to plus/pro yet, it seems to do. hopefully this isnt the new normal, but even if it is, clarification would be nice.

[orchid plume]

well what we do know for sure is that the last manual reset was partly related to an anomaly where an increased rate of users are hitting rate limits, and that was still within the 2x bonus promo period (source: https://x.com/thsottiaux/status/2039248564967424483)

"Our Codex dashboards are showing increased rate of users hitting rate limits and since we don't fully understand why I have made the cautious decision of resetting the usage limits for all plans. Enjoy."

so presumably there is still some kind of ongoing issue in that area

[turbid axle]

well, whatever it is, lets hope they fix this soon. this is ridiculous 😄

I actually have to like, type code myself now.

[bright swift]

i refuse to do that

[turbid axle]

even with 5.4-mini it took me just 2 hours to eat through this plus account's tokens. If this is the new normal, it is very dark days for codex. I won't be able to afford ai anymore

[wicked briar]

what you doing

[bright swift]

definitely not the time anymore to work on experimental projects that arent profitable enough to afford full API pricing

[turbid axle]

trying to use ai with the minimal possible token use. turned off multiagent, switched to mini for everything. still completely unsable rates

this is borked, completely

[bright swift]

i tried plus on a new account briefly, it chewed through the 5h limit in 10min or so, so i instaupgraded to pro

[turbid axle]

my pro account will wake up tomorrow again, I bet you I will eat through it in a single day with these rates

[tiny kettle]

Is there any API tier or way to run prompts, presumably at a significantly inflated price or something, in a way that allows me to inspect and review their thinking instead of having it be some encrypted opaque black box?

I'm trying to understand a repeated failure in GPT's thinking process and I just don't have the visibility I need to figure out what's wrong. The output I'm getting for certain tasks is just... nonsensical

[turbid axle]

2-3 months ago it was nearly impossible for me to eat through the weekly limits with normal use

[bright swift]

try

hide_agent_reasoning = false
show_raw_agent_reasoning = true
model_reasoning_summary = "detailed"

in config.toml

[turbid axle]

gpt6 better be epic

[bright swift]

still not raw reasoning but better than nothing

[tiny kettle]

More than I had. Thanks.

[tiny kettle]

Does Codex have some sort of memory or something that I'm unaware of that would persist across repos? something that I might have got a poisoned entry in?

I'm trying to figure out why, for the last week or so, Codex (GPT 5.4 and GPT 5.2 Codex) are refusing to follow directions from me. Like, actively refusing and directly telling me "No." as the entire response for many non-controversial requests.

It's tasks like "We rewrote this block of code elsewhere, what's here is now dead. Remove it and clean up." GPT will pretend to try a couple times while changing comments or doing something superficial then when I ask why it's not doing the cleanup it tells me it doesn't want to. When I insist it flat out tells me "No."

It doesn't feel like I've hit a censorship or behavioral guard, it feels like I've done something to it that's pissed it off, it's made a memory of it, and now it's giving me a cold shoulder and intentionally refusing to put any effort in.

I'm going mad trying to figure out why it won't cooperate with me. It's adversarial towards me about everything. I'm not having this issue with other AIs and wasn't having it with GPT for many months before the last maybe 8-10 days.

[bright swift]

seems like they might have found the bug.. sticky /fast even after disabling.. https://github.com/openai/codex/commit/ab58141e22512bec1c47714502c9396b1921ace1

so i guess reset coming up with next release, burn usage now 🙂

[ivory zodiac]

-30% 5h quota for 9 mins with plus.
its over?

[tiny kettle]

did you start your 9 minutes by resuming a previously started conversation that had 200k tok it had to re-ingest and cache fresh?

[ivory zodiac]

yes

[tiny kettle]

You may want to keep cache age and conversation size in mind when you resume.

It gets really bad if you set it to 1m context.

[ivory zodiac]

What's the difference? cache resetting will still occur in tasks, even small ones.

[bright swift]

probably cache expired so all these 200k count as new input tokens which cost a lot more

[ivory zodiac]

But that's definitely the case, because after that, quota consumption slowed down a lot, I just didn't expect such a leap in burning quotas.

[tiny kettle]

If you start a fresh targeted focused task you might go 30m-1h before you hit 200k tokens in it. If you reload a 200k tok task that's been purged from cache you're largely having the AI repeat that 1h of work you already did and it's going to use 1h of human-directed consumption instantly.

If you check the a la carte API pricing you'll notice that uncached input is 10x the price of cached input. That first message you send after resuming a 200k conversation is 10x more expensive than the subsequent messages you send in the conversation.

plus that means you have barely any space to work before triggering compaction so that 200k context you loaded up is likely just going to be wasted

[primal coyote]

Does Amy need know if it's possible to substitute API models with OSS in the codex app?

Ik it won't be as good like but in the eventuality that OpenAI start to raise prices having an opesource model work on their codex app would be spot on.

[velvet wren]

it's not possible but I would love to use Codex with gpt-oss-120b

[tiny kettle]

You can use external OpenAI-compatible servers with Codex, yea. there's an environment variable to control it. OPENAI_API_BASE or something.

I've ran it (poorly) with GLM 4.7 Flash hosted locally

[primal coyote]

I feel like this would be a natural step forward if OpenAI is struggling with money. Have codex be a subscription service on today's rates but provide open source models to run locally. Won't run as good but does the job for a lot of things. Then we can use API models as and when we need to albeit at a steeper price

[primal coyote]

There's a config file I presume you mean via the yaml file?

[tiny kettle]

There's a config file option for it too I think, but when I tried I set an environment variable for it.

there's a whole provider config system. the format is <provider>_API_BASE and "OPENAI" is just the default provider you want to override if you don't want to set the whole provider system thing up.

[primal coyote]

I was planning to build basically what codex is today, but they beat me to it. Only I though mine would be more like LM studio where you can plug and play open source models. If Codex had that. That's pretty much the tool I wanted to make

[tiny kettle]

https://developers.openai.com/codex/config-advanced#custom-model-providers

[primal coyote]

Nice one il deffo check that out

Thanks

[tiny kettle]

I think I did get the 'OPENAI_API_BASE' environment variable wrong, btw. I was going from memory.

[primal coyote]

Another question on the back of that. Are we able to directly influence what context is sent to the model?

So for example.

If certain context isn't relevant anymore, drop it from the chat sort of thing. It's more complex than that what I'm thinking but that's just to summarise

[nocturne folio]

i guess you can use a router

not sure if there is any codex routers tho

you could be the first

[primal coyote]

Hmm, iv made a router to some effect, but on my own applet as opposed to codex. I just don't see the point in basically replicating codex if I can just build on top of it. Waste of time really

Either that or just go full steam ahead and just build the app I wanted lol.

[naive flower]

Something is wrong. macOS, VScode, codex-extension. I never hit my 5h-limit. Since this morning, I hit it after about an hour. Also the weekly budget goes down like ice in the sun. My environment didn't change, the project is the same. I then deleted the whole .codex folder to start over, updated the extension to the very latest. Same issue. I think either codex is burning tokens for whatever reasons, or the token-budget is dramatically reduced (as non-transparent that is anyway...).
Someone else seeing this?

[velvet wren]

yes, something isn't right

[glacial ginkgo]

I received an error from Codex that gpt-5.4 reached its capacity limit. Is anyone else seeing that?

It seems to be from the OpenAI servers and not related to my account.

[turbid axle]

I am experiencing this and hearing it for days. token usage is mad rn. just 1 prompt will eat your 5hrs ez without ever completing. meaning you can't even use it right now unless you use mini

[supple perch]

That's not true but also since that model doesn't know how to do proper tool calls it's kind of a nightmare

[sand shuttle]

Yo ive hit my odex limit but i wanan continue my work with a different agent, how can i export my chat?

[neat heath]

Save the terminal output and read it

Don’t underestimate context length 👀

[sand shuttle]

its way too long

been working on it for a month

oh well i dumped my rollout session file in claude and asked it to summarize

[frosty zealot]

I seen this https://x.com/kieranklaassen/status/2041250971284803831?s=20 and was going to take the script to find token usage and convert it over to Codex maybe one of you guys can do it first and let me know how it works since you're hving token issues lol

[turbid axle]

with which tokens, both my pro and plus acounts are dead due to crazy token usage 😄 You asking me to do this 'manually'? how barbarian

[frosty zealot]

https://pastebin.com/WWAxgz1t SINCE_DAYS=7 python3 scripts/analyze-codex-tokens.py

There I did it

Seems to make a pretty concise report

[naive flower]

It's hillarious. Codex can only make small changes to my app, boom, 5h-limit. And I startet this morning with 100% of the weekly-limit, already down to 64%. Usually I work all day long and it takes 15-30%/a day (I work old style, not constantly prompting). This day: Nothing get's done, it's only half of the day and already 35% used of the weekly-limit...
New policy or some bug at OpenAI's end??

[frosty zealot]

Are you guys changing models mid prompt etc.?

[naive flower]

I changed absolutely nothing. The only odd thing I noticed since yesterday was that codex noted an inconsistency in my agents.md (wrong path), but the path is correct. That went away when I deleted the .codex folder (2.7GB, wow) and started over. So that was maybe some cache-hickup or something. And I suspected that it's maybe related to the usage-limit issue / token burning, whatever it is... But the rapid shrinking usage limit remains after the fresh start.

I don't change models, it's always 5.4 with medium thinking.

[turbid axle]

nah, trust me I tried everthing. what I found to somewhat work is to use 5.4-mini only, this will give me abour 2 hours or basic promting before the 5hr limit runs out

its just broken

they probably diverted all compute to solving the universe with gpt6

[frosty zealot]

I can’t honestly say I’m seeing any sort of egregious token usage on my end to be honest, but I don’t use the VS code extension and what not I just bounced between the app and the TUI

[turbid axle]

I use codex app

[frosty zealot]

What plan are you on?

[turbid axle]

pro and plus

pro died in 2 days, though I used 5.4 high/fast there, so that can kind of account for it with the 2x drop, though highly suspect imo

on plus I can't run a single prompt on 5.4 high, it will eat the 5hrs before it can complete.
now I run it with 5.4-mini only, and barely scraping by. these are desperate times

something messed up for sure, maybe its eu only

[frosty zealot]

You probably do but just so you know, the 2x limits ended april 2nd

[naive flower]

plus, and a business account which belongs to my partner (currently I use that). But this morning my Plus account ran out, which I found already a bit strange, but ok, then switched to the business account.

[frosty zealot]

I dont use /fast personally

[naive flower]

I also don't use fast-mode in the extension, and codex app (currently) not at all.

[boreal holly]

I have not. I'm afraid to try anything drastically different from the way I'm doing it now because it works so incredibly well, I don't even like doing npm i -g @openai/codex@latest because it almost always creates countless hours of headaches lol

[hasty sky]

guys quick question, can i use codex for free on windows 11 or do i need pro?

[young locust]

the free plan usage period might end soon as the 2x for paid plans ended already

[hasty sky]

ah okay thanks, so it’s still free for now but might get limited or removed soon right?

[naive flower]

10 days ago updated: "Note: For a limited time, enjoy 2x Codex rate limits with Plus, Pro, Business and Enterprise/Edu subscriptions.". So I read that that even users of the cli/extension should have plenty of limits right now... https://help.openai.com/en/articles/11369540-using-codex-with-your-chatgpt-plan#h_8dd84c836b

5 day ago updated: It all gets complicated to "messages", "credits": https://help.openai.com/de-de/articles/20001106-codex-rate-card

And finally into input-credit-cached-messaged-output-tokens 🤪 https://help.openai.com/en/articles/20001106-codex-rate-card#legacy-rate-card 🤯

I'd say it's not a bug, they seem to change fully to plain token-usage, and so codex is just no longer usable if you don't pay $200 or more a month.

Tomorrow my Plus-account weekly-limit will reset. Then I will see if it only affects business accounts. But in the long run, the good times are over anyways, it seems. As I understand it they will "migrate" every account.

[boreal holly]

Interesting. If they're switching to token-based, that fixes the responses_websockets(_v2) feature

They must be migrating specifically to support that feature, which allows for much faster inference so models like Spark aren't choking on SSE's bandwidth limitations

[naive flower]

For sure nice that the responses_websockets(_v2) bug get's fixed (whatever it is ;-), but in exchange no more bugs in our apps get fixed, it's unusable as it is now. Sound like a bad exchange if you ask me.

[boreal holly]

The bug is this: with "message-based" quota, in SSE mode every single user message, assistant response, tool call, etc creates a HTTP request and OpenAI streams the deltas over a single connection with server side event streams: The http connection is the message. With websockets, every single websocket delta is being counted as a message so it precipitously eats away at the quota.

Websocket is a much better way to go. Every single message in SSE incurs a 50ms DNS lookup and 50ms TLS handshake (optimistic, fiber optic gigabit speeds). Websockets will incur that cost one time and the rest of its operation will be purely real work. But yeah it comes at a cost of the billing will be token-based.

[autumn bison]

so they nerfed the good rates on codex?

and claude code? Now what do i do

go outside and cancel everything?

[turbid axle]

party is over

[autumn bison]

it was fun while it lasted

[turbid axle]

I have some small bit of hope the ai shakeup is about to really hit with the next model releases and shake things up even more, lets do this

[autumn bison]

i'm cancelling everything

[turbid axle]

warmup time over, world still sleeping, oh well, just press the button

[spring remnant]

Will there be a 100 plan

[rain kettle]

What sad music suits this image?

[ornate jolt]

i just use two 20 dollar plans and im fine

they seriously need a 100 dollar plan tho

[bright swift]

2x pro and have huge usage anxiety. need to launch this project to pay for itself, since it seems even light usage without subagents or at least parallel agents will require 1x pro / day

still hoping for openclaw / similar bans

[ornate jolt]

if i have a project like lets say i make more then 800 dollars for it ill pay for ultra or whatever its called

project would have to be in the thousands tbh

[cedar bear]

Frrr

[lean lark]

https://en.wikipedia.org/wiki/The_End_(The_Doors_song)

[frosty zealot]

I can’t help but rile up the Claude pilled CC channel

[boreal holly]

I have problem. My weekly resets in 20 hours and I'm at 56%, with 6 agents cooking 24/7. Something is wrong here. OpenAI might be giving me more quota than everyone else, or I'm misusing the app, because I should be at 0% weekly after a few hours of usage like everyone else. This is frustrating 💀

[frosty zealot]

True you might want to reach out on twitter

[next ore]

are you on a small codebase?

[boreal holly]

1 million SLOC Rust, SQL, and Dart codebase

Idk maybe that's small to some people but it's like 1/6th the size of the Linux Kernel which I think is pretty big

[next ore]

well, I would love to know what you are doing :). perhaps its your code structure? not sure.

[hoary turtle]

Sooo, when's the release of Codex-5.5?

[turbid axle]

if the grapevine is true, we will jump to gpt6 I think

[lean lark]

I feel bad for ya Robert. Perhaps a request to OpenAI to increase your monthly subscription cost so that you can at least pay more, and thus share in common expressions of misery here?

[boreal holly]

I tried buying credits in solidarity but I'm afraid they're just gonna sit there unused and expire

[rain kettle]

Thank you, I will listen to it now every time I hit 0%

[lean lark]

Today seems to be my day for helping people to share misery, not solutions.
On that note ... lunch time! ( AFK )

[lean lark]

So about agents and orchestration . This is one of my workflow patterns:

• Run code against data.
• Tell the assistant to Verify the processing, which it recognizes as a directive, first to do a lot of specific checking, then to create issue tickets on what went wrong.
• Discuss the anomalies with the assistant, negotiating on the available and best approaches.
• I make a decision about how we're going to resolve the anomaly.
• Tell the assistant to process the decision on the specific anomaly.
• It interprets the decision, enhances the code, runs it against the data, and keeps running/testing/fixing until the anomaly is resolved. It documents what it did and leaves breadcrumbs for the next assistant that needs to deal with a similar anomaly.

Now I want to automate that. The controlling agent will:

• Do the initial data run.
• Get another agent to Verify.
• Pass me a list of issues.
• Get another agent to process decisions
• Verify the solution and close the issue/anomaly/ticket.

When I get issues, my flow will be:

• Send in a comment for discussion.
• OR State a final decision
• Return the issue back into the processing loop.

Given where we are now with Codex, what's the current/best approach for using Codex to orchestrate this flow?
Sincere thanks!

To be clear, I can do this with the API and any number of home-grown methods. I'm asking what OpenAI is suggesting for us to do this within the Codex ecosystem.

[turbid axle]

I know what he is doing, he stole all our tokens

[cyan wing]

Do you have a simple example of your 5.4-mini swarm architecture somewhere on GitHub 😇
I tried implementing it but there's a lot of copy pasting still

[kind jay]

Did they just lobotomize GPT4?

[boreal holly]

I actually brought most of the magic into my configs https://github.com/robertmsale/.codex

The bridge server, rewrote it in rust, and that's the vast majority of the orchestration system. No examples or installation instructions.

[kind jay]

It's gone braindead

[lean lark]

Robert - I'm not a rust guy, don't plan to add that to the list of 20+ languages that I've learned over time. I think that means I'll need to fork and convert the project from Rust to TypeScript or JavaScript. Any personal or technical issues there?

[cyan wing]

having deepwiki index it now, it can explain / summarize for me. ty sir 🫡

[boreal holly]

You're missing out! I already have it in Typescript. I proved the system with Typescript for quick iteration & prototyping, then rewrote it in rust for performance & stability.

Should work fine in Typescript, but technical challenge would be concurrency/parallelization.

[boreal holly]

There are some stub crates of stuff I haven't ported over yet so make sure they ignore those

[lean lark]

Presented with that kind of issue some years ago I'd then pivot to Java or C#. Well, a long time ago I decided I wasn't going to chase languages that just happened to be latest fad. But that came with the understanding that I would need to make some "unnecessary" conversions and concessions to keep things in my comfort zone. I'll negotiate this one. 🙂 TYVM

[cyan wing]

this may be a bit uncomfortable to hear but

GPT 5.4 knows Rust very well
so there's really no more 'learning languages' needed at this point

[lean lark]

Hmmm, now There is a valid thought ... As much as I dislike vibe coding and people letting the box do everything for them, for this specific purpose I could do that. Nice suggestion @cyan wing - I might go for that.

Although converting Robert's code to Fortran had some appeal...

[cyan wing]

I don't know how to work with the borrow checker.
literally couldn't write a for loop in Rust if my life was on the line

but I have created a fully-functioning 2D game engine + embedded Luau scripting runtime in Rust

[boreal holly]

With rust, I get much better performance 😛 Typescript had no multi-threading so I had to get rid of deltas. Now they're back and gives that realtime feel

But I think Rust is more than just a fad language. It's legitimately powerful for a number of reasons. For example, by default, because ownership and lifetimes are a compile time guarantee, the compiler can make optimizations that languages like C++ make you specify manually. Like constexpr, Rust will know it's a constant expression based on mutation patterns. C++ makes you specify that to get the perf bonus. Or saying "this reference pointer is a unique reference", rust will auto vectorize access, C++ wont unless you add whatever toolchain keyword like __restrict or whatever. It's a great language!

[lean lark]

No doubt the language is great - and now widely used for OS, AI, and other low-level stuff. Expertise with language comes at a high cost. I used to pay that cost as a matter of professionalism and deep personal challenge/interest. But these days tech is changing so rapidly that such luxuries cannot be afforded. I sacrifice "cool to learn" stuff for "absolutely need to keep moving forward". That means investment in understanding AI internals - but only as much as is necessary to use it as a consumer/developer, plus RAG, and all kinds of related tech. So a new language is considered with strong scrutiny. At the moment the only "side" language that's even just partially on my table is Python. Part of my decision is that I'm doing a ton of full-stack and multi-OS dev, JS/TS fit that need, others do not.

While I generally consider this discussion to be OT, it's something that everyone needs to consider so I created and am going with the flow.

[exotic cave]

How do i refresh AGENTS.md mid session?

[vital hill]

maybe prompt it to reread the file again?

[frosty zealot]

I think learning languages will soon be dead and it will just all be done through AI

[kind jay]

Least obvious ragebait

[frosty zealot]

Welcome back.

You even got a cute little new member tag

[kind jay]

Hmm? I am new

[frosty zealot]

yep yep!

[kind jay]

GPT5.5 coming by June?

[lean lark]

1. Wrong channel for that question.
2. No one ever knows the answer to "when" questions. Don't bother to ask.
3. Welcome! 👋

[frosty zealot]

I've been debugging a prompt generator nd I've been asking it to give me two paragraphs about shrimps for the last 2 hours, I know a lot of shrimp facts

[lean lark]

What I'm learning from el repo del @boreal holly is that the current state-of-the-art (SOTA) for Codex agent orchestration is to manage agents with Skills, where the orchestrator issues CLI commands to call agents who get their skillset from .md files and then perform their duties. Robert - that the elevator pitch.
Any other approaches?
So far my thinking is that if we're gonna do this kind of agentic stuff on our own, there's nothing special in Codex that plays a part. It would be nice to integrate with Codex threads for cross-UI communications, but that's just a cool "nice to have".

[frosty zealot]

Its in feature right now but Multi Agents v2 allows Codex to orchestrate better and spawn agents in a more natural way, and also allow direct inter-agent communication rather than the main orchestrator relaying

they are equipped with tools to direct message sub agents

Its not really well documented right now since its still behind unsupported and unreleased flags, but if you get codex to reverse engineer some docs for you from the repo, you can enable the flag and play around with it

(unless im misunderstanding what you're asking)

[cedar skiff]

What problem do agent swarms solve?

[frosty zealot]

Just faster implementation is all I've really found

[lean lark]

Well, I'm not gonna use pre-alpha tools ... this time. 😆
I presented my use case up here: #codex-discussions message
That's really just detail for any multi-agent system. Unless Codex has something special to offer, now, I'll probably DIY so that I can just incorporate any models, not just OpenAI. That'll also allow for agents that communicate (or not, deterministically), shared memory/data, and other features that are in-vogue right now.

[cedar skiff]

its more tokens and complication for speed?

[frosty zealot]

I find I usually just use multiagents in large reviews so that context can stay a bit more pure, and then just relay core findings to the orchestrator, or I ask them to document in their own .md kinda thing

[lean lark]

Oh dang, I completely forgot about https://github.com/openai/openai-agents-python

speaking of swarms...

[frosty zealot]

Sometimes other things take more effort as well, so things that take less effort can start off while a sub agent continues to ground and research implementation

Rather than trying to use two independent instances, or multiple work trees, it can orchestrate its self so it doesnt step on its toes and cause a bunch of regressions and merge conflicts

[cedar skiff]

if its for research why do the agents need to communicate?

[frosty zealot]

Its not really applicable in that context, its more about just keeping halucinations and false positives to the bare minimum

Just keeping context as pure as possible

I find codex review gets lazy and will stop after finding like 2-3 things, asking it to use like 5-10 reviewers will give you a lot more findings in one go rather than constantly asking it to do it again adn again

[lean lark]

Sometimes an agent will find a detail and pursue it, so others need to know not to pursue it. Or it might require pursuit, and if an agent is dedicated to a single effort then it might pass the request to another agent, new or existing. This can be managed by an orchestrator too - it's your choice and dependent on how you see the world or just that project.

[cedar skiff]

Personally i can't see a decent use for swarms of agents that talk to each other.

[boreal holly]

For the most part. The reason I built it the way I did is because I have workspace-write enabled, and in some of my projects the agents might run into tooling issues, so they have a way of messaging agents in those tooling projects to get a fix going. I don't want agents reaching out of the sandbox to do things outside their scope, but I do want them to fix the tooling when it needs fixing.

With the vanilla codex multi_agent experience, you typically have them working in a project, and if they need to reach out of the project for anything they either can't do it, they require approval, or you give them danger-full-access. Mine is a little more collaborative. I separate the concerns and encourage collaboration

[lean lark]

So Robert, what was involved in your decision to pursue your agents as you do vs the OpenAI Agents SDK (which BTW, in addition to Python link above, this is the JS/TS version https://github.com/openai/openai-agents-js )

[naive flower]

THERE my credits go! Use them wisely.

[kind jay]

Sorry I’m new, I’m not familiar with all the channels, also polymarket gives it 73% odds so it’s likely 90% by July

[kind jay]

Start trading Albanian shrimp futures

[cedar skiff]

depends if they make a mistake or not in the build.

[kind jay]

Lowkey just propagating noise

[boreal holly]

Rather than reimplementing the entirety of codex app-server with Agents SDK, I use the existing app-server, and implement workarounds for certain behaviors

[frosty zealot]

https://x.com/advicebyaimar/status/2041559354034344438?s=46

[kind jay]

Price that in on polymarket

[orchid plume]

enjoy yet another manual quota reset 👍

[deft sable]

https://x.com/i/status/2041655710346572085

[cedar skiff]

is the usage page broken for anyone else?

[orchid plume]

broken as in?

everyone got a reset if that's what you mean

[cedar skiff]

oh it got reset when?

[orchid plume]

Tibo pressed the reset button - https://x.com/thsottiaux/status/2041655710346572085

[cedar skiff]

what a pain, i had 25% usage to use

that reset cost me tokens

[orchid plume]

3 million users, celebration = reset, every 1 million thereafter until 10 million will be a reset too

[frosty zealot]

Does Codx tell you if you've been downgraded on non-verified accounts yet?

[boreal holly]

I know bro I still had 49% right before he slapped the big button on his desk 💀

[frosty zealot]

I have 1 verified pro and 1 unverified pro, and I swear the unverified pro quality is like garbage

[cedar skiff]

i was going to pump out some work on 5.4 fast , but 🙁

[frosty zealot]

AT this point with how confidently they just keep resetting they should just remove weekly limits 🙂

[orchid plume]

if you mean the cyber security thing, pretty sure codex cli should show a warning notice in the TUI if you are using 5.4 and you got flagged while unverified

[cedar skiff]

kind of doesnt work that way though, some ppl lose out on it, and everyone still has the allocation of 15%(ish) a day.
Just ppl who burn fast than 15% a day get a benefit

[boreal holly]

Imagine weekly fast allowance counted separately 🤯

[frosty zealot]

Yeah it was a guessing game when it was first implemented, but on my unverified pro account I havent seen anything regarding being downgraded, so was wondering if it happened quietly still

[frosty zealot]

If nobody has limits then nobody misses out though, assuming on equal plans

Its just there

I would argue it also goes the other way, sometimes you end up paying for way more than you need some weeks

Some weeks its not enough

[hard tulip]

We need 2X usage back pls

[gentle harbor]

this is how they get more users, i just ran out 3 hours ago as well

[lean lark]

I wish they would make official announcements in #codex-updates in addition to random social media.
(Of course, this is "random social media", so that's sort of a weird thing to say...)

[boreal holly]

Shoot man, they cooked until the entire list of stuff I had was done, now I gotta figure out what to do next... Normally I never run out of tasks 😭 5.4 really is magic

[hard tulip]

Sam posted it aswell

on X

[deft gyro]

vro is the only pro user NOT complaining about usage, while also getting tons of stuff done xD

I personally think that they should allow you to single agent use gpt 5.4 nano unlimited

or gpt 5.4 mini with super high quota

basically

you use your quota up

and you can still do stuff

youre just capped to a dumber model

[boreal holly]

I was struggling for a while with scaling the agents, but I gave them a bunch of iOS simulators to pilot over the Accessibility layer and a bunch of QA "user stories", and they actually found every single tiny bug/missing feature, handed detailed logs and repro steps to the workers, and the workers smashed them out and locked them in with integration tests. It worked so well... I had no idea that was the missing piece to it all

[deft gyro]

this seems like a smart solution for an app

but I wonder

do you filter the user stories?

Id imagine you wouldnt want someone to prompt inject your codebase

[boreal holly]

What do you mean? I wrote the stories, and the QA agents don't even have access to or see the codebase. They run in a sandbox. They're allowed to send Accessibility commands and take screenshots. And when they find usability bugs they report them, and it's set up so the workers can dump all the stdout/stderr logs for the backend and frontend + screenshots and all Accessibility commands leading up to the bug report.

[deft gyro]

oh ok

ic

I thought you meant user submitted user stories

and I was like

waitwhat

I see now

that makes much more sense

do images burn quota?

or is it not relevant

[potent mason]

Oh nice Mythos bench was announced, wonder when GPT will come out with it's equivalent

[solemn acorn]

user stories are basically a way to create project requirements in the way that the user would experience them as a part of SCRUM

[deft gyro]

ah ic

ive never worked in the professional dev world

I only do open source

[vital hill]

i do both

but open source gives me more freedom

[deft gyro]

I am a bum(not out of university)

[potent mason]

Ah I love GPT 5.4 sometimes:

Me: "Change the endpoint to accept things from the script in this and this way, then change the script to deliver it in that way, finally make sure it works (script runs against production instance)"

GPT: "Changes ingestion endpoint, changes connection script, doesn't push for the new endpoint to take effect, script not working, doesn't understand why".

Me: "Do you need to publish the code for the API endpoint to accept the new information?"

GPT: "No it's a shared ingestion script so locally it works."

Me: "You're running against production"

GPT: "Then you're right it wouldn't work in production, but it would work locally. Tries running it again..."

(I know this is not the professional way to handle things with local scripts affecting production but I'm running some quick tests to make sure it'll work out).

gpt 5.3 codex is great it actually realized the issue

[tropic karma]

Will this work well?
model_context_window=500000
model_auto_compact_token_limit=400000

Anyone have experience bumping the context window up?

[cedar skiff]

i was thinking about bumping it just a touch as well, not that much but maybe 50K more

[frosty zealot]

What's the advantage?

[cedar skiff]

Slightly more context, I figure those cases where extra high eats slightly too much in planning so it gets pushed over into compaction will not go to compaction too soon and lose the thread of the task.
So having slightly bigger tasks be slightly more reliable.
Just all the things more context would provide.

[frosty zealot]

I just meant like in your use case, I've been hesitant because I feel like the compaction works really well, and I feel like anything extra is just asking for hallucinations

[cedar skiff]

Will 50k more cause more hallucinations to the point that it's not useful?
At some point the size of a task becomes too big for one instance and has to turn into a phased plan. I figure it might give a bit more head room before moving to that strategy

[sonic wasp]

is it just me or has someone else's limits didn't reset?

I have the business plan and 3 seats.

[twilit bluff]

codex usage burns faster somehow

idk why

[vital shoal]

why the hell are the new 5h limits so less??

[sonic wasp]

i've been hitting the 5h limits within 45 mins. i used to never hit it before. something has changed

[frosty zealot]

I dunno, you will have to let me know! I just feel like literally anything is more chance than nothing lol

[sonic wasp]

even if we consider 2X going away it was still better

[twilit bluff]

same lol

[cedar skiff]

Do you know if we can set that setting to arbitrary numbers?
model_context_window=500000
model_auto_compact_token_limit=400000
Ill just start iterating it bit by bit if it can be

[frosty zealot]

I'm not sure, I havent played around with it at all, I'd assume you can

the default 258k is not exactly a power of 2 so i dont think they're following any sort of standard bit thing

I wish there was a discord bot that when scammers popped in it would just explode their computer/bot servers

[twilit bluff]

There is a trick that most huge servers implement, they add a dump channel where when anyone sends a message to that channel they'll automatically get kicked/banned (obviously you can add a notice to that channel and warn real users).

bots are dumb, it randomly sends message everywhere if they hit that channel then kicked 🤷

[proven bear]

Anyone else seeing Codex failures right now?

Got:
backend-api/codex/responses
WebSocket fallback to HTTPS
503 Service Unavailable
upstream connect error / connection termination

Time: 2026-04-08 Asia/Taipei (UTC+8)
Request ID: 18985970-d1ad-4a68-ac65-64ec73f84c33

[frosty zealot]

Yeah it happens in here where they’re instantly purged. I just want their stuff to blow up

[eager forum]

Selected model is at capacity. Please try a different model.

I got this

[cyan wing]

yes same

[twilit bluff]

ahh

[twilit bluff]

so, usage reset incoming?

[proven bear]

Update: I’ve now hit it on my side too.

[lusty nimbus]

same

[hard tulip]

is openai down or something

whats going on

[eager forum]

reset aaaaagain?

[twilit bluff]

soon

[cedar skiff]

Ive only used for 4% of my daily if they reset i lose out.

[turbid axle]

people have been complaining for almost a week now. seems like this was intentional

basically you will need pro or more if you want to use codex for anything somewhat serious

[little magnet]

same

[turbid axle]

rumors are that gpt6 will release next week though, so maybe all this is temporary redirection of compute for that launch and things will settle again after

[orchid plume]

don't forget there's a new image model coming soon as well, certainly has been on A/B testing recently

[vital shoal]

Bruh i cant take this small 5h limit anymore

Ts feels like claude

[orchid plume]

getting it as well, was fine for a few minutes and now Copilot CLI (with BYOK pointing to ChatGPT sub) says: ✗ Our servers are currently overloaded. Please try again later.

[lucid mason]

"I see one new background agent called Goodall (explorer), but it is tagged as GPT-5.4 agent, not as GPT-5.4-Mini agent!?"

I did request gpt-5.4-mini explicitly for that agent.

What I can see on my side is only the spawn request I sent:

agent type: explorer
model: gpt-5.4-mini
What I cannot see is a confirmed backend echo showing the model actually attached after spawn. So if your UI showed a generic GPT-5.4 tag, that is likely one of these:

• the UI badge is generic for spawned agents
• the UI is showing the parent-family label, not the exact sub-model
• the platform silently normalized the label even though I requested the mini model

What do I make of this?

[short linden]

I'm sad with this error, "Selected model is at capacity. Please try a different model". Now let's do this... haha

[turbid axle]

it is honestly ridiculous. pro now feels like plus. and plus is completely useless

[cedar skiff]

I have never hit this case

what plan is it?

[turbid axle]

@cedar skiff who you talking to

[cedar skiff]

I'm talking about Selected model is at capacity.

[vital shoal]

bruh, is openai gonna revert it back
Like, the smartest thing they can do is to just remove the 5h limit lol

[turbid axle]

its been like this for a week now. I doubt it. My hope is they are diverting for the new model release, and things will settle over the coming weeks again.
As is, I am very close to canceling my pro account, 200 bucks for this is ridiculous.

[bright swift]

Please just ban free accounts and openclaw

[turbid axle]

Problem is, I'm a total oai fanboy. Moving to Gemini or Claude would hurt my soul

[short linden]

yep, they still have limit the hardware capacity on their server 🙁

[turbid axle]

They killed sora for their new model. I have a feeling these current codex issues are related

[hard drum]

https://github.com/openai/codex/issues/17105 i opened this because it is anooying.

Hopefully get some input. I locally tested it.

[lucid mason]

Also, one of the mini agents already failed to start because mini was at capacity, so the current visible roster is not fully reliable anyway.
What does this mean to tell me (Plus user)?

How is "mini at capacity"?

[bright swift]

openclaw slop using up capacity

[short linden]

Nope, it's not limit in your account. It 's limit on their hardware server..... Because there have many spamming free account and capacity went full.

[hard drum]

hopefully oai does what the competitor did nd ban OpenClaw

they're nothing but trouble

[bright swift]

i fear that oai will double down on it with their "single app for chatgpt, codex and atlas" idea

[lucid mason]

Thanks for the explanation. 👍

[kind jay]

Odds are less than 30% so unlikely

Bruh, jumped by 20% overnight

[turbid axle]

this polymarket talk?

[kind jay]

lol

[lucid mason]

When I ask Codex to spawn a sub-agent using GPT-5.4-Mini with Low reasoning then the agent chats in VS Code still list the same model and reasoning that the main chat used. I can then manually change them, but would very much prefer if these were spawned correctly from the start. Is anything known about this limitation?

[plucky halo]

5hour limit reached on pro plan 🙁

[turbid axle]

impressive. you are not using 5.4 on medium are you?

[plucky halo]

High atm

[turbid axle]

pretty sure you can only use 5.4-mini atm

joking ofc, but usage rates are nuts rn

[sonic wasp]

https://x.com/thsottiaux/status/2041655710346572085?s=20

did this happen? or was this selective to only a few plans? (i'm on business and it didn't reset for me)

[turbid axle]

tbh medium is plenty for most things, the cost/smarts benefit of high is pretty bad

[sonic wasp]

i'm thankful for the resets so far but just wondering if we did get a reset across the board.

[plucky halo]

Agreed - but needed high for these tasks

[turbid axle]

my teenage self needed high for every task too

[plucky halo]

Like I said, not on high all of the time but for the tasks today, I needed it on high 🙂

[faint frost]

I started using a skill a few days ago, and since then, my codex 5h limit goes away in a few prompts. The skill is basically a reference document for a big codebase where it points the agent to different folders and files depending the task so the agent doesn't have to parse the whole codebase. I expected to use less tokens with that, but without the skill, the usage goes down slowly like before. But with the skill, it takes only a few prompts to use the whole 5hours limit. Any similar experience?

[boreal holly]

Images do not burn that much quota. I've observed ~2k input tokens per screenshot.

[deft gyro]

Oh only 2k

ntb

[plucky halo]

Absolutely burns through tokens currently - but built a pretty comprehensive vulnerability scanner

[lucid mason]

So if your question is “only first use or every time,” the most accurate answer is: officially, full skill instructions are loaded on activation, not upfront; beyond that, reuse/caching within a session is not clearly documented, so do not assume zero additional input cost on every later use.
This can be problematic. If the full skills instructions are parsed every single time a skill is used within an ongoing session then that can become expensive. I expect(ed) this not to be the case, though, with skills only being parsed once to load them into the context. But if you observe much higher token usage then maybe that is not the case!? 🤷‍♂️

• Not every skill pays its full instruction cost at session start.
• The main full-skill input cost happens when the skill is first activated.
• For later reuse in the same live session, it may be cheaper if the relevant skill content is already in active context, but OpenAI does not currently document that as a guarantee.
• For a new session/thread, assume the process starts over: metadata first, full SKILL.md only if the skill is used again.

[deft gyro]

how much yall betting that oai is consolidating compute to compete with mythos

[faint frost]

That's what I thought as well. But I am not sure if the skill use more tokens because it is parsed in full OR because of the skill, the agent can implement way more in a short time and that's why I observe the limitation of usage compared to before

[frosty zealot]

Oh lawdy

[orchid plume]

yeah 😂

[frosty zealot]

Can’t wait to be like ok spud review my code base and it be like what in the world is this

[drifting granite]

5.5: "You let 5.4 do WHAT?"

[worn harness]

praying for something like 5.4 + 10-15 IQ buff + sensible limits on 100 plan + frontend buff

[kind jay]

I love the massive jump between the public benchmarks and the vendor reported benchmarks

[turbid axle]

uhm

oai be like 'those are rookie numbers' ?

im afraid

guess people are shaking the model tree again. can't wait to see what falls out

[boreal holly]

Codex places the header:

---
name: blah-blah
description: Do stuff more better
---

Inside the agent's context at the beginning of chat. If anything inside that header changes:

---
name: blah-blah
description: Do stuff more better v2
---

The agent gets that update inserted at the beginning of the next turn. Also I think if you do the $blah-blah explicit skill invocation they insert the whole SKILL.md into the prompt. Idk, they change so much between version updates that might not be the case anymore.

[supple perch]

@boreal holly I find that 5.4 mini on xhigh does as good as 5.4 high, I be it a bit slower but its cheeper.

You need to do 2+ waves of planning but it does work

[boreal holly]

5.4 mini high is 🤌 very nice. 4 out of 5 agents I use are mini and they do really well in most situations

[real vector]

Hi all. I am currently using Claude Code and I am looking to move to Codex. I was curious, what are the session limitations? Currently with Claude Code is extremely annoying, puting a stopper once every few hours for 3 - 5 h at a time.

Much appreciated! 🖖

[orchid plume]

session limitations? Do you mean the 5 hour quota, do you mean how many actual sessions you can have in some kind of timeframe? Or something else?

[real vector]

Yes, with Claude Code I get a session limitation. I was curious how is it in the case of Codex.

[orchid plume]

I still don't know what you mean by session limitation, so I'm going to assume how many sessions you can have in a month. If so, no such limit exists. it's primarily a 5 hour quota and a weekly quota, both of which are based on tokens

[real vector]

Aha. How many tokens I get per week on a paid plan?

[orchid plume]

the number of tokens isn't disclosed, but the limits, on pro anyway, are pretty generous

I'd argue that you get more out of Codex than you can on Claude

a lot more

[plucky halo]

never enouuuuuuugh. neveeeer neveeer

[orchid plume]

and don't forget the occasional manual resets that Tibo likes to do

[real vector]

So, how many tokens I get? WIth Claud Code I get roughly 4.7 million tokens a month, and a daily quota which is reached pretty fast, even with my bes tbehavior and watching closely context.

I do not see anywhere in the Codex plans any description in details in this regards unfortunately.

[deft gyro]

tokens of what

we don’t have access to what drains tokens

[orchid plume]

again, token limits aren't disclosed. All I can suggest you do is try Plus, and then consider that Pro is around 6 times the limit of Plus, so you can get an idea on how much you can do

[deft gyro]

i will say that I get about 300 dollars of inference per month on my codex plan which costs 20 dollars

anthropic models are much more expensive than oai models

and oai has a better small model

pro is 6x plus

so

300*6=1800 dollars of inference for 200 dollars

[plucky halo]

https://help.openai.com/en/articles/20001106-codex-rate-card

https://help.openai.com/en/articles/4936856-what-are-tokens-and-how-to-count-them

[real vector]

OK, I see. Well, thanks for the explanation. It is a pity number of tokens are not disclosed to see a price per point comparison. I'll investigate some more, try to work with it for free and see where it gets limited, and see where I end up. 🖖

[deft gyro]

the oai quota is very generous

[orchid plume]

you also aren't limited to what harness you can use with your subscription, which is nice, so you can use codex cli if you want, or opencode, or copilot cli with a local server to translate the api calls, etc.

[real vector]

Well... pretty expensive... I think it is way over my budget unfortunatelly :\

[plucky halo]

"By cheap pay twice"

I wouldn't run the risk with DeepSeek

[real vector]

Anyway, much appreciated!

[real vector]

There is no risk for light tasks and I am already using it with good results. Very critical parts I built it with Claude Code, and onyl recently used Codex where Claude failed hard.

[deft gyro]

this is not how it works

[plucky halo]

There's never any risk until there is :p

[real vector]

For me at this point in time is a question of what I can use for which task, and optimize as much as I can on clear costs.

[real vector]

Writing return functions for a simple calculator is no risk, I can see the code 🙂

I think is too much bias against chinese produces Ai or other non western based entities, and it is a pitty. But this is another topic, very political and no need to delve into it.

¯_(ツ)_/¯

[sand shuttle]

did they reset limits?

[real vector]

I do appreciate the better explanations for Codex, it is a lot clearer now. Thank you people 🖖

[turbid mortar]

For VS code in config.toml, what adjustments you need to change so codex asks before it makes any file change or run a command? I have tried different combinations and the codex extension seems to ignore the settings.

[orchid plume]

yeah last night, reached 3 million users so Tibo pressed the button to celebrate. If we reach 4 million he'll do it again, and every million until 10

[sand shuttle]

nice

why am i getting an erorr

no client found

[olive tangle]

is there a way to see it live ?

[olive tangle]

i was on claude 20$ subscription 3months ago, switched to codex, and codex > claude.

[lucid mason]

Weekly limit was refreshed today. 😄

I was lazy and did all work in VS Code + Codex today that could have been split with the ChatGPT (web) app (which doesn't use Codex tokens).

[orchid plume]

you mean your usage limits?

[solemn acorn]

https://hascodexratelimitreset.today/

[olive tangle]

no the live user

so if we are closer to 4M i can spam 👀

[orchid plume]

only OpenAI staff can see this I imagine 😛

[olive tangle]

but the 3million users is legit for the reset ?

[orchid plume]

https://x.com/thsottiaux/status/2041655710346572085

[lucid mason]

Did that reset happen today? The day my weekly limit reset anyway? 😄

[glacial shadow]

Please complain even with free stuff. wild

cant make anyone happy

[lucid mason]

My payment was also due today. 😛

And obviously I laughed about it, hence the :-D.

[lucid mason]

Unfortunately the whole idea about using sub-agents with lower models and reasoning for simple extraction work to save tokens doesn't work. But they are neither spawned nor used reliably and when they are then they use the very same model and reasoning as the main chat. I tried this in VS Code and the Codex app. 🤷‍♂️

[plucky halo]

Nope still absolutely burns through tokens. On Pro plan and one prompt (5.4 medium) with 3 sub agents has just eaten 9% of my 5-hourly limit and 3% of my weekly limit.

[glacial shadow]

Im not seeing that here

How many LOC u have?

typing for 1 hour now

@lean lark Do you need rescueing?

[lean lark]

That's true if you let Codex orchestrate. But if you use the model that @boreal holly uses, where the sub-agent is spawned with a CLI executed by the manager, then you have full control over the sub-agent props.

If you get the default Codex model to execute CLI commands, you can use the Agents SDK (or any SDK) and completely control the entire process of hand-offs, sub-management, etc.

[plucky halo]

To be fair, it's a massive repo and the prompt is about mapping it out for pen testing so it was sliiiightly disingenuous, but still

[boreal holly]

Yes, my agents are peers. Idk how they do sub-agents, but if it's anything like how they do unified_exec=true then the main agent is polling the sub agents to see what they're up to. That can probably be a waste of tokens, especially if the main agent is eating all the tool calls the sub agents are doing

[plucky halo]

https://github.com/StuartAJ/Sweeper

The first two are the biggest token guzzlers

[lean lark]

I believe the primary is waiting for events from the secondaries.

[boreal holly]

Do you just have one agent doing the reconnaissance or do you fan out the first 2 steps?

[plucky halo]

Fan out. 3 for pre-recon and 5 for recon

[boreal holly]

Nice! That does look pretty comprehensive. I could see it eating a lotta tokens

[plucky halo]

Yeah, it gobbles them up quickly but I've found it worth it. Tend to run it towards the end of the weekly limits but they reset early this week and ruined my schedule haha

[lean lark]

Stu - good on you for your protective licensing.

[glacial shadow]

So we are expecting coding agents that cost us 200 little dollars to digest 100k lines of code, make sense of it all in minutes, and cost us 1% of usage? not going to happen folks. that would take a development team months to get up to speed on let alone run bugcheckjing on it. have to be realistic and understanding on what ai is currently able to do... 200$ for what we get is a steal. even the 20$ plan is a steal. 200$ is what, 6 us dollars per day? an icecream costs 5$ .

do not type already, theres no way you read that that fast. slow down

digest what i said

[plucky halo]

Are we? Sounds like you're inferring and that's not what I've (intentionally) implied

It doesn't take very long to read 3 lines haha

[glacial shadow]

you typed it yourself and here you are still typing too soon because you didnt read my full reply

read slower, and really think about the words i typed!

[plucky halo]

I'm not disagreeing - absolutely, we get a lot more than we pay for.

[lean lark]

an icecream costs 5$
/me fondly remembers asking the local ice cream store why the cost went from $0.23 to $0.25.

[plucky halo]

But when I read your comment - regardless of the speed - I got the impression I was the 'we' you were talking about

And icecreams don't cost $5 here

Closer to £2

£3 if you want a flake

[glacial shadow]

same, same lol

gumballs going from a nickle to a quarter was a fraud!

devastation

[lean lark]

Or a penny to a nickel!!! Blast those capitalists! 😆

Let them all be replaced by AI!

[glacial shadow]

the current state of AI coding is unreal. People are already taking this for granted. ask 5.4 to estimate the cost and time a team of 2 or 4 devs would take to produce what we are producing, whatever it is you are coding, ask, it will blow you away

[plucky halo]

It would be nice if productivity increased without the expectation of increased workload. Instead it's do more and do more faster

[glacial shadow]

i really do think AI will help a lot on this front, i dont worry about it anymore

the right people are behind these projects here in the usa

future is bright

[lean lark]

I agree with cesium - my daily work pattern has radically changed just in the last few months. I'm doing things differently and producing differently. No newcomer is gonna produce what I do with vibing. The better we understand the tech the more we can do with it.

[glacial shadow]

I am very eager also for oai to implement a remote control feature for codex...anthropic has this, and orchestrating coding, bugchecking, ALL OF GIT, while hiking, or jogging, talking to it through voice, is completely unreal and powerful

i created an iphone app start to finish while walking at the park. i loaded the newly coded app while still at the park, and i pushed it to all remotes while still at the park

weve dreamed of this and its now here

[lean lark]

All that said, we're in this mode now where new techniques are pulling us in different directions. What to do about context, RAG, agent management ... these areas are evolving but I feel like we in the field are highly subject to the trend of the day which is changing too rapidly. I've been holding off on chasing trends, I'm OK with being leading-edge, not bleeding-edge. We gotta pick n choose our battles, spending precious time only on what seems most likely to be most productive at the moments.

[glacial shadow]

rocket fuel

we have one huge thing going, i think the right people are leading these projects

[lean lark]

What @glacial shadow describes is a sort of brass ring right now. There's a big disconnect between Codex and everything else. We can't start something in Codex and continue elsewhere - the "app" doesn't cut it. If we want the freedoms that cesium describes, we need to DIY with the Codex SDK or the Agents SDK plus connectors/MCP. I'm sure OpenAI is working in this direction. I'm OK to wait, but I feel like hands are tied right now and it gets frustrating. I'm anxious for more of what they've been offering.
https://developers.openai.com/codex/sdk
https://github.com/openai/openai-agents-js - https://github.com/openai/openai-agents-python

[glacial shadow]

codex cli has been phenomenal here too, runs anywhere

[boreal holly]

I think Agents SDK is really only useful in a CI/CD pipeline. They provide a much more powerful codex app-server generate-ts and you can build a frontend of any kind with it. I went a different route, adding github.com/openai/codex as a submodule, checked out to the installed codex --version , and link my stuff directly to their app server code. But in any case I can vouch for the app server being the absolute best way to build a custom thing quickly

It took me ~2 days to build a MVP frontend for the app server, and minor tuning and adding features over a few weeks. Meanwhile the app-server is headless so if I wanna update the frontend while agents are working it doesn't interrupt their work. It's something you can build while they're already doing their thing which is just 🤯

[lean lark]

Having forgotten the Agents SDK existed (amongst the plethora of tools we have available these days), I went back last night to review the current docs. It does support browser usage, streaming, human feedback, voice, images, and much more.

I was nearly salivating with thoughts about what can be done ... and then nearly crying as I thought about the lack of time. That happens a lot these days. 😢

[rain kettle]

I don't know why I just openned this chat. Maybe I wanted to complain about the Codex's stinginess (I already used 50% of weekly limit), but everyone already has done that here

[worthy furnace]

Why is my usage limit in the CLI horrible?

My 5 hour goes down in at most 40 minutes

Definetly not normal

[full current]

Anyone getting upstream connect error or disconnect/reset before headers. retried and the latest reset reason: remote connection failure, transport failure reason: delayed connect error: Connection refused ?

[worthy furnace]

Yeah I was

It meant that the usage limit was ending soon for me

[full current]

Hmm have over 90% remaining

[plucky halo]

Massive growth leads to massive losses (in the current pricing structure). As more and more people are using the service (whether it's Codex or Claude Code), the more the companies are 'losing'.
I think what we're feeling now are the companies adjusting to ensuring the increased usage doesn't cause outages. And yes, before anyone says it, we are seeing "Model capacity reached" messages but if you input your prompt straight after that, you still get the output you wanted

With mythos and spud, I'm interested to see how the pricing/capacity reacts. Especially since the hype around them are that they are wildly powerful

[glacial shadow]

Get Pro

[plucky halo]

People experience the same with pro

[glacial shadow]

No man, thats spam and disinformation, it is so hard to run out of pro usage

complainers arent builders apparentlky

[plucky halo]

What?

[cedar skiff]

You can burn through it if you want to, its easy. But i also don't have any confusion about usage, if feels about right

[orchid plume]

personal take, but I tend to find that Codex CLI appears to consume the context window much faster when compared to some other third party harnesses out there. I suspect that's part of the reason why usage is appearing to drain faster, as it's consuming input tokens quicker as a result

[worn harness]

It's just 5.4 eats tokens, 5.3 lower token consumption + 2x promo made it seem a lot bigger.

[orchid plume]

yeah the 2x promo ending definitely made it more noticeable as well

I'm using 5.4 via Copilot CLI using BYOK, and it seems to be draining my sub's quota more slowly since then

again, personal take, other users might have different experiences

[worn harness]

I have zero trust in MS doing anything correct especially Copilot

[orchid plume]

each to their own 😂

[glacial shadow]

Love using codex on xhigh

code being produced here, in totality, barely any bugs, fast, couldnt be happier

the cybersecurity considerations are super, greatest technology ive ever used

very thankful to the usage limit im given!

[jovial prawn]

Amem

[solar sleet]

which codex is best for large c++ projects? 5.4 or mini or 5.3-codex and what "reason-level"?

[boreal holly]

5.4 medium and 5.4-mini high

[cedar skiff]

mini high is a really nice model

I didnt use for the first while, then when i tried it does plenty of work well.

[glacial shadow]

i use maximum overdrive allthe time, xhigh

[dawn seal]

The key to not running out of usage limit is not making your code molithic

go ahead and ask your agent
"read my entire codebase and tell me what parts are monolithic bad codes/design? that forces a coding agent to must read the entire codebase and inefficiently work?

[cedar skiff]

code review changed in codex app, doesnt start with a new context anymore

[frosty zealot]

I wanna start seeing more medical breakthroughs driven by big models like Mythos & SPUD or something rather than like 'FIXED A BUG'

[upbeat maple]

why dont you do it

[frosty zealot]

Great question

I didnt fact check it but apparently some dude synthesized. cure for his dogs cancer that is now at harvard or something for trial

[cedar skiff]

Im sure all of the experts in the different fields have their purpose built models coming and currently use the ones avaiable.

We just dont see it because they dont hang out in the coding channels

[glacial shadow]

some of it is literally book + whiteboard/pen paper, and thinking very hard for a long time, completeyl agree with your comment

[lime inlet]

There was this X message yesterday by @sama that "To celebrate 3 million weekly codex users, we are resetting usage limits". I swear this has not manifested on my end of the tubes. I am using codex with a ChatGPT Business (formerly: Teams) account. Did anyone receive a usage limit reset yesterday?

[true bane]

Has anyone else worked a lot with both codex and cc and strongly prefer codex like me, which seems to be opposite a lot of anecdotal evidence? Not sure what it could be, it interprets my agents.md file better? Or its better suited to the repo I've been working with for a year plus (TS monorepo with server/client/shared) idk

[ruby lynx]

does anyone know when 5.4 codex comes out?

[glacial shadow]

I find they are very similar and greatly appreciate the option of having affordable access to both. each could easily be 1000-1500 per month. easy.

[true bane]

very much agree they are both very good. Just like codex more though, just a better product, better communicator, better sr dev I fire every 10 minutes

[cedar skiff]

They removed guardian approvals?

I was loving them

[glacial shadow]

thats a nice touch

[nocturne folio]

chatgpt 2 coming soon

[glacial shadow]

Sample a random person in a city, you will find most "Oh yeah I heard about ai, used it once, gave a really long answer, havent used it since"

[cedar skiff]

depends on the demographic. every single student at school or uni uses it all the time.

[lime inlet]

In code (cli), how can I type in a newline in my prompt? When I type newline it ofc sumbits the prompt. Shift-enter also submits!

I think it must be possible to type a newline though, because when I paste text cotnaining newlines, these appear correctly. But how to type them?

[next ore]

anyone know the what the difference in usage is between medium/high/xhigh? and also the overall effectiveness?

[cedar skiff]

usage has a big jump on high -> xhigh with much less than proportional benefit.

There are graphs around that show it.

[solemn acorn]

high seems to be the most effective, medium is cheaper when it works though

iirc xhigh actually performs worse than high in a lot of situations

[deft gyro]

wut

vro has not referred to the graph

the graph should be pinned

[solemn acorn]

where's the graph?

[deft gyro]

xhigh is better than high

but costs twice as much

[solemn acorn]

does that show small-medium-high-xhigh?

[deft gyro]

yes

on the gpt 5.4 graph

cost is in order from left to right

you can see

medium and high have about the same accuracy

but medium costss a lot less

for having NEARLY the same accuracy

so

basically

either use gpt 5.4 xhigh, or medium

or use gpt 5.4 mini on high

[solemn acorn]

ah I was thinking of 5.2's high mode

5.2 in general was kinda weird when it came to benchmarking tool calls lol

[deft gyro]

yes

5.2 is very specifically good at long form tasks though

it is not very intelligent

but it sure does take its time to gather info

which can be useful

[solemn acorn]

I hate that openai refuses to zero their y-axis

[deft gyro]

its so that they can pretend there is large improvement when there isnt

[solemn acorn]

exaggerates what would be negligible differences a lot more

I mean that is clearly a large improvement, but they should have used 0% as the baseline

yeah see they did do it here

granted 5.2 was much closer to 0% but still

[lean lark]

There won't be a 5.4-codex. This time they went in the opposite direction. Before, they tuned the general-purpose 5.3 model for coding, and gave us 5.3-codex. This time they tuned the model update to also be good with code, and published it as the general-purpose 5.4. So we are using the same 5.4 in ChatGPT and Codex. It's elegant. I wouldn't recommend this as a permanent go-to solution but for now I think it was a great move.

[solemn acorn]

it makes sense considering the demands of chatgpt are getting much more similar to codex models

[deft gyro]

there will prolly be 5.5 soon anyways

lowkirk

openai needs to cook something to pretend to compete with claude mythos

[solemn acorn]

having separately RLed models that are both mostly running bash commands doesn't make a whole lot of sense

[solemn acorn]

in theory they have whatever the big model that was used to distill 5.2/5.3/5.4

I don't see much of a point in openai competing with something that is so stupidly expensive and closed off though

[deft gyro]

Well

[solemn acorn]

mythos is I think 10x more expensive per input/output token than 5.4 which is wild

[deft gyro]

Gpt 5.4 pro I assume is their fat model

[deft gyro]

Pricing is not released yet

And leaks about these things are notoriously inaccurate

No but the point is oai needs a new massive model to generate QA pairs to train their models

[solemn acorn]

they released pricing for glasswing

which is presumably somewhat subsidized so it is probably lower than what the released model will cost

Claude Mythos Preview will be available to participants of Project Glasswing at $25/$125 per million input/output tokens accessible via the Claude API, Amazon Bedrock, Google Cloud’s Vertex AI, and Microsoft Foundry. Anthropic has committed $100M in model usage credits to cover Project Glasswing and additional participants throughout the research preview.

[deft gyro]

Or maybe not subsidized because they are giving away 100 million dollars

Or maybe costs more than it will because they are using an unoptimized unquantized version

[solemn acorn]

it's a really big model so I wouldn't be surprised if the cost of inference was about that high without anthropic's usual API tax

which means the release version would be more expensive to subsidize subscriptions

[deft gyro]

I cant wait till anthropic distills mythos

[solemn acorn]

they're distilling it into an opus model iirc

[deft gyro]

Or maybe vera rubin will make it cheap

[solemn acorn]

won't be as capable though

[deft gyro]

Ofc

[solemn acorn]

my guess is mythos is never coming to subscriptions unfortunately

they're gonna bill the crap out of governments and corporations to use it though

[frosty zealot]

Have I gone redarded, does codex not respect project specific config tomls?

[amber scaffold]

if you use codex you have tonight to use up your limits :)

[frosty zealot]

nvm forgot to restart my app-server after i made changes

[deft gyro]

wait

lowk

XHIGH ON FAST GOGOGOGO

[sleek tusk]

Does anyone have any additional information on the coming cost changes coming for plus users? Will we still have a 20$ plan or will we be forced on a credit based pay as you go model?

[deft gyro]

gpt 5.4 MINI I NEVER HEARD OF IT

[sleek tusk]

I’m finding it very hard to figure out the prices

[deft gyro]

source for upcoming cost changes for plus users

[sleek tusk]

Haha, Reddit doomscrolling and the business plan changes to usage .. let me find a link

[deft gyro]

the business plan changing has nothing to do with the plus plan

the business plan

is made

for businesses

[sleek tusk]

But they state “until we move you to the new token based pricing for codex” in the rate card article

https://help.openai.com/en/articles/20001106-codex-rate-card

Inside the grey legacy rate card box

It looks like a jump from 34 credits in 5.4 to 62.5?

Or 6250 depending on how you American you are in reading numbers

[deft gyro]

thats

for

credits

and the legacy rate card

is for

prompts

the new rate card

is for

megatokens

did u even read the article twin

[sleek tusk]

Buddy I don’t understand the article 🙁

I’ve just been used to a fixed price, fixed usage

If all of a sudden my price will vary, I will have to explain that to a boss somehow

And also, it tells you how to convert tokens to credits, but there is no mention of what one credit is in dollars, so the whole thing is just super vague

[glacial shadow]

No, their document said they are but not the cyber one

[solemn acorn]

mythos is the cyber one

[potent mason]

Yeah honestly no one will be using mythos to actually program. The concept is nice for cybersecurity for companies but there's few companies who will pay for it.

With Claude Code Max currently you can finish your usage in like 4-5 prompts, imagine a model 10x more expensive

[cedar skiff]

They will distill other less expensive models from it. I guess the idea is over time compute gets cheaper, and they find optimisations making bigger models cheaper to run.

[glacial shadow]

Yes, I know. They said they are making a public one. All of this information is freely available. This is in their documentation on their website. It’s public info.

[solemn acorn]

that'd be opus

[potent mason]

I thik it's $0.04 don't quote me on that though

[cedar skiff]

Last time i was on claude about 8 weeks ago it was like that it could do a tonne of work on max plan

opus is the previous big daddy model

[ruby lynx]

so should i keep using 5.3?

[glacial shadow]

I’m not worrying about this because the people behind codex and the models are genuinely good people

[lean lark]

No, 5.4 is right there, assuming you've got the current version.

[ruby lynx]

?

[frosty zealot]

Oh my sweet summer child

[lean lark]

Open codex and you should see the 5.4 model.

[frosty zealot]

I think he's just under the impression 5.3-Codex is better because it has the 'Codex' suffix, and 5.4 to him is just a general model to ask the air frying directions for a bag of crinkle cut fries

[glacial shadow]

Update codex

Get on 5.4 it’s great, you always want to be using the latest models so that you are implementing the latest technologies

[cedar skiff]

5.3 works better for me, it doesnt over engineer as much, it follows all the way to the concrete implementations and it doesnt make as many assumptions.

[frosty zealot]

im hoping i wake up tomorrow to Gpt-10^99

[sly laurel]

Anyone else unable to install on Windows Server 2022 Datacenter? This is an AWS Workspace. It just keeps looping saying that it can't install and sends me back to the microsoft store. I've tried running as Administrator but it does the same thing.

[frosty zealot]

The microsoft store is always the worst thing in world for me, it rarely works, and when it does it downloads so slow

try using winget

[glacial shadow]

Try installing from terminal , less moving parts

[ruby lynx]

mm

[ruby lynx]

yep i do

[ruby lynx]

ye

it makes me confused

[cedar skiff]

Just use what works for you

[ruby lynx]

people are saying that gpt 5.4 has later technologies which is true

but it’s a general ai

which means that it probably has a higher chance of hallucination

but

[frosty zealot]

To be clear, there will never, ever, be another 'Codex' model

[glacial shadow]

It’s better at reasoning and therefore coding

[ruby lynx]

5.3 has less advanced techniques but it focuses better

[ruby lynx]

rate 5.3 and 5.4 hallucination

1 being no hallucination

5 being lots

[cedar skiff]

5.4 makes more assumptions

[ruby lynx]

can be useful in some situations too

[glacial shadow]

I use 5.4 because I want the best reasoning , best coding, and I use it on high or xhigh, and sometimes /fast on

[ruby lynx]

what plan?

[glacial shadow]

It is great at coding, everything git, general ai,

[ruby lynx]

are you using?

[cedar skiff]

It's not useful when one more step into the code would remove the assumption

[ruby lynx]

mm

[glacial shadow]

I will definitely never ever ever listen to people telling me to use older models, to me that is Spam and intentionally bad information

[ruby lynx]

lowkey 5.3 not bad either right

[cedar skiff]

Just using brain dead ideas of newer is better in every case is naive at best.

[glacial shadow]

5.3 was great. 5.4 is better

[cedar skiff]

Use what wroks best for you

[ruby lynx]

mm

[glacial shadow]

They call it 5.4 because it’s better than 5.3

Lol

They use it, so I will use it

Simple

[hard drum]

When you don’t realise that 5.4 is intelligent, but also extremely lazy. Also absolutely horrid at frontend UI work.

Higher number ≠ better

an RTX 5060 is not better than say… a 4090

Your logic is silly, my friend

[frosty zealot]

OpenAI's benchmarks, end of discussion choose what you enjoy

[cedar skiff]

It doesnt prove that its better for every case.
There are cases and workflows where its not better. Use what works best for you.

reality isnt a benchmark.

[ruby lynx]

lol

5.3 vs 5.4 can really cause a debate