codex-discussions

[oak trellis]

it again compated after another 2 message ..

never happen

i asked a question and it used 30k tokens ..

something off

[crystal flint]

I just tested out the new limits. It used 13% of my 5 hour limit in 1 message on extra high (worked for 13 minutes)

[oak trellis]

yes its garbage now ..

they screwed up something on the token usage

1 message 123k token! never happen

so what happen is they screwed up something on token usage itself .. limits are the same but now 1-2 message it has no more cache .. might cache bug

100% a bug now

nowhere i have that issue not even in cc

lol ... it compacting again

[soft glade]

does anyone else have codex CLI keep changing permissions? like i have repeatedly set permissions to default and it repeatedly, seemingly at random, switches to some custom permission set. it sometimes is read-only and other times it just asks before every tool use. its insanely annoying

[oak trellis]

lol so it compacts read the file again uses straight 136k tokens again ..

wow i loop with 4 message through now 500k tokens

yah we got rug pulled by codex

not usuable anymore

same stuff in cc consumes 5k tokens max

even if

[potent mason]

Thats interesting is it trying to read like a huge file or something?

[oak trellis]

golang files

i mean its the same files i work on since 7 days

[potent mason]

Maybe its the tokenizer they messed up (since they changed usage from message to tokens iirc)

[oak trellis]

cc consumes the same files 6 files .. 3-5k tokens max .. modification when in the session 300 or something

[oak trellis]

its the cache iam in a compacting loop

it keeps reading the files then iam half in 136k tokens .. 5-6 files .. do something .. 75% .. and then its compacting ..

wow ..

token usage is insane

[fathom kite]

Recently my VSCode Codex diffs (in the Codex chat window and VSCode Codex Diff editor window) have lost their background colors (red for deleted lines, green for added lines), and instead just display the line numbers in red/green with an accompanying (always green) changebar. The source lines themselves use standard formatting. So it is very difficult to actually pick out the changes. I don't know if this is a VSCode or Codex problem, but it only started a few days ago and I'm sure both VSCode (1.115.0) and Codex (26.406.31014) extension have changed in that timeframe. Is there a VSCode/Codex setting that controls this, or is this a recognized problem that (hopefully) is already being worked on.

Does anyone else see this behavior or is it just me?

[oak trellis]

this is bad ..

yes it used 70% of my limits burn through my tokens like no other model

the 100 usd model will burn through it via the same rate

[high girder]

The plus sub does seem to burn through the 5 hour limit faster now, but the weekly feels fine

[oak trellis]

check the token usage there is a bug. it feels like the cache is disabled

[brave bramble]

It did feel bugged to me but I just figured the $20 sub was cut back for codex usage.

[potent mason]

Not gonna lie I think it might just be on your side (based on the error description I don’t think I’ve seen anyone else speaking about it today)

The 5 hour sessions being super short though is a problem everyone’s having

[oak trellis]

check the token usage

i literally monitored it

│  5h limit:             [█████░░░░░░░░░░░░░░░] 25% left (resets 08:34)           │
│  Weekly limit:         [██████████████████░░] 88% left (resets 03:34 on 17 Apr)```

5 prompts

unusable ..

trying to get chinese models now

[frosty zealot]

run the checker script and see whats eating your tokens

[potent mason]

I haven't had it compact a single time today (83% context left after 2 prompts)

[oak trellis]

where can find that ?

[oak trellis]

i was in a compact loop

[frosty zealot]

https://pastebin.com/bn7Zn1Hr
SINCE_DAYS=7 python3 scripts/analyze-codex-tokens.py

#codex-discussions message

comes out looking like ^

[oak trellis]

moment checking

[frosty zealot]

be curious to see what it says

[waxen bramble]

I just tested it and the adjustments to the limits are already working. The 5-hour limit was reached quickly... it's the first time I've hit the 5-hour limit on the Codex.

[oak trellis]

wow so it always burned throught tokens .. something is off .. Top sessions:

• Apr 3 — 175.8M tokens
• Apr 8 — 158.0M tokens
• Apr 3 — 12.3M tokens
• Apr 3 — 0.8M tokens

[frosty zealot]

Whats the blurb at the very bottom say?

[oak trellis]

- Context replay dominates output: input/output is 220.3x and cached-input/output is 199.0x across the whole report window.
- A small number of sessions dominate spend: the top 5 non-subagent sessions account for 346,814,695 tokens (100.0% of total usage).
- The heaviest static instruction payload observed was in `019d532a-8791-7223-afaf-86dc6a728168` (cantwin) at 19,197 characters, or about ~4,799 tokens before any repo/file context was added.```

[frosty zealot]

But between Apr 3 and Apr 8 the 5 hour limit is like instantaneous?

[oak trellis]

let me check

[frosty zealot]

Or wait

todays the 9th

Whats it say for today?

[oak trellis]

wait iam on berlin time (bali)

[frosty zealot]

oh

[oak trellis]

so its 8 for germany i guess ?!

let me verify moment ..

nah wait today is 10th

[frosty zealot]

lmao

VAS?!

[oak trellis]

got confused it cant find 9-10

lemme dig in

i think a session what compacting since 8th so yes

i used same session what compacting since 8th ..

[fathom kite]

I confirmed this is a change from 26.325.31654, which shows the expected diff formatting.

[oak trellis]

let me get all infos now

[frosty zealot]

• Context replay dominates output: input/output is 220.3x and cached-input/output is 199.0x across the whole report window.
  This is good but also inefficient

High cached-input ratio (199×) = ✅ efficient reuse (cheaper + faster than recomputing)
Means caching is working as intended

This ratio is very high, and usually signals inefficiency:

You’re sending way too much context per request
The model is doing a lot of “reading” vs “thinking”
Symptoms:
Higher latency
Higher cost
Diminishing returns on quality

according to chatgpt

[oak trellis]

something off because it didnt compact after 4 messages yesterday. Always was super smooth

will read the report moment

[frosty zealot]

it coudl very well be bugged too

you're like the 100th person today to say the 5hr limit got smoked

im on the $250 plan so im not sure

[oak trellis]

so fast it was scary ..

[frosty zealot]

i personally havent noticed

[cedar skiff]

well pro limit feels like an endless bucket today, i think its on double usage

[oak trellis]

they try to push everyone out from the 20 to the 100 i think.

[cedar skiff]

yeah seems like it

[frosty zealot]

realistically with the 10x rate increase, if you can afford it, even if its just during hte promo, its really worth it imo

[oak trellis]

it made no sense whats happen on my session 0 it feels like its a bug..

the 100?

[frosty zealot]

yeah

[vital hill]

accept the blessing bro

[oak trellis]

i would go there but not if its like the old 20

[frosty zealot]

im pretty sure they said its 10x usage limits on the $100 plan

till may 31st or something

[oak trellis]

then.. get rug pulled haha.. cant be there something stable its headache always changing models

[potent mason]

I'd be more worried about the bug that your codex has than the plan cost.

I'm on the $200 plan right now and it feels basically unlimited, I'm running two open claws, multiple projects all the time, etc.

And bearly getting close to 10% of weekly right now

[frosty zealot]

not in the great world of technology

[oak trellis]

yes that is the point .. would go with the 100 .. was waiting for that. Mention that they should create a 100 plan ..

[cedar skiff]

Yeah like wise, i think its double usage again

[frosty zealot]

@kind jay already confirmed double usage for the pro plan

[oak trellis]

4 files 2 searches 50k .. (new context window)

now exploring .. 6 files 100k tokens ..

8 files 111k i litreally told it please check the git changes if we have any issues on those changes.

yes that is totally garbage

how is it on your 200k plan @frosty zealot related to the context window

[frosty zealot]

What do you mean

[oak trellis]

sorry i mean 200 usd plan

[potent mason]

He means if you're having context issues probably

[last willow]

This is my same experience

[frosty zealot]

I'm in the same boat as the other $250 plan users, I feel unlimited right now

[oak trellis]

Do you have context window limitations ? Here i made 1 prompt running over 8 files from git changes ..lol

that is not normal

[frosty zealot]

How many MCP's/Skills do you have?

and how many AGENTS.md's

and how many LoC in them

[oak trellis]

same as yesterday lemme check 0 skills, 1 agents 136 lines, 1 rule 59 lines... max

[potent mason]

Oh by the way I wanted to know from someone who's tested this more than me.

Do plugins take up context or are they only called when you / them?

I put in a bunch of them in case they would ever be useful but I forgot to ask if they where context rotting my prompts

[oak trellis]

i have 0 mcp ..

[frosty zealot]

is this like a fresh prompt or days or what?

[oak trellis]

did a fresh one on target files (git changes)

[frosty zealot]

how big are the changes though like whats git diff --shortstat read

[oak trellis]

wont work that command, but it changed 7 files +51-54 .. not much at all

[frosty zealot]

are they uncommitted changes?

[oak trellis]

yes

[frosty zealot]

hmm.. should have worked

are you in a worktree and running the command in main?

[oak trellis]

in main

[frosty zealot]

if you have a dirty repo it should def print out the total LoC changed, weird

not sure

[oak trellis]

wow so cc has 1% usage on the same

[cedar skiff]

has to be a bug

report it

[frosty zealot]

Is it a visual bug or is it actually stopping you at 0%

[oak trellis]

it consumes literally almost real time after every greb

[frosty zealot]

But at 0% it's telling you to wait until reset?

[oak trellis]

yes

[frosty zealot]

Inchresting, yeah def sounds like a bug, like I said earlier you're like the 100th person I seen say that today

[oak trellis]

LOL so yes cc 1% i let it run on it on fresh window same question

lmao

let me check how much token cc used for the same ..

[frosty zealot]

But, theres always people complaining about usage like every day so I usually don't pay any heed to it

[oak trellis]

i never had issues even the last time they put limits in its fine, codex always was generious ..

everytime i complain like last time it was cc i think eric saw me coming here .. anthropic made an anouncement and they booted openclaw

i monitor my stuff extremly haha

cc seems balanced now since last time.

with 100 usd current week resets today on 53% ..

i suspect they drive out users from 20 usd to 100 usd .. then they adjust the limits from 100 usd back to the previous 20 usd limit from feb.

[frosty zealot]

All you can really do is make a bug report https://discord.com/channels/974519864045756446/1070006915414900886 or open an issue on the repo unfortunately

Or upgrade 😛

[oak trellis]

will do that now. And yes i would like to upgrade but not getting rekt again haha. Wanted the 100 usd plan for a while anyway

[frosty zealot]

I sorted the bug reports by codex and theres literally one post

so I think every one just reports the issues on the repo lol

[oak trellis]

no permission to post there

[frosty zealot]

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

[oak trellis]

thx

[oak trellis]

did get a refund of one of my accounts .. nice .. will then get the 100 usd plan for one month until the rugpull comes in haha

You have been granted a full refund. Your active subscription has been cancelled. The refund will show up in your bank statement within 10 days.

[cedar bear]

Could one of you briefly explain to me which plan makes the most sense from plus to Pro 200?

Um currently a plus User

And I didn't quite understand what changed at Plus now that there is one Pro 100

[pearl kestrel]

If you use GPT Codex a lot and you use GPT Pro model more than a very little, ChatGPT Pro $100 looks much better than 5 Plus accounts.

It makes the business accounts look quite poor value based on the cost of additional usage. If you spend $100 on Business + extra codex or GPT Pro you seem to get quite a bit less than you would on the ChatGPT Pro 5x account for the same $100. You get workspace/admin/privacy/team features, but less Codex or GPT Pro use for the same $.

[frosty zealot]

https://super.engineering/ the background on this website is sick

Anybody know what you would even call that I might have to itterate on it 😛

[pearl kestrel]

Where I live it's currently:

• ChatGPT Plus for $20 + VAT
• ChatGPT Pro x5 for $100 including VAT
• ChatGPT Pro x10 for $200 + VAT

That makes x5 really good value. I don't know if it's the same in other places.

[pearl sierra]

The limits are dropping crazy fast

[pearl kestrel]

That was for upgrading my plus account. For a new account, 5x is $118 and 10x is $200. Their pricing is deeply confusing.

[pearl sierra]

If my limits are reached, would the session still continue generating or would it stop generating mid?

[real blade]

how long can someone use codex in plus in 5 hrly limit?

[frosty zealot]

That is basically an impossible question to answer there are so many variables

[cedar skiff]

5 hr at most

[frosty zealot]

lmao

[real blade]

Fr

?

[cedar skiff]

after 5 hours it resets so thats the max

[oak trellis]

there is a mcp on claude what will get you a clone 🙂

so for token waste .. i found something user told me on reddit: handoff.md

you create under rule that agent should always handoff

and i found this what you think ? https://github.com/rohitg00/agentmemory

since i use handoff and codex 5.3 .. i did way more work used only 75k tokens .. and 8% on the daily .. but still sucx

[cedar skiff]

i always feel like those kind of things carry baggage that isnt needed for the current task

[oak trellis]

agree .. this time it helped a little bit

[kind jay]

I’m not sure

[frosty zealot]

Is it morning where you are?

1:30am here i go eep now

[oak trellis]

ok i think i found the bug .. compact instant draines immedalty your tokens ..

so never do compact open new session ..

before i was like 1 hour under 50% .. now within 3 min 43% down the new session and it also drains the 5h ..

[short linden]

because compact will pass some important detail to next context window.

[oak trellis]

but way more then yesterday way more ..

[orchid plume]

And probably because cached input tokens are ultimately cheaper

[oak trellis]

they changed something from yesterday today ...

and that could be ..

[short linden]

... I'm not surprise any problem, because I already have billing issued while using some credit....and contact support to refund later ( that's wrong auto-recharge the credit ) so just let's go, haha

[kind jay]

No, it’s night

[oak trellis]

35s · ↓ 19 tokens - dream usage

[kind jay]

Skill issue

[safe holly]

Why is codex now completely unusable? The 5h usage is draining like a leaky sieve

[plucky halo]

The 2x has made it LOVELY

[hidden glade]

didn't they literally just resetted the usage and gave 2x until may 31st? I don't understand the usage complaint, go try claude code and see your usage absolutely gone in a blink. the number of people complaining about usage is like 10x there

[plucky halo]

The 2x is only on the pro plans, to my understanding. So plus users won't be seeing the benefit

[white furnace]

anyone here tried the 100 dollar option, how is the 5x limit?

[safe holly]

Because usage is now half what it used to be?

[hidden glade]

ahhh okay that's unfortunate for them then, thx for info

[safe holly]

I paid for this product, the product has now materially changed

The absolute lack of clarity on ai billing is absolutely criminal

[oak trellis]

haha

[hidden glade]

trust me when i say things are worst in claude code land this past 2 weeks, hence why i am here now

[oak trellis]

yes lol

sam style

[oak trellis]

still ?

[white furnace]

tbh openai does what they want, they change models on the browser like new clothes often and codex is just an extra to the chatgpt sub so they can do that to codex

they gave free quota resets often too

[hidden glade]

yeahh like the past 3-4 days everyone's hitting their weekly usage for the $200 plan by like midweek and also the quality of outputs are noticeable worse, they probably have a really high load on their GPUs for other "projects", although some people are saying today is a bit better. I haven't tried, still waiting for my weekly limit to reset on claude code

[oak trellis]

they get free bail out too soon or later

hook up the world .. abuse vc money .. get tax money bail out

he try to get bank status ..

[safe holly]

But im not paying for claude? This is nothign to do with Claude

[hidden glade]

i was chatting with the other dude

[oak trellis]

thats crazy

[safe holly]

Its a bit of a wild west atm

[kind jay]

“Usage limit, usage limit”, break out the wallet and get that pro subscription

[safe holly]

And what when they suddenly kneecap the pro sub limits?

[white furnace]

true, but tbh the whole ai thing is wild west and quite jumpy, they had issues with ai changing personality often so they couldnt control that and nerfed the models a lot, looks like they cant control much of the stuff lol

[oak trellis]

soon or later first KPI is offload 20 usd into 100 usd

[kind jay]

Buy another pro account

[hidden glade]

i will say i tried plus plan for a bit and i burned through it pretty quick but pro plan feels unlimited or it's really hard to bring a few percentage down for normal use

[white furnace]

ngl it was draining yesterday a lot more the 5h cap, not sure what happened but looks like its draining much less now

im on plus

[kind jay]

Yeah but what happens when that isn’t true anymore?

[hidden glade]

cry in corner or load up gemma 4 or smth not sure

[safe holly]

Exactly, so keep throwing more money and get less

[kind jay]

Or learn to code, idk

[oak trellis]

never could imagine that some evil corp now control my codding .. i never could go back line by line coding

[plucky halo]

I paid £200 and in the past 30 days hav euse around £3700 worth of tokens. I'm pretty happy with the pricing

[oak trellis]

i have 0 cc issues .. damn always cc <-> codex eather one the black sheep .. thats why i like competiton

healthy

[cedar skiff]

meta just got a half decent model as well, not sure they are going to try and move forward with coding though

[safe holly]

Can’t imagine they’ll focus on code. I guess they’ll want to use it to start processing the vast quantities of user data they hold as well as generating content for their dying platforms 😂

[lucid mason]

Wasn't 1M the context size limit for Codex in VS Code? Seems like I can set arbitrary number, although I am not sure if that's really used or not.

[oak trellis]

there was a 1mill ..??

[hidden glade]

I thought 1M is available to codex cli as long as you set the flag / limits for it? but I haven't bother changing the default

[oak trellis]

aaaah so the old had 1 mill ?? in codex

[cedar skiff]

5.4 has 1 mill i think does 5.3? They are terrible apparently

[lucid mason]

In Codex and the API, GPT‑5.4 is the first general-purpose model we’ve released with native, state-of-the-art computer-use capabilities, enabling agents to operate computers and carry out complex workflows across applications. It supports up to 1M tokens of context, allowing agents to plan, execute, and verify tasks across long horizons.

[oak trellis]

thats crazy ..

so they swapped it out of plus up to 100/200 .. and i get hit with non stop compact

i guess 100 usd is the way now ?!

[hidden glade]

I'd be careful with context rot though, even with claude opus that has wayy better need in haystack benchmark scores, i still try to keep my usage 300k-500k max, don't want to risk the quality of the outputs so for codex ive been sticking to the default 200k-ish

[lucid mason]

In VS Code you can set the context size via config.toml, but to my surprise it - at least superficially - accepts values larger than 1M. Don't get your hopes up too much, once you close in to 500K the chat starts misbehaving and once you go towards 700K it can become senile/lobotomized.

[oak trellis]

500k is fine

so in config.toml we can set the limits ..

also for codex in cli ?

[lucid mason]

This is my current project when I start a new chat prompt and tell it to read in all necessary instruction files.

I don't use CLI (and don't even know how), so I can only speak for VS Code.

[oak trellis]

i use the store app

[hidden glade]

not sure i did not try but i thought i saw docs / comments about it. But someone correct me if i am wrong

[oak trellis]

lemme figured that out bc the compact killing it for me

[cedar skiff]

#codex-discussions message

[oak trellis]

you figured ?

[cedar skiff]

nah still havent tried

[lucid mason]

Question remains, why can we set sizes larger than 1M and do these actually exist in reality then?

[short linden]

1M window context still in beta test... So use at your own risk. It could less performance than default context.

If you want to try, just start with 500K context.

[lucid mason]

Yes, but the question was about being able to set sizes much larger than 1M. Why is that and is that real?

[hidden glade]

extra context is sent to the shadow realm or gpu servers on the darkside of the moon

[cedar skiff]

its about training window size

[short linden]

more context = more scan project , but it might incorrect results when you edit the whole project instead of split by milestone.

[lucid mason]

What does this mean?

[cedar skiff]

when they feed the model during training the size of the input in tokens is the possible size of the context you can use.

[kind jay]

It doesn't make sense to use context so large

[cedar skiff]

But there is degradation at higher context, so its not like you can just train on 10 million and have it just work

[lucid mason]

So context size can arbitrarily set above the 1M limit and that is not surprising behavior? And the context size is real then, not just a GUI confusion?

[short linden]

So in default context, agent will automatics split into milestone that suit with current context size, but more size doesn't meant you can finish the project in one promt, it's just make AI learn your relationship of your project.

[cedar skiff]

it cant use above what it was trained on

I mean there is some ways to do it, like some context extending methods

I cant remember what its called, but they have to support it.

[white furnace]

does steering prompts make the model focus on the new task and forget the old task?

it feels like its forgetting the old task for some reason and focuses on new task completely.

[lucid mason]

So this is a GUI issue?

[cedar skiff]

where do you set it in gui?

[lucid mason]

config.toml

[cedar skiff]

that isnt gui

its just a config file

[lucid mason]

And the GUI then displays a size value corresponding to what is set in config.toml.

[cedar skiff]

you can set it to abcd if you like

[oak trellis]

eric you a good one .. just wanna tell you that!

🫡

[cedar skiff]

I'm not sure what i am looking at

[lucid mason]

The output of the /status command.

In VS Code.

[cedar skiff]

Ok you mean how it says 5273K?

[lucid mason]

Yes.

[cedar skiff]

I see what you mean now

[lucid mason]

And my original question was: Wasn't 1M the context size limit for Codex in VS Code? Seems like I can set arbitrary number, although I am not sure if that's really used or not.

[cedar skiff]

It wont use over the 1 million - assuming 1 million is it's limit.

[wicked briar]

what street

[lucid mason]

Both the VS Code tooltip and /status appear to read the same runtime session field. Your config sets model_context_window = 5550000 in config.toml, and your current session started with model_context_window: 5272500 in the session log. That is the exact 5,273K shown in the GUI, so the UI is definitely using it.

Inference from the local logs plus the official docs: this is a real client-side budget value that Codex uses for status/accounting, but it does not change the backend model’s actual hard context limit. For gpt-5.4, the official model page still says 1,050,000 tokens: GPT-5.4. The tooltip text “Codex automatically compacts its context” is generic; it does not mean the model truly has 5.273M usable context.

[forest crypt]

Has anyone noticed if with the recent update of the OpenAI accounts, the rate limits per 5h has decreased? Mine appear to be decreasing a lot faster

[worn harness]

they already said it is so

halved pretty much

allthough they seem to be nice in the sense that if you have a long running task they will let it finish instead of crashing it during the middle when the usage limit hits (unlike a certain orange competitor)

[spiral gorge]

plus sub now has the limits of free version and 100usd sub is the new plus

how to upsale 100 million customers

[inland bramble]

That is happening from one month ago: https://github.com/openai/codex/issues/14593

[short linden]

5hr decrease but weekly limit still same. this prevent spamming the massive usage on multiple account during $100 plan campaign.

[spiral gorge]

oh?

that makes no sense what

[lapis cove]

There is no plan to make codex available for Intel Macs is there

[velvet wren]

nope, Intel Macs are done

[severe mason]

Yo anybody using Hermes? If so how does it fit your workflow and what model do you use?

[sharp gazelle]

so sad that they once again forgot about business plan users with the latest rate limit reset…

[wicked briar]

2x usage only for pro lite and pro plans

not for plus

[wicked briar]

cli better

[spiral gorge]

https://i.imgur.com/St8L4MB.png cringe... hitting a patch size limit? help?

[foggy wren]

Has anyone noticed that the usage of the plus plan has decreased significantly? I only used it for less than 2 hours, and the available amount of 5 hours is only 11% left.

[waxen bramble]

https://github.com/informigados/analyze-codex-tokens

[celest stag]

You can sandbox codex to WSL so it isn't affected by command length limits (which isn't a simple process if you aren't familiar with Linux) or wait until OpenAI to fix the bug. Last I heard they were working on a workaround.

I've heard good things from people who spend the time setting it up to work in WSL, but as someone who works in a windows dev-env using a lot of windows-only code, I haven't spent time trying it yet.

[spiral gorge]

YES YES. plus sub now has the limits of free version and 100usd sub is the new plus.
how to upsale 100 million

[spiral gorge]

i say no to linux

[foggy wren]

so they just do this for the 100usd pro user

[spiral gorge]

theydo what?

[foggy wren]

Cut the plus plan’s limit

[spiral gorge]

yes. they do

[foggy wren]

that’s ridiculous, I already run into 0 for 1.5h

[boreal holly]

I love how on Windows literally everything is a workaround

[celest stag]

Well, OpenAI seems to be a mac-first dev company, for whatever reason

[boreal holly]

It's a total mystery

[supple perch]

@boreal holly I mean companies are able to make those decisions. IE we made the decision to not support Windows

[ornate jolt]

think every company is mac first for the most part

i got a mac just for coding so if your a dev get a mac you can get one for like 800 bucks

[ruby pond]

I just received this email. But when I look at my credits I don't see the one-time credit boost they mentioned. Anyone got the boost?

[boreal holly]

hier
Hallo

I would be skeptical

I would check the sender domain. My guess is if you load remote content (e.g. images), one of the things that gets loaded is a "did this person open this email?" (did this image get loaded) and they'll know they reached a real person, at which point they'll send you phishing attempts.

I have a personal Plus account and I did not receive that email

[short linden]

"hier", is that real email ???

[harsh atlas]

hi guys!

[indigo robin]

because mac is superior

also why am i constantly getting stream disconnected before completion: websocket closed by server before response.completed , this didnt happen before now its almost all the time

[deft gyro]

2 spelling mistakes, probably scam

[deft gyro]

fellow golang fan detected?

peak

[spiral gorge]

https://i.imgur.com/5TOdlVQ.png

who has tried the new 100usd plan?

how are the limits?

[harsh atlas]

but with 5x or 20x we also have gpt 5.4 pro?

[hidden glade]

Mac is a staple in any tech company in the bay area, unless you work remotely for big enterprises, like the day my soul died when i received a few year old HP laptop for my internship for a fortune 500 company (actually probably closer to top 10/20) because big enterprise was still stuck with cisco webex (couple years ago) when everyone is using zoom/google meets or at least MS teams

[hard drum]

Cisco 🤮

I have not-fun-memories of Packet Tracker, and I condemn Cisco

[true bane]

everyone else on plus is getting crazy high 5 hour usage movement last few days right? I think the party's over sadly

[narrow wave]

are the steer and queue buttons currently swapped? enter and tab enter seem to be inconsistent, sometimes a 'queued' prompt won't actually wait for the turn to finish. also, would be a good idea if you could queue a separate thread so that it starts as soon as there are no active threads, i could queue a build and then a fresh context review/cleanup.

[patent patio]

So I was on the free plan and reached the rate limit for 5h/weekly (both were the same) and it said to upgrade to plus to use more. I upgraded to plus and I still have no usage... what.

[potent mason]

Maybe try logging out and logging in again? It’s probably that it thinks you’re still in free plan

[patent patio]

I think that's fixed it, it took a few attempts but I think it's now just a visual error, saying 100% usage but I can send messages 😂

[potent mason]

100% usage should mean you have 100% remaining

[sullen swift]

look what i managed to do

codex programmed a bridge that allowed a fully automatic randbats game (sd hosted locally) between two agents

pretty cool right

[lean lark]

Assuming you're using VSCode and the extension:
Rather than doing a queue from the same side panel, you can F1 for "Codex: New Codex Agent". That opens a window that's just like the sidebar - with a new thread.
The window can be repositioned, and some might prefer it to the sidebar. In earlier versions we could open more than one window. I recent fix of another issue locked it to one for now.

The Tab/Enter thing has always been wonky. There's also the weirdness of the composerEnterBehavior setting.

[true bane]

There's zero chance something hasn't changed starting Sunday/Monday. On plus I'm at 0% 5 hour left with 3 hours remaining. No matter what I would never be there before this week, same work flow, same stuff - I'd be at 50%, at worse.

Party's over

[kind jay]

@frosty zealot

[frosty zealot]

How did you know I literally just woke up

[kind jay]

I have access to your webcam

[frosty zealot]

Inchresting

[glacial agate]

so is it cheaper or more expensive with it being tied to the api now?

[oak tinsel]

weirdo

[kind jay]

I’m not talking to you

[pale rose]

"Rebalancing" I too will be rebalancing. Fun while it lasted, but I can't spare the extra cash for a hobby that's bringing me no income

[frigid shadow]

Hello, I’m working on improving my setup and I’m looking for a way to inspect the actual context being sent to OpenAI.

Ideally, I’d like a tool (possibly with a web UI) that shows the full prompt/context, including things like system prompts, skills, MCP data, and any additional injected content. This would help a lot in understanding trade-offs (e.g. skills vs MCP, or when sub-agents are actually useful).

Does something like this already exist?

If not, I’m considering building it myself (likely without modifying Codex directly). Any suggestions, ideas, or prior art I should look into?

[quasi summit]

where are we supposed to send feedback on the codex gui app? cos the codex repo is for app server and cli

[rocky fog]

got the same email
no spelling mistakes
I guess he had it tailored to his language but not fully? lol
from
[e-mail]

also no "one time credit boost" yet
nothing in history
unless thats supposed to mean something else than actually getting more credits
but then I also switched to pro few weeks ago

[wooden halo]

https://discord.gg/openai

[lean lark]

👆 um, we're already here. 🤔

[wind spoke]

So, is everyone else getting hit by this reduced 5-hour limit too? Do we know whether this is permanent?

[worn harness]

yes honeymoon is over

[glacial shadow]

I havent hit any limits at all

[boreal holly]

It's OK, when you buy credits they work on the 5hr limit too

[glacial shadow]

This technoology is the greatest thing ive ever used

[silver cliff]

hi guys

idk if this is where i should talk, but Codex in VS Code doesn't work

[stiff panther]

Anyone noticed that Codex Plus allowance went noticeably down even though "As a thank you for being a power user of Codex, we've provided a one-time credit boost that's available now in your account."? Consumes all the 5 hours worth of credits in like 10 minutes

Was thinking to go PRO but 10x / 20x this is still incredibly low

[silver cliff]

I litteraly give ONE prompt and then I run out

[stiff panther]

Oh I'm not the only one then

[umbral junco]

is possible use gpt for change stuff on website, download filess from cloud etc just like claude cowork?

or really i have to pay Anthropic

[dusk thorn]

What model are you using??

[silver cliff]

wait guys

I keep seeing people

say they use GPT 5.2 instead of 5.4

like for coding

why?

[dusk thorn]

Extra high 5.4 mini is good imo

[silver cliff]

is it good?

[dusk thorn]

Got 5.2? Not sure

[silver cliff]

yes

so

what should I use

cause i know 5.4 is very good, but

yk

[boreal holly]

Dang dude Codex is sloooow right now

I guess we're gonna get a reset soon when they hit 2 million active users or whatever lol

[dusk thorn]

I find gpt 5.4 mini to be able to do a lot on extra high reasoning for little token taking

[silver cliff]

ohhhh

maybe this will help me

[dusk thorn]

Did a bunch of backend work today and used like 5% of my weekly limit

Hourly took up I think 30%?

[silver cliff]

idk bro i type 1 prompt and poof

and also

i dont even have a 5 hour limit or whatever

why

[dusk thorn]

Depends on your use tho because it can take a bit longer than other models

[dusk thorn]

Really?

[silver cliff]

yea i only have weekly

[dusk thorn]

Plus right?

[silver cliff]

oh..

that figures it out

[dusk thorn]

I have plus and I got a 5 hour

[frosty zealot]

Last night for the first time I got a 'This model is at capacity try another one' message

[silver cliff]

yea that happend to me too

that's when I knew the downfall had started

[frosty zealot]

just people hyping the 100 plan, and with all the extra usage theres probably a lot of overlap which normally wouldnt happen

[short lynx]

I've been using the Codex app on MacOS heavily and I am getting a but frustrated by all the subtle and not so subtle bugs. Also, there seem to be bugs in the feedback system itself! Sometimes I get a failure to submit the report, sometimes I get the report-ID (good), and sometimes it seems to take the feedback, but there is no acknowledgment.

TL:DR; I wish the Codex team was on top of all the feedback and pumping out new versions to address basic issues. Thinks like 'Can't generate a commit message' in a lengthy thread.

[boreal holly]

"Command failed successfully" lol

[lean lark]

1. Check the GitHub repo for recognized issues.
2. The team is very active there.
3. Create a new issue there if you can't find it.
4. The team does occasionally peek in here but this isn't an official vector to the company, so don't count on notes here to translate to product changes.

[solemn acorn]

filing github issues is da way

you could even use codex itself to reproduce the issue

[kind jay]

@frosty zealot lunch looks yummy

[frosty zealot]

want some?

[kind jay]

No

[jolly lily]

How many times higher limits does the plus plan include than the free plan?

[pliant notch]

For Codex? Free doesn't have Codex access. For ChatGPT access, around 15x.

[rain kettle]

I used Free Codex in VS Code once, and the limits were insufferable. It's only good for really small tasks

[jolly lily]

So compared to the ChatGPT free access, it's around 15x more right?

[pliant notch]

Give or take.

[jolly lily]

Thanks

[pliant notch]

np

[lean lark]

"free doesn't have Codex access"
"so if I pay for it I get about 15x more?"
"give or take"
🤣

[frosty zealot]

Don't let dreams be dreams

[lean lark]

Everyone at some point should also consider using an API key. For hundredths of a penny you can process a lot of tokens. No subscription. No minimums.

[frosty zealot]

do i get banned too if I say

Check my Bio 🙂

[kind jay]

If you’re lucky, sometimes you can find them on GH, as a little treat

[spring remnant]

When is codex 100 plan, the 200 is too steep

[dusk thorn]

It’s out already

Released yesterday

[spring remnant]

Thank you

[dusk thorn]

https://community.openai.com/t/introducing-new-100-month-pro-tier/1378752

[boreal holly]

You can tell precisely when they changed from message-based to token-based 🤣

[boreal holly]

I wonder why Plus folks keep running into "1 prompt, 0% 5h" if that's the split

[frosty zealot]

https://developers.openai.com/codex/pricing?codex-usage-limits

its straight from here so im not sure

[boreal holly]

Wait a minute, it says the difference between 5x and 20x is twice as many messages

[frosty zealot]

maybe its one really big message lol

[frosty zealot]

Sueing for false advertisement

[boreal holly]

Ayoo where the 80x plan at? I want twice as many messages as 20x

[frosty zealot]

I wonder why code reviews cant be done by 5.4

[frosty zealot]

Switch to GPT-5.4-mini for routine tasks. Using the mini model should extend your local-message usage limits by roughly 2.5x to 3.3x, depending on the model you switch from. I

I wonder what they consider a 'routine task'

[supple perch]

umm its always been token based... gh copilot uses messages...

[frosty zealot]

Negative

As of April 2nd, we’re moving pricing to API token-based rates. Credits remain the core pricing unit that customers purchase and consume, but usage is based on tokens consumed, calculated as credits per million input tokens, cached input tokens and output tokens your workspace consumes. Read about tokens here.

This format replaces average per-message estimates for your plan with a direct mapping between token usage and credits. It is most useful when you want a clearer view of how input, cached input, and output affect credit consumption.

Under this model, actual credit usage depends on the mix of input, cached input, and output tokens in each task. The new rate card is displayed in the table below, and is currently applicable to new and existing Business customers, and new Enterprise customers.

New and existing customers on all other plan types should continue to use the previous message based rate card, until we migrate you to the new rates in the upcoming weeks.

[supple perch]

Okay thank you! I still would love to hear Roberts side <3 (they know more about what I do <3)

[oak trellis]

Good morning from Bali

So 100 plan is bad too ?

I know plus is dust now

[potent mason]

Hey @oak trellis how's it going

Did you manage to fix your issue with codex?

I'm honestly wondering what was going on with it hahahaha

[hollow holly]

You guys notice any showstoppers using Codex Hooks on v0.118.0?

The MCP tool layer is erroring my screen reader calls as if "speak the final response to the user" == "agent hallucinating to random external service", but I don't wanna spend time plugging it into hooks yet if they're too early-on and super jank or smth

[tiny kettle]

wait did they just learn about the axios compromise?

that was like 10 days ago

[balmy acorn]

prob.

and these people has the most advance in house ais and still get this attack

[torpid blaze]

so codex is unusable now?

with 20$ plan

is it still worth it?

[balmy acorn]

no u can still use it just prob hit the weekly instantly

[torpid blaze]

thats what i mean

[balmy acorn]

if your gonna use it alot 20 plan isnt worth it

[torpid blaze]

like how much is alot

i'm just doing side projects as hobby

like developing small to medium projects

not enterprise grade

[potent mason]

I think they said weekly limit was unchanged, the only one they changed was 5 hour one, so you now get more sessions with less usage

[naive tree]

1 prompt waited 5 min 50% gone of 5 hour usage, and week usage is from 100% to 70% with 1 prompt. I will definetly cancel subscription after this move.

@placid cairn Yeah, it’s really like that. After just 3 prompts I got a notification that the 5-hour limit was reached. I was like “what the hell that’s impossible,” but it looks like the Plus plan has basically become the Free plan—just paid now 😄

[potent mason]

Well two things happened.

The first is they stopped basing it off messages and started basing it off tokens consumed so I'm guessing shorter messages == less usage, longer messages === way more usage (since you can go up to like 50 turns without a problem).

The second is that Pro $200 changed from being 6x usage to 40x usage so they probably had to modify the base unit in some ways

[naive tree]

same here brother

[potent mason]

Claude is even worse right now, I have the $100 plan and I can finish a 5 hour session in 3-4 prompts (it's supposed to be about 6x $20 plan on 5 hour sessions and 8.33x on weekly limits)

I would try the $30 plan from copilot, I've heard that one is pretty generous right now, but besides that yeah GPT is probably still the best

Yeah np, if you do test out copilot tell me how it goes to see if I should stop suggesting it to people who are currently on GPT $20

[torpid blaze]

how about google?

is gemini's plan better? like they give gemini-cli, antigravity etc.

[wide hull]

Did they increase rate usage for Plus users? I'm somehow already approaching the weekly limit and I've only been using it for two or three days.

[torpid blaze]

token based approaching causing this prob

everyone is complaining about this

[wide hull]

do they have any plans to fix it? I have a feeling that they wont because they want people to upgrade

[torpid blaze]

nah

[wide hull]

ofc

welp thats gonna slow things down a lot

[torpid blaze]

they cant because if they wanted to why they even switch to token based approach at first

[torpid blaze]

ai coding agents problem is not enough computing power sadly

they always reducing usage limits to handle more users

[dusk thorn]

Hey, codex has a 100 dollar plan now

[potent mason]

Well theres nothing to "fix", they intended the change (unless they are somehow miscalculating something)

[torpid blaze]

they forced probably

[dusk thorn]

5x rate limits as plus, and now through may 31st I believe it’s 2x that so 10x rate limits as plus

[torpid blaze]

message-based approach is literally not making any sense

[wide hull]

i would be more open to buying tokens if it wasnt $40 for like, a few messages or whatever it is

[torpid blaze]

try minimax 2.7

[wide hull]

whats that

[torpid blaze]

like 1/30 cost of tokens but it tries to mimic top tier models performance (people says its literally doing great work not as much as top-tier but great)

[potent mason]

Yeah agreed, specially as agents become more capable we are going to be moving from 1 message -> 30 minutes of work to 1 message -> 8 hours of work

[dusk thorn]

True, think they reduced plus

[torpid blaze]

stupid than top-tier models by %20-%25 but way cheaper

[dusk thorn]

Sucks it was a major selling point but now they going full anthropic it seems

[wide hull]

hm

[torpid blaze]

if you dont need top-tier work its great to consider

[wide hull]

i mean... im trying to use it to help speed up programming for game dev in Godot so

[torpid blaze]

and you can also just make it do heavylifting and use your premium weekly limit to review it

[wide hull]

idk how well itll fare with that

[dusk thorn]

Codex becoming like Claude in limits is a nightmare hope it doesn’t happen

[torpid blaze]

it can be mental blow on this specific work i think

it's optimized for general workflows but using godot and programming a game is even hardwork for codeex and claude

[wide hull]

I mean it works ok if I give it the documentation as context

[torpid blaze]

but gemini is better at game programming and overall godot

[wide hull]

but the harder game dev problems is kind of difficult to nail it down

[torpid blaze]

but it's other thinking skills are not good

so it's a deal

[wide hull]

how much is it?

[torpid blaze]

gemini?

[wide hull]

ya

ive never used it

[torpid blaze]

it's half price for first 2 months

[wide hull]

but whats the price tho lol

[torpid blaze]

it's full price is nearly 20$

[wide hull]

so the same as GPT Plus then

...hm

[torpid blaze]

yeah but first 2 months half price

if you didnt buy it before

[wide hull]

I mean i use GPT for more than just coding so I kinda need the subscription for that too

i dont have a ton of money, GPT Plus is one of my only subscriptions

idk if i could do two of them simultaneously

[torpid blaze]

yeah gemini is good at game developing and godot overall but not good at other things like asking questions, architecture etc.

[wide hull]

but thats good to know regardless

[potent mason]

Its not even close to anthropic drop in usage but I get what you're saying, I guess for me it just feels like everyone knew this was coming they're giving $4-5k usage on a $200 plan

[torpid blaze]

gpt is overall better decision if you're having questions etc

[torpid blaze]

ya ya always good to consider options

[wide hull]

rip

[torpid blaze]

btw

if you use windsurf kimi 2.5 is free

[wide hull]

is that a local model?

[torpid blaze]

if you want to. but windsurf offers it tons of billions parameters model as cloud

for free

for limited time ofc but it's still active

[wide hull]

hm

[torpid blaze]

not giving top-tier work but gets the job done

[wide hull]

OH WAIT i forgot I have GitHub Copilot accessible too

thanks college lol

i can probably use that to help a bit

[torpid blaze]

gh copilot is the worse in my opinion

but if its making your work, then its up to you

[wide hull]

idk ill see what happens

mainly bc the thing im adding is getting pretty complicated, idk if i could keep working on it on my own

[torpid blaze]

current ai coding models are just try-and-see work model

[wide hull]

lots of moving pieces

[torpid blaze]

godot...

[wide hull]

?

[torpid blaze]

i thought you were talking about godot cuz i used it one time and it was hella complicated

[wide hull]

i mean its the thing that im trying to get working IN godot thats complicated

[torpid blaze]

oh

[wide hull]

im basically trying to make a mario-maker styled game editor in 3D

[torpid blaze]

oh that's cool af

god help you cuz ai probably not going to

[wide hull]

currently tryng to fix collision detection problems with rotation

[wide hull]

i mean its gotten me this far lol

[torpid blaze]

yeah the problem always begins after that far

in my cases

I make things, it goes so good then suddenly my highly intelligent super developer ai model goes dumb

[wide hull]

thats about how it feels rn

tho then again I probably wouldn't have fared much better

[torpid blaze]

yeah without ai me too couldnt even make this project this far

i'm looking for good subscriptions but codex,claude everyone complaining

[wide hull]

I dont mind having AI help with/do the coding bc there's only so many ways you can program something to have a certain outcome

[torpid blaze]

i'm stuck with gemini

[wide hull]

definitely dont want any Ai generated creative assets in my project tho. There are infinite ways to be creative.

[torpid blaze]

Yeah ai is hella good for developers that doesnt wanna learn all the language for simple job

Like I was trying to do some project but it needs me to use another language in that job so i use ai to make it and it's done

[wide hull]

I mean I mostly get the language, its just the "spending hours trying to find the exact configuration of features in the engine" part I want to skip lol

GDScript is a lot like python

[torpid blaze]

yeah I was using kotlin and project made me learn rust/go instead i just asked ai and it made both of them seamlessly i tried one and picked best result, no need to learn all that for such small work

[torpid blaze]

if its python-like actually minimax 2.7 can do a good job, cuz it's overly trained in python and optimized for it mostly

but the thinking phase, i pass that.

[wide hull]

can it get context from txt files?

[torpid blaze]

It's designed to be an agentic model like codex or claude

so it can take all type of context

[wide hull]

I have godot's documentation downloaded as a 10MB txt

[torpid blaze]

It can do it but i'm suspicious about it's applying phase

[wide hull]

hm

[torpid blaze]

even claude struggling to keep up with context files

it just forgots about it after some prompts

probably need a skill file or another approach

codex is good at context things

it doesnt forgot anything you say

[wide hull]

i see

oh shoot Codex actually got it right before the rate limit hit lol

[torpid blaze]

lucky guy

[potent mason]

Can we see how many users codex has? I want to plan my usage for the 1m resets

[wide hull]

maybe i should reduce the reasoning effort a bit when using codex

im assumign that reduces token usage

[potent mason]

I would avoid that, if you really want to reduce usage there are better strategies

GPT 5.4 high has a high improvement from medium. (dont use xhigh though)

[wide hull]

oh I've been using 5.4 on xhigh

:|

ill try regular high

[torpid blaze]

bro using xhigh and complaining about tokens

xhigh is like -> come and take my tokens

[wide hull]

I didnt know they changed it 😭

[potent mason]

Oh wow I'm actually flying through usage today on $200 plan, yesterday it felt unlimited

[wide hull]

Xhigh was one of the only ones that felt like it would work half the time

[torpid blaze]

what work are you doing bro i've done all my rust/go projects with medium and never had any problems

[wide hull]

also this is what i was working on. basically working now

[torpid blaze]

good work but its not really lookng like it needs xhigh

[wide hull]

🤷‍♂️ idk when i was testing it early on it felt like xhigh was the only one that would get it done right and fast

[torpid blaze]

maybe its about godot idk

in my rust/go projects it was doing everything right with medium

[wide hull]

idk, if it works it works

[torpid blaze]

yeah

maybe i should buy codex again

[wide hull]

and at first, xhigh was the one that worked

[torpid blaze]

maybe the right approach is reducing thinking after milestones

like at first when doing first basic skeleton of project -> xhigh, then after modifying it a bit -> high after everything done and working editing it with medium

maybe if i do it like this i dont hit weekly limit that fast

[oak trellis]

handoff.md works actually for me never let it compact .. just use handoff.md .. design your own template .. works well saving tokens for me

never believed in that stuff but it works

[quiet furnace]

What about an Android app for codex? I wanna get things done on the go!

[torn lodge]

I mainly use ChatGPT Plus for hobby programming, not for professional work, and the recent change from something that felt closer to per-message usage to token-based usage has made it much worse for me.

The problem is not that I am working on huge projects or doing anything extreme. This happens during normal everyday coding use. Even with regular prompts, ordinary back-and-forth, and standard coding help, I can hit the limit much faster than before and then get blocked by the 5-hour window.

That recent change is what frustrates me most. Before, it felt much more usable and predictable. Now, with token-based usage, even normal coding sessions feel cut short too quickly. For hobby programming, that seriously damages the value of Plus.

I am not willing to pay more just to keep casually programming under these new restrictions. For my use case, this recent change made the service feel much less worth the price.

[daring yew]

has anyone found out a way to get rid of all the bloat code the models generate? i dont want the best practices i just want the smallest code that works. verbosity didnt fix it

[rocky fog]

keeps resetting the usage
whoops, we need to lower max daily/weekly usage

[short linden]

what's model you using ? Medium / High / xHight ?

High will act as professional coding.

If you want to use the high, you need to promt the detail to generate small code, don't need model/controller style.

[daring yew]

Tried with all versions after 5. Used it with all except low. Don't think medium was better but didn't use it much

The cheap solution is to ask it to make the code as small as possible without breaking functionality

[cedar skiff]

Ask for minimal indirection with as little abstraction as possible. Give it a limit of 3 for cyclomatic complexity unless it it unavoidable, limit nesting to a depth of 3. Have an architecture that you adhere to. Build a topical-critical-architecture-guardrails skill that the agent must read every time and build it with reference files. As you encounter common mistakes add them to the skill.
In agents.md or in config.toml developer_instructions add instructions:

• when planning searching or implementing it should step through relevant code until it reaches concrete implementations.
• For planning and implementation follow local coding conventions - find 2-3 local code exemplars that match the implementation target if possible.
  This is short hand for what i do, but essentially this is how i get idiomatic code.

[short linden]

Or just : /plan mode with small detail "create feature...." with small code : that's fast solution, and check if agent planned to create model style, then you can change the detail here.

[daring yew]

I tried agents.md with copilot but completely ignored it, dunno if it works in codex

[rocky fog]

yes it does

[toxic vortex]

Isn't Copilot AI?

I thought that Copilot is AI, like ChatGPT

Or Gemeni

[daring yew]

What's cyclomatic complexity?

[rocky fog]

https://developers.openai.com/codex/guides/agents-md

[short linden]

Are you using codex in copilot ?

[daring yew]

No I meant the vscode extension

[toxic vortex]

oh ok

[cedar skiff]

lots of conditions in a method - this is a question for codex 🤣

[toxic vortex]

this one?

[daring yew]

No I meant I used copilot and now try codex

[cedar skiff]

a methods with a cyclomatic complexity of 4 and nesting level of 4 is a real pain to work out whats going on when you look at it

[toxic vortex]

I use only GPT, because i have a Pro subscription

[daring yew]

Honestly I like the ux of copilot much more

[rocky fog]

yes, or even in project root is enough

[toxic vortex]

Alright

I think that Pro subscription is too much expensive

103 euros/ per mounth is crazy

maybe 70-80 euros is enough

[short linden]

in codex, just start create AGENTS.md and PROJECT_STATE.md that and make rules with AGENTS.md like "prefer small codebase style, not need model/controller".

[rocky fog]

yeah just tell codex to create it and explain what you want

[toxic vortex]

It's simple isn't it?

[daring yew]

Why two files?

Don't they both get fed in the prompt anyway?

[rocky fog]

probably only agents.md would
and if you would tell agents.md to check the project_state.md then it would check it

[hard drum]

it's probably because of the God-awful VAT

now, imagine 24% VAT on top on a digital service

[short linden]

yes, it's simple, agent will automatics implement the rules into AGENTS.md
PROJECT_STATE.md will keep some detail/planned of your project.

[hard drum]

🤦‍♀️

[toxic vortex]

Ok, but i think 70-80 euros is enough

[hard drum]

it should be

[toxic vortex]

Yep

[hard drum]

I wonder if it's really not possible for OAI to not have VAT applied for EU customers on something that is literally digital

paying 24% (or wtv is your country's VAT size) on top for access to AI models is just unfair for all of us EU users

and then it's giving NA users an advantage in price

[daring yew]

Vat is clear thievery it should be outlawed

[rocky fog]

in the api you get the vat as credit
add 50$ but pay 60.50$ and 60.50 gets added to your api

[hard drum]

It's the question of, "why should we pay more for the same service?"

[hard drum]

I wouldn't mind... i don't know... A 9% or something VAT, but 24% in 🇪🇪 for example...

It is the question of, "the hell am I paying for out of pocket?"

[toxic vortex]

My neighbor country

[hard drum]

let me guess... 🇫🇮

[daring yew]

The only solution is to use a company

[toxic vortex]

No

[hard drum]

Yes, as if any hobbyist had a company

[toxic vortex]

Guess again

[hard drum]

🇱🇻

[toxic vortex]

yees

[hard drum]

OOOO

you guys have amazing kvass

[daring yew]

Maybe if a friend or family have business vat

[toxic vortex]

It's pretty nice

[hard drum]

also Karums

Karums cannot forget

[toxic vortex]

Thanks, and i love zabka

[hard drum]

zabka...? the POLISH zabka? that store?

[toxic vortex]

YEEESSS

[hard drum]

ZABKA YESSSS

you guys have zabka???

[toxic vortex]

i love poland so much

[toxic vortex]

No(

[hard drum]

(

Go I love zabka as well

OH, and Forum Gliwice

[toxic vortex]

omg

i already forgot about it

[hard drum]

Gliwice is such a pedestrian-friendly city it's insane

compared to the disgust that is Krakow

[toxic vortex]

I was in Warszawa, Krakow, Gdansk

[hard drum]

I was in Kamieniec, Gliwice, Tarnowskie Gory, Bytom (for like a few times just to go elsewhere), Katowice, Krakow (twice--never again hopefully), Warszawa, Pyskowice, Karchowice (once for a few days)

moutnful, innit?

[toxic vortex]

WOW

Good to hear

[hard drum]

almost everything is Silesian voivodeship there

because Silesia > rest of Poland

[toxic vortex]

Ok

[hard drum]

you get the public transit really nice

PKP Intercity, Transport GZM (often seen as ZTM locally), and so much more

It's... like second home to me

[toxic vortex]

[hard drum]

I may not know the language, but I can easily navigate most of it as if i lived there

[toxic vortex]

Nice

I need to code

Right now

I'll be back

[hard drum]

you need to codexificate

[toxic vortex]

Yes

bye

[kind jay]

Bye

[toxic vortex]

Also you can find me in #prompt-engineering #dev-chat

[rocky fog]

did they just remove the fast mode (also) for pro in the latest vscode prerelease codex extension?

its gone
also from settings
also /fast is not a slash command now
https://developers.openai.com/codex/speed

[runic breach]

Guys, did Codex introduce lower 5h limits? Last week I never hit one, coded like the whole day. Hit weekly limti, waited for reset and now after 20-30 minutes I have that message ~25% 5h limit left

[rocky fog]

thats what people have been mentioning if you scroll up, yes

[runic breach]

Bruh, moved from claude because of the limits, now here too :X

[rocky fog]

everyone is hammering the coding tools 😄

[orchid plume]

yeah, Plus's 5 hour limit was reduced so that overall it lasts longer in the week, but at the end of the day to increase the quota for Pro 5x and Pro 20x they had to change the balance

[runic breach]

As always, I am late for the party 😄 Subscribed to Claude, introduced the awful limits after ~2 weeks, now here after 1 week ahah

[rocky fog]

oh it just suddenly reappeared without even updating anything -.-
must be serverside deciding to make it appear or not

[orchid plume]

at least the limits on the Pro 5x and 20x plans are very good, 20x at the moment feels like it near enough shouldn't bother showing quota bars 😂

[short linden]

it's still work.

[rocky fog]

it was completely gone for me for like an hour after I updated to latest prerelease vscode codex

[short linden]

oops, I only use stable version.

[toxic vortex]

is /fast type of command that makes codex think faster than ever?

[rocky fog]

https://developers.openai.com/codex/speed

[toxic vortex]

Okay

[tiny fulcrum]

so what is the idea with business codex accounts now? the limits are kind of not usable if you want to use it for software development work

[toxic vortex]

Is it free?

[rocky fog]

when you dont expect to use your usage fully then you can just enable fast mode to use it more
that is if you care to have it faster

[tiny fulcrum]

talking about business, we ordered 20 business accounts, but they became basically almost useless now, quite a few people need a 5x Pro subscription, but it is not available for business

[toxic vortex]

Thanks

[short linden]

nope, it's not diff, just 1.5x and priority queue connection, and 2x credit usage. Look worst than normal speed.

Example : during days time, there have many user using gpt5.4, if you using "/fast" mode, it's will send command to priority server instead of global server.

[oak trellis]

wow

[tiny fulcrum]

well its not a big deal, I think you still get plenty of value, just can't work with this normally, 5h limit is depleted after 1-2h of work
I'm just trying to understand why the 5x Pro account does not exist for business, we just want to be able to manage everything at one place

[oak trellis]

20 accounts and they empty after few 10 minutes..

[orchid plume]

as a pro 20x user, I don't really notice a speed difference usually between normal and fast mode. Possibly because pro has priority processing which makes it faster than without anyway. But between the choice of fast mode for twice the cost, and using more subagents or concurrent work, I prefer not using fast mode and just utilising subagents or concurrent sessions where possible

[exotic cave]

That's because Pro x5 and x20 are x2 until the end of May... So enjoy it while you can...

[orchid plume]

even when 2x ends, that means I've only used 4% though, I used to use about 10% to 15% per day 😛

[exotic cave]

Then you are golden, hopefully they won't nerf things after the promotion ends.

[orchid plume]

🤞

[past prairie]

Guys do you know if using gpt 5.4 pro in the openai platform use usage that we see for codex ? It's wrote general usage in codex but idk if it's specific to codex or no

[orchid plume]

separate quota, ChatGPT's quota, whatever that might be for GPT-5.4 Pro

[past prairie]

Ok that's crazy

[toxic vortex]

Okay guys, let's talk about MCP, i made a guide for this using ChatGPT:

1. What is MCP?

The Model Context Protocol (MCP) is an open standard for connecting AI models to external tools, data sources, and services in a consistent way.

It defines a unified interface so that models can:

Access files, databases, and APIs
Call tools (functions) in a standardized format
Retrieve structured context dynamically
Work across different applications without custom integrations for each one

Originally introduced by Anthropic, MCP is now widely discussed as an emerging interoperability layer for AI systems.

2. Why MCP exists

Before MCP, every AI integration looked different:

One app uses custom function calling
Another uses plugin APIs
Another uses bespoke JSON schemas
Each tool requires separate glue code
MCP solves this by:
Standardizing tool + data access
Making AI systems portable across apps
Reducing integration complexity
Enabling reusable “AI tool servers”

Also, MCP is that protocol, that has code, here is an example:
{
"name": "search_files",
"description": "Search local project files",
"input_schema": {
"query": "string"
}
}

2. Handler logic:
   Accept request
   Execute operation
   Return structured result

3. Transport layer
   Usually JSON-RPC over stdio or HTTP

4. Best practices
   Security
   Restrict file/system access
   Sandbox tool execution
   Validate inputs strictly
   Design
   Keep tools small and composable
   Avoid “god tools” that do everything
   Return structured outputs, not raw text
   Performance
   Cache frequently used resources
   Minimize large payload transfers
   Stream results when possible

5. When to use MCP

MCP is ideal when you are building:

AI-powered developer tools (IDEs, CLI assistants)
Enterprise assistants (docs + databases + APIs)
Multi-source research agents
Long-context systems that need dynamic retrieval

12. Summary
    MCP is a standardized protocol for connecting AI models to external context and tools
    It introduces a consistent architecture: host → client → server
    It unifies tools, resources, and prompts under one system
    It complements OpenAI-style tool calling rather than replacing it
    It enables scalable, reusable AI integrations across ecosystems

And that's not all the features!

[velvet wren]

are you pasting from ChatGPT again?

[toxic vortex]

No, i make it in Word, by choosing sentences in https://developers.openai.com/codex/mcp

ChatGPT helps me with prompts, coding and the problems that i have in my life

[toxic vortex]

And also, I have experience. I have 5 years of working with ChatGPT and programming languages ​​such as C++, Python, C#

[velvet wren]

ChatGPT hasn't existed for 5 years

[toxic vortex]

I mean AI

[velvet wren]

which AI did you use before ChatGPT?

[cedar skiff]

whats your goal with this?

[toxic vortex]

Tell people what MCP means

[cedar skiff]

No one here needs that though

[toxic vortex]

GPT-3, Instruct-GPT

[toxic vortex]

Ok

[cedar skiff]

If they want to know about mcp they can do the same thing you did

[toxic vortex]

I'm sorry

[toxic vortex]

And they all from OpenAI

[velvet wren]

not for 5 years though

[toxic vortex]

It is

Trust me

You can also google it

[velvet wren]

how did you obtain API access 5 years ago?

[toxic vortex]

I can't remember that, but i'm sure that i have worked with GPT-3

[velvet wren]

check rule 12 of the #server-rules

[toxic vortex]

Okay, and?

[velvet wren]

tell me you have worked with GPT for 5 years or apologise for making a mistake

[toxic vortex]

i have? how can i trust it?

Send you an API-key?

[velvet wren]

I can't be bothered now, carry on

[toxic vortex]

Okay, nevermind. Just forget about it

[waxen bramble]

Script that analyzes CODEX usage

https://github.com/informigados/analyze-codex-tokens

[marsh thicket]

Did codex die?

[velvet wren]

working fine for me right now

[exotic cave]

OpenAI released the first version to the world in 2022...

[topaz phoenix]

He got a point

[exotic cave]

It's funny how the 20$ plan users are crying. at 20$ don't expect for much.

[velvet wren]

they actively reduced the limits for Plus, I could do a lot more with Codex on Plus than I can do now

[exotic cave]

Plus was never supposed to be used for daily coding, only simple work. I wouldn't expect it to write full features and new modules a day.
It's there for simple stuff, and for 20$ it's still good.

[cedar skiff]

Doesn't change the way expectation were managed.

[exotic cave]

I was here kinda warning about it... My worry is about being able to use models not about if 20$ will give me the power to write fully featured apps.

[toxic vortex]

No, look for GPT-3

[exotic cave]

I read your comment about ChatGPT which was released in 2022.

[toxic vortex]

Ok, just nevermind

[hard drum]

better than competitor for 20 bucks

these people complaining have not been on Claude Pro || Gemini Plus/Pro which are about the same-ish price, but way less limits overall in many areas

OpenAI is doing a huge thank here

[muted tinsel]

What's with this reduction in CODEX usage of at least 80% or more? Isn't that a bit much?

[tiny fulcrum]

the problem is the business accounts have no 5x or 20x licenses... if they fix that everyhting is fine but now business is literally useless...

[normal spruce]

💔

[nocturne folio]

might buy the glm coding plan just for this

[normal spruce]

eh

i was on the $20 but canceled my plan a few weeks ago

I think i have a few days left anyways

[boreal holly]

Pro 20x 🚀

I'm not gonna complain. I use the heck out of Codex.

[normal spruce]

sheesh

yeah im like that with claude x5 plan

and geminis free 15 mo pro plan for students

[boreal holly]

The way I see it, I have til May 31st to adjust to the new quotas. Gotta find more token savings. Always happy to share the findings here, but yeah the burn is tangible now

[exotic cave]

The burn is real, and as i said yesterday we are still on the x2 🔥

[boreal holly]

Exactly, that's why May 31st gotta adjust for token efficiency. I read about this thing called extended prompt caching that lasts up to 24h instead of 10-15 mins which might be useful (exists in the API, idk about subscription)

[short linden]

Previously, only paid credit will count as token base usage, now subscription will also count as token base usage too. That's why it's burn.

[normal spruce]

darn it

thankfully it finished

[rocky fog]

yeah I think when I had it run out, it still finished task as well

[normal spruce]

codex is great, just prefer other models for everything else chatgpt offers

[boreal holly]

Yeah I noticed on Pro when I hit 0% weekly and started using credits, the credits would evaporate very quickly. Now the quota is about the same as credits! Makes sense!

I think the short term for me is more gpt-5.4-mini. It appears to use 3x fewer "credits" on output and ~6x fewer on input. It's not a bad model either, just requires more frequent corrections

[deft gyro]

I have been on gemini pro

And the quota on gemini pro is better

Same 20 dollars per month

Jules alone has better usage limits, and im not even counting the cli or ag

[short linden]

Hello, try this "config.toml", this should reduce the token usage a little bit.

model_verbosity = "low"
model_reasoning_summary = "none"

If you "/plan" and don't care about verbose/summary from agent.

[boreal holly]

Nice, I have not changed verbosity but sounds potentially helpful

[deft gyro]

I have just been using mini on high mostly and only like 5% of 5 hourly per prompt

So

Ntb

[hard drum]

but their models are marginally worse

[exotic cave]

True, last time i used Gemini Pro 3.0 i turned back fast, it forgets things doesn't stick to instructions too. I guess not much improved with 3.1?

[hard drum]

nope lol

[orchid plume]

last time I tried 3.1 it was awful for agentic work

[hard drum]

their best model was gemini 2.5 pro 2511

I assume 2511 stands for Nov 2025

[exotic cave]

Also gemini 1m token context window is a big trap.
Source Anthropic when they genrally released 1M context window: https://claude.com/blog/1m-context-ga

[dusk thorn]

Release the super app

[exotic cave]

I think you can try to build it with codex no? OpenAI slogen is "Just Build It"?

[nocturne folio]

should i get the zai code plan

[nocturne folio]

"you can just build

[exotic cave]

Thanks for correcting me

[exotic cave]

I believe there are some benchmarks around, but then again those are synthetic and doesn't give the whole picture.

[dusk thorn]

its coming like

next weeik

week

with openclaw agents leaked

and all that

gonna be insane

[exotic cave]

Now seriously what is this super app? does it come with a new model? And about next week, any hard proof?

[torpid blaze]

are you mentioning the "super app" that merges codex and atlas

or smth else?

[exotic cave]

I don't know to be honest, Maniac has said something about a super app.

[torpid blaze]

i only know about the "super app" thats gonna merge atlas and codex

now i'm interested too

[exotic cave]

I am interested in any news of a new model, OpenAI been quite. Wondering what they are cooking to counter Mythos from Anthropic.

[torpid blaze]

there are rumors about they're making mythos opponent

[velvet wren]

5.4 hasn't been out very long

[torpid blaze]

before 5.4 openai was so quiet so 5.4 cant be considered new probably they're cooking smth else background

[naive tree]

just a simple 1 prompt with Codex Plus, it feals like a Codex Free 😄

[torpid blaze]

switch token based open src models

and use bigger models for just reviewing

or just pay more.

[hard drum]

no, because they as a provider quantise the GLM models to lobotomy

you get extreme garbage at around 30-60K context window

ridiculous

[naive tree]

You speak wisely, but I still think that sometimes it’s better to say nothing than to say something that doesn’t make sense.

[dusk thorn]

Tibo, An actual dev at codex, and every other codex developer soft leaking it for the past week

That’s a small part of the big plan btw, which is also super nuts

[torpid blaze]

so peak

[dusk thorn]

You guys gotta get twitter even if you hate it it’s the only place you get super up to date ai news and devs dropping stuff

[torpid blaze]

why openclaw is even there

[dusk thorn]

Because OpenAI I believe got the maker of it to join them

I think

But openclaw will be super useful with OpenAI safety because it will be able to help you do so much stuff

[nocturne folio]

i wish we could have a more natural benchmark that shows a model's competency

might go try to make it

[topaz phoenix]

do i get like this if i set codex at low ?

[dusk thorn]

LMFAO

Why tibo making us so impatient 😭

[deft gyro]

You made a statement about quotas, not model quality.

[nocturne folio]

gpt 5 released on a thursday

[dusk thorn]

?

No just a pro plan

Spud is gonna go against mythos for intelligence

[nocturne folio]

like gpt 5 released on aug 5 which was a thursday

im manifesting that gpt5.5 will release on a thursday

[hard drum]

yes, and they give you more quotas for crappier models

so it definitely isn't worth more

[boreal holly]

GPT-5.4-mini is really not that bad! I think as far as the code it writes, it's really solid. I think the only thing mini does a bit worse than the big model is in a stressful situation (e.g. tooling issues, tricky troubleshooting) it gets "stressed out" and will sometimes go in circles trying to figure out the issue. The big model handles stress a bit more gracefully. So the key to getting the most out of mini is informed steering during stressful situations.

[exotic cave]

Thanks, now i am interested what are the big plans xD

[orchid plume]

to name a few possibilities:

• GPT-5.5 or 6
• GPT-Image-2
• The all-in-one "super app"

[lavish vector]

Is there any way how to fix it ?
It used 30% of limits and did nothing

[silver cliff]

switch models ig

it's extremely annoying

[torn lodge]

GPT Plus and Pro x5 feel pretty unusable for programming right now, unless you’re willing to work very slowly because of the new token system. Does anyone know of another AI that’s comparable, or nearly as good, but with a setup more like the old system? Now it’s basically: work efficiently for half an hour, then go old-school and spend four and a half hours digging through code before you can be productive again. After that, you have to wait a few days before you can work properly again. At least with GPT plus.

[calm sigil]

Has anyone achieved good performance on using more than a 258k token window?

[lilac relic]

yes it’s useful to increase the window for huge repo’s but it’s token costly and can take a lot longer

[calm sigil]

Token cost I am not worried about, quality and drift. I use very little tokens, codex can run for hours and use nothing.

[lilac relic]

it’s worth it to just upgrade your prompting instead the more variables you fill out in your prompt the less reasoning it has to use to figure things out the window usually isn’t the issue

[calm sigil]

I have heard that it can be poor results though, if boosted to the 1M limit.

My prompts are large scale projects, its needs the patterns of the architecture.

I have reduced my token usage by 60% through "methods".

[lilac relic]

it can but codex as far as i’ve been using it i use it a LOT, its very dependant on the operator to fill variability, efficiently use tooling and focus on doing sliced runs rather than huge 1 shot multi action prompts

increasing the window is only worth it IF you can actually cover majority of the issues that cause drift at that scale

[calm sigil]

I am doing slicing runs, just need the context to isolate the problem. I feel like context compaction is removing history and then causing drift based on rules.

[lilac relic]

what i’ve done for things like that is a custom skillset that organises repo context PERMANENTLY as it works so that even if i reset its knowledge mid task, without spending a long time or risking drift again just trying to find everything

it checks the organised logs that it generates

this is also really good for long term projects that may a lot of the time require you to double back on things

[calm sigil]

Yeah been doing that maybe be more ridged, are you specifically asking for those md updated every turn?

[lilac relic]

compactions works like music compression you still hear the main song but OVER TIME small details leave, that’s the purpose of the custom skillset context is LOGGED in a specific manner that’s effectively lossless and not a giant 100k word log of everything

[calm sigil]

Yeah, I know exactly what compaction does, I do doctoral research on AI memory. The issues are:

1. I do not know how the latent space was trained for chat GPT, so we do not know if it natively uses 1M token are is doing Matrix broadcasting across vectors.
2. The Agent loop is hidden, so what are thier prompts and instructions and algorithms for compaction.

[lilac relic]

i’m tired af forgive me if i answer indirectly but in its instructions there’s specific triggers i have set for when it acfuallt does the log writing, it’s set to pause while it does that it usually doesn’t take long regardless but the only downside is you can end up with GB of codex run logs that pile if you don’t obviously manage them if they’re not relevant

[calm sigil]

Now that's valuable. I do see your cascading and data sprawl.

[exotic cave]

This will answer your question, accuracy suffer badly.
Find it here: https://openai.com/index/introducing-gpt-5-4/

All the while anthropic have quite good accuracy for their 1M context implementation.
Source: https://claude.com/blog/1m-context-ga

[calm sigil]

exactly what I thought, the Model is not trained on that size of a Latent Space matrix. Therefore, they are taking a 1M token matrix and folding it on itself and "hoping" that they weights of the vectors stacked on top of eachother improve latent variabel search.

Which this seems it does not.

[lilac relic]

see those i’m not sure of currently that’s interesting but i do know it’s been more beneficial on my end to assume that certain functions are not agentically perfect and creating skills and triggers for them that replace functions in tailored ways

[calm sigil]

It works for Diffusion models sometimes. Comfy UI does that, the token window of a model is 77 Token, if you put more than that, they overlap (broadcast) the numbers on top of that 77 token window. Effectively adding them together.

[lilac relic]

has effectively replaced certain things

[potent mason]

Yeah no offense if you’re going past 170k tokens because you made more than one prompt it doesn’t matter the model you should’ve started a new one, keeping them at 100% is important

[lilac relic]

it’s just a bit tedious because you do have to figure out all of that on your own accord

[calm sigil]

I feel the tooling is doing the heavy lifting.

[exotic cave]

That's not how it works, read Rune messages.

[calm sigil]

My 1 prompt is that large. Its Examine this code and do this to it.

[exotic cave]

Depends what you need Claude might be better for your work for recall from a big context window.

Though make sure to look at Claude IFBench it's not so good xD.

[lilac relic]

oh yeah 100% tooling integration is like the core foundation of its success imo, for people advanced enough to use every feature you actually can really abuse codex for just about anything, it just shifts a lot more of the automation away from codex and onto you

when i’m bored my favourite thing to do is make it do really non general things

today i tested its capacity to be an anti virus and im not gonna lie it did pretty good (cli)

[calm sigil]

Yeah, I just think the models are not there yet, they have not been trained for large projects, its not in the Stack overflow dataset lol, its all one fixable item. Models are only as good as the example training data, and the one thing we dont have is 1 person looking at a entire repo documented. Its not something humans do, so its not training data we have.

[calm sigil]

https://openai.com/index/harness-engineering/ its OpenAI's own philosophy.

Trend of improving the determanistic tool suite.

blah.. let me fix that.

[lilac relic]

one extremely powerful thing i’ve noticed the codex cli is actually remarkably good at doing (tasks like this you do wanna make the context window larger) but isn’t a mainstream use is full system tech support, it returned my old laggy laptop from a rust bucket to almost running brand new and tuned the cpu voltages, it even managed to diagnose an issue on my friends pc remotely that multiple computer people and part swaps were not able to solve, it was some bios thing and a few other weird caching logic issues

provided your trusting enough to give it full agency and just sit back

[calm sigil]

Yeah, thats cool I have been wanting to experiment with that.

[lilac relic]

trust me do it 😭

i’ve found more enjoyment in making it do niche tasks than actually developing with it

[calm sigil]

Sweet I will, did you have it run Powershell commands or just run through codex desktop app?

[lilac relic]

mine is cli set up at my computer root using pwsh 7 and it’s able to run anything including powershell commands yes

i have the app too but i use cli a lot more often

[calm sigil]

I use the VScode plugin so I can work and check things in the CLI, is there a benefit to the CLI over the plugin? Right now mine is working in a pure linux environment WSL. Thats where I Code and create stuff.

[lilac relic]

i also suggest installing certain things MANUALLY that aren’t default on windows and baking the increased command capability into its knowledge by default

like for e.g there is a lot of pwsh functions that have existing alternatives that are much more efficient

like when it has to scan big file paths for certain content certain pwsh commands are HEAVY as f on tokens despite other command types you can add doing the same thing much more efficiently

i noticed on 115 this was actually a major issue that standard codex would end up in where it would end up doing MASSIVE uneventful get-child chains to find information it wasn’t sure of

i only even thought of replacing certain command functions because of that

[exotic cave]

If i am not mistaken the plugin just runs Codex CLI behind the scene, but i might be wrong. Also the Plugin might lag behind.

[calm sigil]

DM me we should swap solutions, I have a suite of tools I am building for Linux for this exact reason.

[lilac relic]

my experience with using it through vscode wasn’t fun your better off actually using it externally and just pointing it at the repo when needed

[calm sigil]

It does, I just havent been leveraging the / commands much, I think its a habbit when you use the CLI directly.

I feel I have to ask the AI why something is happening how its different before the agent loop actually does that itself and then adds that back into the context. This seems like a specific flaw in the Codex agent loop. OpenAI if your listening, you may want to fix this.

For anyones information, the Agent loops are desinged to navigate a LLM latent space which can find answers givent determanistic text. Therefore, we are kind of doing gradient decent withing the latent space of the LLM which in turn finds latent variable (IE, the answer we want)

[lilac relic]

default codex behaviour is a bit like that yea but these things are changeable manually using developer instructions and the agents.md files atm

like i wouldn’t say it’s very beginner friendly yet

unless your making a random number generator

[calm sigil]

True, but have you seen their default codex prompt at the top of the session file?

ITS 15KB !

[tiny kettle]

Anyone else having issues with GPT refusing to do basic things through Codex? As an example, I asked it to do a review of an issue with a directx shader producing dimmer output than expected, it ended up responding to a question it asked itself while looking instead of what I originally prompted. I ask it to provide a response that fits my original question and it directly tells me "No."

It then took thirteen messages back and forth for it to quit refusing the request and actually complete the initial task. I had to walk it through stuff like "What do you think is worth refusal here" and "why exactly would that prevent you from completing the task" before it finally acknowledged it could do the task and began.

I'm having this random denial behavior almost every single session I start lately and I can't figure out why. I'm getting it across repos and with no AGENTS.md configured.

It's been doing this for all kinds of random tasks over the last ~2 weeks. Using defaults, 5.4 on medium thinking.

[undone lava]

Uhm, why? The website shows 4% aswell...

[calm sigil]

That next prompt with the current tokens you have on that white circle on the bottom right there would go over the 4%

[lilac relic]

you probably have a weird twin cli setup issue this is actually interesting of all issues i’ve never seen this

[undone lava]

there is no way, "test" would use 4%

[lilac relic]

maybe some weird sub agent settings

but honestly to diagnose your own issue

you could fresh install at a new location

another brand new cli install

and have it investigate 🤣

[toxic vortex]

Do you have Pro subscription?

[calm sigil]

Codex agent loop is realy build around goal chasing and planing. Ask for a plan to fix something, then execute the plan.

[undone lava]

no plus subscription