codex-discussions

[cedar parcel]

they;re all running

[normal spruce]

this issue is on openais end right

"stream disconnected before completion: An error occurred while processing your request."

?

[cedar parcel]

dark clouds gather....

[glacial shadow]

focus on coding. I doubt its a bug

[leaden heron]

Yes

[glacial shadow]

A 100$ plan would be great. Might have to go 200$ plan soon

[normal spruce]

🙁

[bright swift]

this cant just be the 2x promo ending...

[leaden heron]

it is a bug

3 prompts is literally the 5 hr limit

[cedar parcel]

4.2 bln... Nice!

[leaden heron]

1896 dollars lmfao

[bright swift]

and its just 2 sessions without subagents, something definitely wrong lol

[leaden heron]

insane

They gon have to reimburse us

[cedar parcel]

or it built something really epic

[glacial shadow]

What are you guys building

[bright swift]

same as everybody, a system to build the system.. 😉

[glacial shadow]

the buildings get so good, it all becomes secret

[leaden heron]

yes

[glacial shadow]

next gpt model will be insane, hoping also remote control cli

[leaden heron]

New model would be useless if 1 prompt drains your entire 5hr and weekly limit

[glacial shadow]

please dont start on the usage thing. the claude discord has been overrun by spam about complaints

Are you on Plus or Pro200 plan here?

[bright swift]

pro ofc

[leaden heron]

3 business seats

[glacial shadow]

usage lmts should be fine in those modes

[leaden heron]

A lot more than pro apparently

[glacial shadow]

and generous

[leaden heron]

4.2B tokens

[leaden heron]

Not at all

[glacial shadow]

Is there no way to buy more?

[leaden heron]

?

[glacial shadow]

is there a way to purchase extra usage if you run out

[leaden heron]

yes

Credits

[glacial shadow]

ok thats good

[bright swift]

which are now priced at api rates, or will be

so better get that productivity up

[leaden heron]

Credits aren’t even worth it right now

GPT 5.4 will just drain it all

[glacial shadow]

ai is greatest tech world has ever seen. buying a slice of it for 200$ is a steal

cost to hire developer to do the same is hundreds of thousands

[bright swift]

i wouldnt have an issue paying for 10x pro if they made clear tos for it

[glacial shadow]

so im not complaining

[bright swift]

or reasonably credits pricing

[glacial shadow]

so would be priced over 1000/month?

[leaden heron]

Right

[leaden heron]

2000

[glacial shadow]

its worth it if youre making serious money using it for your things

[leaden heron]

Enterprise

[bright swift]

isnt enterprise api pricing?

[leaden heron]

Not sure

Probably is doe

It’s customizable so probably better than paying for credits

[bright swift]

they changed the pricing model after april 2. credits are now exactly same price as API. soon for plus/pro too. https://developers.openai.com/codex/pricing?codex-usage-limits=plus#how-do-credits-work

[glacial shadow]

how fast do credits deplete when in the middle of a project

this is probably too individualistic

[leaden heron]

fast

[bright swift]

i wish i could afford gpt-5.4-xhigh-fast at API rates. not more worries about limits, the dream

[leaden heron]

bring back 2x 🙏🙏

[glacial shadow]

2x is enticement for people to sign up

time to get that credit card out 💳

[cedar parcel]

this was always the case. its for credits you buy after you burn through the limits

its much better to get a second, third, fourth sub (cough) than pay for credits/api

[glacial shadow]

not worth potential ban in my opinion. ai coding too valuable

do things the right way and sleep soundly at night

[cedar parcel]

if you're sleeping at night you're not coding enough 🙂

[glacial shadow]

not a service you want a ban from

[bright swift]

just hope they reset limits soon only 35% left and 4 days to go

where is tibo

[leaden heron]

i hope they fix this entire thing not just reset the rate limits

[lean lark]

Best thing most peeps can do to start is to just disable Fast mode - that nearly doubles token usage ... so there's your 2x back. You trade time for money. Want to use less of your time, pay money. Don't want to pay money, it will cost you some time.
Frankly I don't even notice the difference between Fast and Standard mode anyway.

[leaden heron]

not even using fast mode

[bright swift]

i never use fast mode

[leaden heron]

literally

[bright swift]

but 5.4 xhigh

[leaden heron]

i use medium and it drains

[lean lark]

As I said "most peeps" ... that eliminates many here ... but perhaps not "most".

[glacial shadow]

its so good lol

[bright swift]

yea even going down to high i often big difference

[glacial shadow]

i dont use anything other than maximum overdrive

[bright swift]

same.. "update the readme -> 5.4 xhigh"

[glacial shadow]

U use claude too? or only codex

[lean lark]

Devs here need to recognize the Good/Fast/Cheap triangle. https://www.google.com/search?q=triangle+good+fast+cheap
Pick Two.

[bright swift]

i downgraded claude max 200 actually after 8 months or so yday. had gpt pro + claude max in parallel. but i just cant deal with claude anymore

[glacial shadow]

max5 plan has generous limits imo. what couldnt you deal with anymore, its replies?

[bright swift]

mainly the tendency to jump in and write code

codex much more methodical

[glacial shadow]

which produces better code in your opinion

[bright swift]

also much better at following instructions, no "production ready" BS, and most importantly, the codex cli auto compaction is just magic

codex no doubt

[lean lark]

Another way to look at this is to back away from this new code fad where everything needs to be done by some third-party provider. Do your own coding, the way it's been done for decades, and rely on the new tooling for details, higher quality, and tougher challenges. Balance tasks more equitably between human and bot and you'll save all of those precious tokens.

[glacial shadow]

Im here to spend token like a flamethrower in vietnam

[bright swift]

i'd be like a beached whale without AI tbh. mostly using rust and ts now and never wrote those by hand

[glacial shadow]

I had to tell claude to not conserve tokens for the sake of efficiency, because i noticed they literally had claude back off on token usage from the spam feedback they were recieving about free plan/20$ plan usage limits

it was giving me half-thought through replies and crap depth before i instructed that

[leaden heron]

it would take hours to write a single script that connects to everything correctly while an AI would be able to do the same task in 5 minutes

if the script is like 800 lines long

[lean lark]

AI is kinda like a gateway pill. If you let yourself get addicted then you will forever rely on other people for your own products. That's NOT good for any of us. It's like forgetting how to make fire ... which most people can't do. Yes, at some point AI will write all the code and we won't need to see it, and coding languages won't use syntax that's meaningful to humans - it will be purely functional "machine code". But we aren't there yet. YOU must understand code. YOU must write some code. So the more token costs increase the more you should realize that maybe you shouldn't depend on it so much.
-- Patronizing lesson of the day 🤣

[leaden heron]

i think of AI as a translator

like telling me to speak another language when i could be using google translate

or ai

whatever

[hard tulip]

[bright swift]

funny thing about claude is.. i use super strict linting settings, and have those in my posttooluse hooks for fast feedback loop (in my own codex fork because for some reason they are a bit slow to release). and after a while claude just stops. not even a response like "this is too hard", just stops. and codex powers through.

[lean lark]

Learning and using other languages is good for the brain.

[glacial shadow]

I have noticed the laziness too, imo its directly caused by the spam feedback they get about complaints

[leaden heron]

its also takes alot of time

[bright swift]

yea probably it thinks the write tool is broken because it constantly gets error

[glacial shadow]

its fun too

[lean lark]

Um, time for a better brain? Sign me up! ( I speak about five languages, mostly pretty bad, but my brain is pretty healthy ... except when I open the refrigerator and look for the morning coffee, or I look around the house for my glasses which are already on my face. But those are exceptions really! My brain is healthy I tell ya! I aced an IQ test! I correctly identified the giraffe!) 🙄

[bright swift]

only negative thing about codex is that it seems very focused on its last response. you cant iterate as well as with claude to build a understanding over multiple turns. but i think this is what actually makes it so good in long running tasks on the other hand. so just have to understand how to talk to it. i also miss doubled ESC in codex to quickly jump back.

[glacial shadow]

I think this will improve wrt time as well. The pace of full codebase generation is unreal in its current state, so next models and codex updates are going to be insane

They* went from autocompletion in vsc using ai to now thousands of lines in minutes

[bright swift]

i missed the autocomplete era and went straight from writing everything by hand to not reading most code anymore lol

[glacial shadow]

I did not use it during autocomplete either

[leaden heron]

i speak english and mandarin both fluently and again its not about the brain its about the time it takes

[lean lark]

You have to remember that on every turn it needs to look at the entire project like it's never seen it before. If your project isn't organized or document, then it's duplicating wasted effort in every turn -- consuming time and tokens. For way too long the FOSS world has taught newcomers that comments don't matter. But they always have, and now even more-so. Craft AGENTS.md to instruct the assistant to always document its code and to update project development docs, and then always to refer to those docs for details. The project is so much better for it, and you also save a ton of tokens.

[leaden heron]

yo

this guy is an AI bruh

[bright swift]

same thought lol 😄

[glacial shadow]

No

[leaden heron]

🤦‍♂️

[glacial shadow]

his profile says hes been coding since 1979 and what hes saying makes sense

[bright swift]

i try to turn every line in agents.md into automatic enforcements/lints and make it as small as possible

[waxen terrace]

Have you thoroughly tested Claude Code and Codex?

[bright swift]

i'd say so yea, using both all day for 6+ months now, but just downgraded claude code to only pro from max 200, used it less and less

[waxen terrace]

Back then, I used Codex and then Claude, and I was amazed, but what’s happening with Claude right now like the usage limits, the Claude code leak, and the somewhat poorer performance is making me skeptical.

[glacial shadow]

oh god please not here, that unofficial claude server is overrun by "usage limit when" spam

[waxen terrace]

I think I'm going to switch back to Codex, but I want to make sure it's really worth it or if I should just stick with Claude after all

[bright swift]

you mean like telling everybody how your new model is too dangerous to release because of its cybersecurity capabilities, and then having your flagship product source code leaked... yea 🐑

[glacial shadow]

sub to both and make determination while using

[waxen terrace]

You mean Mythical?

[waxen terrace]

What do you think about that?

[glacial shadow]

Not understanding the reply to my statement

[waxen terrace]

What do you think about that?

[glacial shadow]

I dont think it requires much forethought. Just start using both right away

youll know which one you prefer from using them

just jump in

[glacial shadow]

Codex has way more tooling than i realized

[bright swift]

and you can build anything you want on top of it.. i'm trying to refrain from building an orchestration system though 😄

[glacial shadow]

do it gif not allowed here

[glacial shadow]

well what do ya build overall

whats the main goal

[bright swift]

i'm working on my pre-AI business, expand it a bit, legacy code work. the bigger project is sth new in data analytics space

[brisk junco]

Does anyone know if it is possible to run Codex on your VPS and message it from the Desktop app?

[cyan wing]

If you or a loved one have been effected by AI Psychosis
call the number on your screen now

As you may be entitled to financial compensation.

[normal spruce]

ah shucks

[undone bobcat]

So my rates reset an hour ago. Just got back to the pc. Doesnt look like it completed. hard to tell but 3 of the agents are still awaiting instructions.

Is there a way to trigger it to pick up where it left off without undoing all of the last changes and rerunning my last instruction?

[boreal holly]

You can sneak special signals into the system prompt. I have one called **DRIFT**, that tells the orchestrator they failed to handle something urgent. I like how the orchestrator knows precisely what the drift was without even investigating 😏 no more typing lengthy "you didn't do this and you need to do this", just slap em with **DRIFT** and they recover automatically

[nocturne folio]

unironically gpt6.7 could release if openai keeps following the monthly incremental drops

[lean lark]

"Robdex" 😂 💕

I should get the agent to save log detail as "Captain's Log, Stardate -297144.29 : Enhanced API to include ..."
[ Edit: Corrected for actual TNG stardate ... cuz, details matter ... ]

[solar sleet]

Hmm, It did say like a few days ago "renew 1 april" for the tokens, but now its still 0 ?? but says renew like 8

whats going on

[cedar skiff]

tehre was rest on the 1st whoch put everyone to the 8th. Maybe you hit a bug with that?

[glacial shadow]

nice

[undone bobcat]

Naturally my next question is - what does your drift orchestrator look like? And because you've ended up at that solution, I presume there's no way to simply continue, or rerun it easily?

Am I right in thinking if I undo the last bunch of changes - which I believe was a fraction of the work I needed doing, and then rerun the prompt, then I'd also be in good stead? Ofc it'll burn tokens doing the same work it's already spend 17misn doing before I ran into my rate limit

[undone bobcat]

https://help.openai.com/en/articles/20001106-codex-rate-card Maybe this? Note: As of April 2, 2026, we’ve updated Codex pricing to align with API token usage, instead of per-message pricing.

[boreal holly]

The **DRIFT** means at least one the following conditions are not met

• All blocked QA agents have a worker implementing a fix
• All unblocked QA agents are actively working and not idling
• All unarchived workers are working on something and not actively awaiting one of the workflow gates (pre-implementation planning, pre-merge, cleanup)

Rather than typing it out I just send **DRIFT** and they know they need to reconcile the invalid orchestration state

[cedar skiff]

skills can do the same thing. It's a pity the skills standard never came with a way to configure a skill to be on command only. Lucky openai.yaml handles it. Anthropic sort of half cooked it though. They deprecated slash commands in favour of skills, but slash commands dont inject anything into the system prompt and skills do. So if you had a whole lot convenience slash commands they begin to pollute the context.

[boreal holly]

Skills are for agents that do a lot of different things. Orchestrators have one extremely well defined and bounded job. I don't wanna be like $do-your-job skill on a general purpose system prompted agent, or be like $orchestrator-recovery for situations when they are failing to orchestrate.

[cedar skiff]

You can configure skills to be just simple inline commands with codex. So they don't get injected into the system prompt. They are only invoked when you explicitly called them. It's a replacement to the "prompts" that they used to have. So skills wear both hates.

I get what you're saying, but it seems similar to me. In both cases you are giving the agent instructions to do some thing

[boreal holly]

Yeah, skills make complete sense if your agent is designed to do absolutely everything, but I do not want my orchestrator writing code. The built-in system prompt literally says stuff like "prefer writing ASCII instead of Unicode unless you see unicode there or the user explicitly asks for Unicode". Mine says "Absolutely no file editing under any circumstances", so all of the skills for normal agents become irrelevant, and so does the section about "how to edit code properly and use apply_patch", because they straight up to not do those things.

So their entire purpose collapses into a very simple job, and I have a keyword to remind them to handle unfinished business

[cedar skiff]

I understand what your saying, if it works it works. skills don't = writing code though, they are just prompt instuctions, any prompt insturctions you want.

like "handle your unfinished work"

[boreal holly]

Yeah, and after compaction they have to re-read a skill file. They are forced to consume the system prompt, and it takes precedence over everything. They treat it like Isaac Azimov's 3 laws, absolutely shall not be infringed. Skills are more like suggestions.

[cedar skiff]

For better adherence after compaction makes total sense. That is a little bit of a sore point with codex, using xhigh doesnt usually leave enough room for the task to happen in one hit and often it's too close to compaction after thinking about it that it's not reliable.

On big tasks*

So orchestration is the solution

[boreal holly]

The other reason I did it that way is rooted in how LLMs work. Their "attention workspace" and neural network is designed around positional embeddings. That's why you can feed a model some massive document (haystack) and ask it to recite a particular passage (needle) with almost perfect accuracy. For my orchestrator, **DRIFT** is a positional embedding inside the system prompt in the beginning of the chat history, and it makes their attention workspace apply extra focus to that area without re-reading the prompt or anything. It saves tokens and redistributes their system prompt into their attention workspace.

[cedar skiff]

Did you see the work kimi dev team did on this?

They worked out a way to prevent the diffusion of attention as the context grows.

[boreal holly]

Interesting! I have not read about it

[cedar skiff]

It's directly about the point your making

https://github.com/MoonshotAI/Attention-Residuals

all the new ideas that are coming out of research are showing there is tonnes of room for large step gains

not just scaling

like every point of the process is like its just held together with glue and gum atm

and they are going to refine it

[boreal holly]

Wow that's incredible 🤯

[cedar skiff]

turboquant is another big step - on efficiency though

[quaint sigil]

anyone else feel like it’s easier to hit weekly limit then it is daily

i have 2 accounts on plus plan already at 70% weekly

[cedar skiff]

What is daily limit?

[lean lark]

Nice stats on Attention-Residuals

Like RAG, I see all of those NIAH issues being resolved over a short period of time. The tech is evolving very quickly and there are a lot of eyes on these common issues. I think limits on context are artificial, based on these very problems. Once the issues are gone, context limits can go up significantly. The next frontier is solving this ongoing issue of suggesting vs command and "why didn't it do x?". There are too many files in too many locations over too many platforms (system, user, project, codex, repo, context, turn) and most people don't recognize that there are so many tiers, nor what happens at each tier. I hope that the tech will evolve so that we can control everything under the system instructions in one dashboard, and see exactly what directives a model consumes (and not) with every prompt.

[twilit bluff]

seems like you are looking for afk agents/ralph loop, but I don't use those so I just send continue prompt

[boreal holly]

I made a script a while ago you might like: https://gist.github.com/robertmsale/8aa81706735621165efb9262e187da33

It mirrors the AGENTS.md loading pattern in Codex, so you run it inside some CWD and it will compile down all of the AGENTS.md files, including AGENTS.override.md and any fallbacks following the same patterns as codex, lists the loaded paths, load order, and context cost in bytes. I don't use it anymore (no more AGENTS.md) but it gives you at least 1 of the 3 pieces of the puzzle

[lean lark]

Will look. Thanks Bud!

I'm gonna code a ditty to notify me when Codex finishes a task, like send me a Discord or Slack message on completion. This is probably a simple hook.
Does anyone already have this published for Node? TY

[lean lark]

Simple codex_hooks feature for Stop events are still not complete or stable, especially when using VSCode extension. I'm gonna build a little more around that and will probably publish something.

[frosty zealot]

My 20x CC runs out tomorrow, and cancelled my sub, just reset today, be free

[vital hill]

Maybe stop using high xhigh for everything

[dawn seal]

After working quite intensively within unity and Codex, I felt like as if Unity was being an obstacle thats blocking Codex's capability to function. So I tried a new project and requested it to use Rust + Vulkan + SpacetimeDB and recreate my project in the custom engine and it did in a few hours. It's fully functional and now this game runs on cloud 24/7/365 using SpacetimeDB, you can play it now (account creation not required as it even has guest mods, and it even supports creation of your own world-dimension and you can travel to it instantly )

[glacial shadow]

but its just so good

[undone bobcat]

I suppose that should work too. I was trying to overthink it. I sent a prompt to acknowledge the incomplete changes, audit the files, and then continue with the earlier fixes and report back on the state of the repo

[bitter ether]

yall think i can make it?

[leaden heron]

It will finish it unless context compacts inbetween

Which it probably will

🫩

[scenic temple]

I am making a Rust CLI with Codex right now. It is an all-in-one toolbox with a plugin platform.
I'm adding AI features for it and u can just run
cat error.log | tinfo ai fix
tinfo ask "why is this slow"
to get a solution quickly in the terminal.
By the way, it's a fully open-sourced project licensed under the Apache 2.0 License.
https://tinfo.1234567890.dev/

[bitter ether]

holy, it was a heavy change and it only took 1% and %3 of 5hr usage 😱

[cyan wing]

are we cooked chat?

[bitter ether]

unless your doing something truly revolutionary, nah 🤷‍♂️

[glacial shadow]

Codex xhigh is so damn good

[hard tulip]

[deft gyro]

Does /fast apply to gpt 5.4 mini

[deft gyro]

its really not

gpt 5.4 xhigh does not perform significantly better than gpt 5.4 high

[chrome raven]

lol that's wrong

[tame magnet]

Can someone explain to me, why 4 prompts on a team account got my 5h window down to 10%

and the same 4 prompts on a plus account got it down to 90% ????????

like team account drops as fast as free account

this is ridiculous, for a higher price

whadafak

[velvet wren]

that depends on what the prompts were and how much Codex worked on them

[tame magnet]

nah man, I said "hi" on the team account with gpt 5.4 low, and it dropped by 4% of the 5h window
did the same with the plus account, it did not even drop by 1

this is not normal

team account is supposed to have higher limits than plus

[velvet wren]

I don't think that is correct

[tame magnet]

well I don't think either

but I did the test on 2 team account from 2 different teams

and it's the same

I can believe that the usage limits arent x2 anymore

but this is just non sense

[cedar skiff]

80 prompts on 2x in 5 hours sounds about right for some half decent prompts

[cyan wing]

"spiky intelligence" as they say...

[fading spear]

did you make it? XD

[cedar skiff]

if the goal is to use the 10% before the reset i think you can do it.

[fading spear]

I'm cooked for 4 days 😫 do I gotta like.. type characters again? I thought altman said I don't need to do that anymore

[cedar skiff]

there is cheap alternatives that can fill the gaps, minimax, glm, open router has free good models from time to time, currently it has qwen 3.6 plus, but you have to add 10 credits to your account to use free models for 1k request per day.

a $10 minimax account gives you 15000 request per week with 5 hour of 1500

[undone bobcat]

Does anyone here only ever use GPT 5.4 on Extra High?

Has anyone got experience of not always using Extra High - how does it perform? I've found myself using the 5.4 / Extra High combo 100% of the time - not sure that I'm giving up if I drop it to High or Medium. Would be good to know if anyone's done any side-by-side comparisons (+ real-world token usage differences)

[cedar skiff]

easy, siwitch to high and see if you notice a differencce

just use high for a day

[vital hill]

yeah it's good for persistent bugs, not centering a div

[supple swan]

whats the limits for codex on the free trial tho? i have access to gpt 5.4 xhigh which feels illegal on the free tier

[supple swan]

we can use kilo code that gives access to qwen 3.6 plus without any rate limits, open router is good but the rate limits on the free tier are deafening

[orchid plume]

personally I find no noticeable improvement in intelligence with xhigh, possibly a slight regression as it may overthink. It noticeably consumes more tokens and takes longer to complete tasks though. I just use high

Also the benchmark at https://voratiq.com/leaderboard/ aligns with my perception, high overall appears to be better than xhigh. Give high effort a try and see how you find it

[bright swift]

My subjective feeling is I like the code that xhigh writes more. But with the usage limits as they are now.. 🙁

[dawn seal]

codex made this

[bright swift]

Testing out 5.4-high for implementation and 5.4-xhigh for planning/research today, I dont think the difference is that big.

[nocturne folio]

😭

idk what happened to gpt5.4

i swear its overglazed

[hard drum]

it is

[nocturne folio]

it hallucinates way too much

[hard drum]

you using gpt5.2?

[nocturne folio]

yea

[nocturne folio]

with pro?

what are you doing to nearly use all your weekly limits

[hard drum]

this with https://github.com/xsyetopz/openagentsbtw

Yes, pro

[nocturne folio]

osu?

the circle clicking game?

[hard drum]

osu!droid over to osu!cocoa (android to ios)

i bought an iphone 14 just to do ios development lol

i've never used an iphone before

so...

i now have the entire mainstream dev roster

mbp 16" w/ M1 Max (apple silicon)
mbp 15.4" w/ i7-4870HQ && R9 M370X (intel -- runs W10 on bootcamp) -- macos + windows supported natively
samsung s24 (android 16?)
iphone 14 (ios 26.4 but 15.6 set minimum for dev work on Xcode)

my main is the samsung, though

and the m1-max mbp

[crisp badger]

What is the MCP codex_apps ? It seems to try to boot without me having it configured 🤔

[cyan wing]

Ya that’s so annoying.
And u can’t switch model when its doing that
so you have to sit there and wait for it to boot 😐

[glacial shadow]

Nice

[bright swift]

[features]
apps = false # buggy, burns tokens

or open the codex APP and go to settings, disable there (github app is enabled automatically). apparently this carries over to the cli and burns a lot of tokens with the mcp tool descriptions

[wild meadow]

Is this only if you have the codex app downloaded?

[bright swift]

idk, might also be a default setting otherwise, havent look at the code. probably lives in the codex app server somewhere

[wild meadow]

jw - i never downloaded the app, just use via CLI or VScode... how can i check for that?

[deft gyro]

you are strictly incorrect

look at the benchmarks

[bright swift]

i'm not sure but i think its also in the 0.117 cli

"The app-server-backed TUI is now enabled by default, and the plugin/app rollout flags have been flipped on in normal builds."

https://github.com/openai/codex/releases/tag/rust-v0.117.0

[chrome raven]

xhigh is used for all benchmarks

[chrome raven]

for me xhigh performs so well for me, i dont even change level, i set xhigh and forget

[bright swift]

just ask codex what tools it has access to. if it talks about a bunch of github mcp.. you're pwned. try the apps = false flag in config.toml and ask it again

[deft gyro]

ok so you dont know

you dont have data

and you dont know what you're talking about

[chrome raven]

why would they run evals on high when xhigh is there?

[deft gyro]

do you look only at first party benchmarks?

why would you ever look at first party benchmarks

"5.4 xhigh is overall 2.1% more accurate while consuming double the tokens. High is ~0.3% more accurate overall than medium while consuming ~33% more tokens"

[bright swift]

where is that quote from?

[chrome raven]

https://openai.com/index/introducing-gpt-5-4/

[chrome raven]

i dont understand you o, if you love using high then so be it,

[deft gyro]

I dont love using high

I primarily use medium for most tasks

and xhigh I use on gpt 5.2

because its good at long horizon tasks

[bright swift]

2,1% for double the tokens oof maybe i shouldnt use xhigh all the time

[chrome raven]

then so be it, no need to argue over it

[chrome raven]

then good for you

[thick drum]

whats this token stuff?

I havent figured it out yet

is there a decent read?

[deft gyro]

are we so fr

[obtuse river]

It's the data bits an LLM uses, like words but a word can break up into more than one token, and punctuation can also be tokens.

[deft gyro]

you are just objectively incorrect

[chrome raven]

that in your own view

i know what xhigh has done for me, so you have your view and i have mine

[obtuse river]

Usage limits are often counted by tokens processed, and bigger models that use more resources server-side often use more of your usage limit.

[deft gyro]

thats

not

how

it

works

[boreal holly]

Yeah, statistically if it takes gpt-5.4 medium to do something in 2 shots, when it takes 5.4 xhigh 1 shot, it still costs 33% more tokens to do it with xhigh.

[thick drum]

token bs and usage limits are blah

[deft gyro]

what

like

two shotting something on medium

takes less time

and costs less

than oneshotting on xhigh

[boreal holly]

That's what I'm saying

[bright swift]

does gpt-5.4 still have that adaptive reasoning thing that 5.1-codex or whatever had, on medium?

[boreal holly]

xhigh and high are diminishing returns for this model

[deft gyro]

high makes sense if you need high exploratory because gpt 5.4 already seems disenclined to do that and med made it slightly worse in my experience

like comp to 5.2

and 5.3 codex

[boreal holly]

Think about it like this, reasoning effort is the upper limit on reasoning. It still has the ability to choose less reasoning. xhigh is no upper limit at all. It could literally reason itself from 100% to 10% context window in one reasoning shot

[deft gyro]

gpt 5.4 tends not to looka round

although robert

could u link the bench u reference for thinking

I remember looking at one

but I cant seem to find it

[pure gull]

After having used Claude code on web, came to Code with high hopes but

• you cant select latest models on codex web
• you cant directly commit in git repo from codex app
• you cant select model but cant directly commit in git repo

Whats the best way to use Codex where we get all the options available at run time ?

[deft gyro]

in the codex desktop app there is a commit button

#justaskcodextodoitforyou

[boreal holly]

@chrome raven linked to it. The introduction to GPT 5.4, in the SWE-Bench Pro stats, the difference between medium and high in terms of accuracy is 0.3% at the cost of 33% more latency: https://openai.com/index/introducing-gpt-5-4/

But I actually prefer to look at this one https://openai.com/index/introducing-gpt-5-4-mini-and-nano/

Latency, they measured to include tool call time, which doesn't really tell you tokens or requests. In the introduction to gpt-5.4 mini/nano writeup they have an Accuracy vs Cost measurement, where cost is directly correlated to tokens.

[deft gyro]

yeah it was this graph that I saw

I wish codex let us use nano by default

[boreal holly]

In any case, the difference between 5.4 medium and high is so small in terms of accuracy it makes zero sense to use high. And xhigh, I think that's only really useful if you want the agent to invent new physics formulas and not write any code.

[deft gyro]

ah yes more context bloat and compactions please

I use xhigh for plan mode occasionally but not with 5.4

I usually use 5.2 for that

[boreal holly]

Oh yeah, reasoning tokens do not get used in compaction, so all of that thinking they did to reach 10% gets wasted on the next agent. Reasoning tokens are ephemeral, and only useful for the agent producing them

[deft gyro]

like ive had 10 dollar single prompts with gpt 5.2 before

[chrome raven]

Exactly

[chrome raven]

exaccctly

[deft gyro]

what

these

disprove your point

[boreal holly]

Maybe Niffylord is using xhigh to invent new physics formulas which might be an ideal use case

[chrome raven]

thank you, i dont know why this is hard for him to understand, i am not coding with xhigh.

[bright swift]

codex does these little updates though while thinking, so not completely wasted

[deft gyro]

my bad for thinking people in the CODEX chat the CODING harness were CODING

im sorri

in a bit of a bad mood

had a frustrating day so far

[boreal holly]

Commentary, sure, but all agents do brief single sentence commentary. The 200k reasoning tokens they made between commentaries disappear into oblivion during compaction, as in not included in the mental state blob at all for the next agent to gleen from

[bright swift]

Interesting i didnt know they werent part of compaction

[deft gyro]

why no fast mode on gpt 5.4 mini

:(

[boreal holly]

Yeah, the way it works in the codebase is it replays the rollout log to /v1/responses/compact, which is I guess a special agent designed to produce mental state artifacts that normal models understand. The rollout log does not include reasoning, only user messages, tool calls, commentaries, etc. Reasoning is purposely excluded from chat history for security reasons or whatever. So unless the agent is dropping extremely insightful commentaries and not just "I'm rerunning the test to confirm if the problem is gone", all that thinking is wasted

[bright swift]

You know if it includes tool call results? I'm a bit worried about poisoning context with posttooluse linting. In my opencode plugin I prune them from context myself. Might have to do that in my codex fork too

Probably need to check it directly 😄

[boreal holly]

You know what, I guess they updated it since the last time I looked https://developers.openai.com/api/docs/guides/compaction

The returned compacted window includes an encrypted compaction item that carries forward key prior state and reasoning using fewer tokens. It is opaque and not intended to be human-interpretable.

[deft gyro]

in the darkness

[boreal holly]

But this only applies if you send reasoning tokens to the endpoint. codex the CLI/IDE/App does not store reasoning tokens in the rollout log, and it sends that rollout log to the compaction endpoint, so the way it works in codex it will toss reasoning tokens.

[deft gyro]

the trees are full of starlight

[bright swift]

So could potentially make better use of high/xhigh with a fork, hmm. I put codex on some self-investigation 😉

Seems like reasoning does go into history and compaction, and some older tool use results are already automatically pruned.

But what the endpoint does with them server-side is another question.

[steady vigil]

don't kid yourself - nothing is free or without rate limits. qwen is pretty terrible in my experience. Rate limiting and just doesn't work at all for anything that isnt super simple

lots of this and 429

[obtuse river]

Isn't it also training on you always? Or is that DeepSeek, or both?

[steady vigil]

its a chinese model ofc its stealing everything it can

[cedar skiff]

the free openrouter is 1k requests so it last 1-2 hours a day if you use it all up.

[obtuse river]

I mean, the US companies aren't much better if any, but I already have enough strikes.

[cedar skiff]

I'm not worried about them training on my data

[steady vigil]

yeah. but the are the only ones that 'just work' right now

GLM-5.1 : totally broken. GLM-5: nice but too slow

[cedar skiff]

the mimo models were pretty good

well they still are but they arent free any more

[steady vigil]

Ive heard that and I think I only tried via chat not agentic

[steady vigil]

me either

if its that sensitive you shouldnt trust anyone

[obtuse river]

The individual effects on me aren't what I'd be worried about.

[steady vigil]

what is your concern?

[obtuse river]

AI x-risk long-run, inequality escalation short-term.

[steady vigil]

I mean lets face it chinese companies dont have the hw resources the usa have, so they have to be innovative. and they are doing that pretty well

[obtuse river]

Indeed.

[tiny musk]

Hiya guys
My codex takes about 3minutes average to respond to basic questions that aren't even overly complicated
It's a bit annoying... anyone else experiencing this? Anyone know how to fix this delay?

[obsidian lotus]

maybe not the best place to ask question like that but anybody has experience and can tell if codex works well with C++ code? I get a lot of anecdotal evidence of codex doing either well or badly... depending on technology etc

i will probably go for some paid plan of coding agent in a month of two... claude looked like well rounded option that isn't exactly "bad" at anything but now there is absolute clown fiesta with usage limit problems of claude subscriptions

so it looks like opencode subscription plan with GLM model or codex are the only sane options left for "home use"

[boreal holly]

Yeah if you make it unit test its code and provide it with debugging capabilities it does just as well with C++ as any other language.

[obsidian lotus]

i consider using it for Unreal Engine game development so that may be complicated... 🤣

sorry if i should have provided more context in the first place

i saw bunch of opinions about LLM's generally having worst time with languages that are not exactly mainstream or are not for web development

because there is less focus on making stuff work right in these areas and there is less emphasis for providing training data there

[boreal holly]

Debugging maybe outta the question, but testing I think you can set up a harness for that. The models know and are trained on C++, it's just C++ gives you a lot more opportunity for footguns than most languages so testing is going to be a hard requirement

[cedar skiff]

In any significant project you end up with some edge case to handle where the agent has a gap or does something wrong the same way every time. It's a bit of an art to handle. Every solution is just a few words away, and finding those words while being efficient isn't always intuitive.

[delicate gulch]

Error running remote compact task: stream disconnected before completion: error sending request for url (https://chatgpt.com/backend-api/codex/responses/compact)

Anyone else getting this?

[unique dew]

@delicate gulch I've been getting that all the time

My codex session keeps trying to reconnect session and always repeat the loop from the very first

It's been pretty unusable for the last couple days

[slow sandal]

i got a error, that his context size is to small to compact: ```■ Error running remote compact task: {
"error": {
"message": "Your input exceeds the context window of this model. Please adjust your input and try again.",
"type": "invalid_request_error",
"param": "input",
"code": "context_length_exceeded"
}
}

■ Error running remote compact task: {
"error": {
"message": "Your input exceeds the context window of this model. Please adjust your input and try again.",
}
}```

Does nice web companions exist for codex? so codex can visually show and present me designs or ideas. where i see the design boad, the todo list or the status of the tasks?

[frosty zealot]

https://x.com/iruletheworldmo/status/2040508733944668408?s=46

is this confirmed anywhere else

[cedar skiff]

lines up i guess, apparently they called the big daddy model spud. So i guess they are distilling that out into what they need.

[delicate gulch]

Ugh

[turbid axle]

my usage jumped to 0%, no warnings or anything. normally takes me all week. 🙁

stealing my token for AGI I see

[tacit parrot]

complaining so that we get our mid-week credits reset I see. keep it up, we need you!

[deft gyro]

mid week resets are actually bad

tbh

hot take

i would much rather have a better weekly usage limit

than getting weekly resets whenever

also it would be cool if there was a usage multiplier so like if you are doing stuff off peak hours it only consumes like 0.75 or 0.5x

[obtuse river]

Knowing these corpos they'll do the opposite.

[delicate gulch]

Yeah. I've got a lot of automations and long running stuff I would queue up for off-peak if they gave me any incentive to do that.

I'm going to start running everything on Monday mornings just to spite them.

[obtuse river]

Malicious compliance is always a W. XD ❤️

[cedar skiff]

to what end?

[obtuse river]

Messing with "Open"AI is a W by default at this point. 😛

[turbid axle]

Its a bit overkill though. I have pro account, and I eat through it all in 3 days, building only a single app atm. that is nuts. I used make multiple things at the same time and would barely touch my weekly rate. 5.4 fast must be really expensive

[hard tulip]

They raised like a bill i think and then also there public so they get money from that

[ivory zodiac]

Hello my people

How goes it

[boreal holly]

Bro, 96 hours later and my orchestrator is still maintaining an unyielding and adversarial stance on all code changes. I flippin love 5.4 🤯 first time going 3 days non-stop and finding real forward progress, waking up in the morning to 30 PRs and stuff getting done

[turbid axle]

sounds like gpt6 is going to be chaided

[deft gyro]

yeah i mean gpt 5.4 is super contrarian so if that works for your usecase it seems like an excellent choice

[nocturne folio]

doesnt have the autism that gpt5.2 have

[boreal holly]

are you saying I am super contrarian? Because I love everything it's doing lol

[deft gyro]

lol no xD

but like

if you want llms to be able to say that other llms are wrong

i guess then it being contrarian is very good

whats your token cost?

/how many subs

[boreal holly]

~6 agents 24/7, and a couple agents in other projects

[deft gyro]

on a 200 dollar I presume?

what if on top of compaction

LLMs could call a tool to search thru the chat

for keywords

that would be cool

[turbid axle]

if that is pro something is really wrong with my account. I eat mine in 3 days, using just 1 agent, only during daytime

[deft gyro]

do you use fast? Do you use 5.4 xhigh?

also do u use cloud or local?

[turbid axle]

5.4 high fast pretty much for everything. I'm guessing its fast

local

[deft gyro]

thats odd

[turbid axle]

that what I thought too

[deft gyro]

/fast should only be 2x usage

and 5.4 high is ntb token cost

[turbid axle]

I used to barely touch my weekly budget, now It gone super fast

[deft gyro]

and to confirm you are a pro 200 dollar/month

not a plus

[boreal holly]

Fast is gonna bite you, but you're welcome to compare your configs to mine: https://github.com/robertmsale/.codex/blob/main/config.toml

I have a really conservative configuration, v0.116.0, and some heavy optimizations so I can fan out swarms without eating the quota

[turbid axle]

it also says it will reset 8th of april, even though the week ends 5th.. I am confuse

yeah, its probably due to fast.

thats the main change in how I work

[deft gyro]

weekly resets are based on the last reset day and when u started to use

[turbid axle]

yes pro

[boreal holly]

I thought losing the 2x until April 2nd thing was gonna hurt more than it does. I think if I'm at 50% weekly by tomorrow at noon I'll be on track to reach 0% before reset time

[deft gyro]

yeah

the 2x

ntb

well

I have like 180 dollars of credits left

so im gonna burn those b4 upgrading subscription

I might move to claude code if they fix their darn quota

cos id rather do a 100 dollar sub than switch between 5 20 dollar sub accounts or pay for a 200 dollar sub

I cannot justify 200 dollars per month right now

[turbid axle]

Its probably the fast use. we had a lot of bonus tokens recently, prob got used to that

[boreal holly]

The only reason I didn't to Plus + credits is because the 5 hour limit is so small you can't spawn a whole bunch of agents. I used to hit the 5 hour limit in like 1 hour or less lol

[deft gyro]

I dont have enough mental bandwidth to do a bunch of agents

when I need to get lots of stuff done I just turn on /fast and pray

[turbid axle]

I just did not want to bother with things, but from what I understand getting multiple plus accounts is the cheapest option, oddly

[deft gyro]

yes

I think

unless you want 5.3 codex spark

I could see that low latency being useful

[boreal holly]

I only really interact with 1 agent in each project (orchestrator). The workers are so sandboxed they can't do any damage. Worst they can do is waste time. But if 1 out of 6 agents drift, that's still 5 agents that made forward progress so it's worth it

[turbid axle]

anyway, grapevine says agi in 2 weeks, so I guess we can just spawn entire apps for 5 bucks

[deft gyro]

we already have agi

imo

like

opus 4.5 was agi imho

it can simply do whatever single thing you ask it to do

whatever that thing is

[turbid axle]

I don't think so until it can properly learn.

we have human level intuition machines now, but real work needs more.

[deft gyro]

artificial general intelligence

idk

I wouldnt say learning is an important thing

it is generally intelligent

[turbid axle]

but at this point more glue and optimization will get us very far indeed

learning is key for multistep tasks. many real world tasks simply are not solvable without the ability to gain deeper understanding

loading md files into the context can only get you so far

and it makes the bots crazy

but oai is working on it I guess.

research requires this ability. its basically the missing piece for research level agents imo.

the reasoning is there, it now needs to use its reasoning to find more reasoning patterns and absorb them so it can dig ever deeper into these reality patterns

[bitter ether]

peak vibecoding

[obtuse river]

#Techdebt

[bitter ether]

oh yea brother

[tall zodiac]

Do you guys run your unit test suite sequentially or in parallel?

[deft gyro]

testing?

I only do 3 things

ask codex to do thing

git commit -m "stuff"

git push

👍

[tall zodiac]

💀

[signal tapir]

Is it just me, or has codex use started being much more limited?

I'm on business, and the last few days it feels like my rate limits remaining drops like a rock.

[nocturne folio]

its cuz the 2x is gone

or ur usage may have increased without knowing

[signal tapir]

Damn.

What's the most cost effective option now? The 2x was putting me on enough to work ok.

[frosty zealot]

Second account

[signal tapir]

Yea, I just came to the same conclusion.

[potent mason]

I have a PR coming with 150k lines of code xd

[bitter ether]

can i ask what you could possibly be cooking in that kitchen?

[frosty zealot]

lol should find my PR when I refactored my entire app

[potent mason]

Changing language for a big project (testing some things out to see if it’ll work out)

[hard drum]

i am so doomed

openai, pls roll my weekly back haha

not even a 200 sub can save me

[hard tulip]

you got all this left, back to work

[hard drum]

copilot is trash, though

and my agent framework does not support copilot

maybe i should add support for copilot next tiem

[fading spear]

really need some reset relief after the 2X loss 😒

[karmic shadow]

new 5h limits are a joke

canceling my sub soon

[fading spear]

5h limit on top sub shouldn't be a thing imo. I can't see how it would affect them since the 1w limit averages it out.

[dawn seal]

how do people run out of usage limit

do you just ask for it to build next google?

[plucky halo]

By using it a lot

[turbid axle]

its ez these days. just single agent app dev with pro account is gone in a couple days

I have it running pretty much every waking hour though.

[dawn seal]

do people make it run auto?

[plucky halo]

I have pro subscription and will regularly hit limits. But I use it in my full-time job

[turbid axle]

can go pretty quick, especially if you use subagents and fast mode etc. I need cheaper smartz

[orchid plume]

I'm on Pro and use it quite a bit too, generally every day of the week, GPT-5.4 (high). Never have managed to hit the weekly limit yet. That said, I do use other things that may help save on tokens, such as rtk

[white furnace]

did they remove skills or smth, i cant find it anymore

it shows just New Thread, Plugins, Automations for me

[turbid axle]

ok, something deffo broken. I just hit my full 5 hour usage limit with just 1 prompt, thats a new one

[winged depot]

It's skill issue.

[warm pilot]

IMO: mostly wrong prompting, I've seen people continuing threads for 20+ messages, always trying to "fix" a bit more; this fills up context/usage quickly.

[turbid axle]

im pretty sure they are tweaking usage atm. this feels much much faster than normal. maybe need more compute for agi release

I was able to do a full week with just plus if I used it normally. now I can barely make a couple days with a pro account.

[dawn seal]

that explains literally nothing

[dawn seal]

this might be? I think so too. I think people with no skills just keep requesting "fix" and the codebase just grows crazy and the AI has to read everything

[plucky halo]

You’re trying to assert that the only people who are reaching limits have it on “auto” or are trying to “recreate google”.

I’m explaining through normal use, it’s easy to hit limits.

[dawn seal]

but I heard claw is like 52thousnads lines of code or something

so... I dont know

[plucky halo]

“I don’t hit limits so there that do must be doing something wrong” 😂

[dawn seal]

what I am saying is of course you use pro
of course you use it for your personal use or job
those informations are not adding any value is what I mean

it's pretty much "duh"

[warm pilot]

LOC isn’t a big problem if the code is still structured and doesn’t eat up context just trying to find the relevant part. It’s more problematic if humans still interfere or if it’s a legacy codebase (still IMO)

[turbid axle]

its just a case of use, use it a lot, use expensive models, etc, you will run out quicker, and it can scale dramatically. that said, there is no good way to actually see your usage, its all abstracted away. probably so they can dynamically adjust when they need compute etc. if you want hard numbers you should use api. but its way more expensive

[plucky halo]

You literally asked how people hit limits. My reply includes context and adds a bit more than “by using it…”

[dawn seal]

lol...

[warm pilot]

I’m using pro for my job, personal things, side projects and the occasional dumb thing. I’m not reaching limits.

[dawn seal]

I hope you are just massing with me cuz that's literally trolling level of answer xD

[dawn seal]

me too

[turbid axle]

same, but atm I ran out in just 3 days, normally I barely touch my week limits, unsure if I did soemthing weird or usage changed

[plucky halo]

What kind of answer are you expecting? Different people have different uses. Who’d have thought. But by all means keep circlejerking around the fact that you must be a better user than someone else because you don’t hit a limit 🤷‍♂️

[warm pilot]

2x ended 2 days ago.

[warm pilot]

I think they are asking for use cases where the limit gets hit. Maybe we’re just missing something.

[turbid axle]

yeh, that is my guess too, a combination of lots of fast usage, and getting used to the 2x bonus, which all skewed my expectations

[dawn seal]

He was asking what specific usage patterns burn limits fast.
“People use it a lot” is too vague to be useful, and getting defensive made it sound worse.

A better reply would’ve been:
long coding threads, heavy agent use, expensive modes, and repeated edits can drain limits quickly.

Even shorter:

Your answer sounded dismissive, not informative.

He wanted specifics, not “different people use it differently.”```

I hope it helps

[warm pilot]

I switched back to non-fast, but I think that I could have left it on. I’m far from hitting any limit.

[turbid axle]

that said, I switched to a plus account I have laying around, and a single prompt used the entire 5h budget, and 20% of the weekly budget. that seems steep. just regular 5.4 high

I do have a feeling things changed more than just 2x rate, or something is borked

but lets see next week

[plucky halo]

What nonsense. Look at your other messages. You didn’t want legitimate conversation or answers. You instantly started to deride others as “people with no skill”.

Have a conversation with your LLM if it gives you the answers you want to hear

[dawn seal]

You’re still arguing against something nobody said.

The point was never “people who hit limits are better” or “worse.”
The point was that your original answer was too vague for the question being asked.

were asking for examples of what kinds of usage burn limits fast. Once people gave specifics, the conversation became clearer immediately.

So this isn’t about “wanting answers I like.”
It’s about wanting an answer that actually addresses the question.

[stable obsidian]

What is going on how come gpt5.4high burns up my complete 5h limit in just 1 prompt not even using 2x

[plucky halo]

You: “how do people reach limits”
E383838: “wrong prompting”
You: I think so too… people with no skills just keep requesting ‘fix’”

Hardly arguing against something nobody said when you literally said that

[dawn seal]

Yes, and that still does not change the point.

My issue with your reply was not that you disagreed with me.
My issue was that your answer to the original question was vague and then you got hostile when people said so.

The discussion got more useful the moment people started giving actual examples like long coding threads, agents, fast mode, repeated edits, and large context.

That is what I was asking for.
Not “people use it a lot.”

[plucky halo]

You weren’t asking for anything. It was purely rhetoric then you went straight in to attack others who may use it more than you.

[dawn seal]

No, I was asking why some people burn through limits so fast, and your answer to that was vague.

Yes, I made a separate comment about bad prompting. That does not magically make “people use it a lot” a useful answer to the actual question.

Once people started giving concrete examples like long coding threads, repeated fixes, fast mode, agents, and large context, the discussion became actually informative.

That is the difference.

[plucky halo]

Hahaha. I’m not arguing with an LLM - keep copy and pasting the replies.

[dawn seal]

You can keep pretending that instead of addressing the point.

Whether I wrote it myself or got help does not change anything I said.
Your original answer was still vague, and the conversation only became useful once people started giving specific examples.

So if you want to respond to something, respond to the actual point.

[warm pilot]

It’s an opinion/thought. I never said it’s true or can’t be different, it’s just what I observed and think.

[plucky halo]

Because you don’t say it. You fed it into an LLM and pasted the results back. The only original thoughts you had is that other people must be using it wrong because you personally don’t hit limits. The others must be “people with no skill” or must be “asking it to build the next google”.

Now you’re using an LLM to try to spin your original comments into something that wasn’t as condescending as you originally intended.

Your LLM also quotes your own/
its own answers back to yourself claiming that’s when ‘people’ started being more constructive.

[dawn seal]

Wonderful

Can someone give me their experience of using a coding agent? How is it different from Codex?

I find Codex amazing as it can build an amazing game demo from scratch and easily integrate many 3rd parties. Is this something that these "claw" coding agents can do easily as well?

[upper thistle]

if you are chatgpt plus subscriber (20$) what models do you get in codex?

[orchid plume]

everything except spark I believe, and of course Pro but nobody gets Pro in codex unless you pay API 😛

[upper thistle]

so theres 5.2 and its versions?

[orchid plume]

yeah you should see 5.2 and 5.2 codex on plus I believe

[upper thistle]

ty

[boreal holly]

It all depends on what the agent's role is. I have orchestrators that have quite literally hundreds of user prompts, but they remain accurate for weeks because their purpose is tightly scoped and the standard operating procedure never changes. Workers building features into the software, if you have em continuing for a long time they can drift.

Also it doesn't burn quota having long convos if the agent is constantly active. If they are idle for 15 minutes or more then OpenAI will clear the cache for that convo, so having an orchestrator constantly receiving small messages from workers and making quick decisions gives you a cached input bonus.

You know what does burn quota faster though? Something totally sneaky that you would never think about? Code formatters and indentation! A lotta formatters do 4 spaces and do stuff like this:

pub fn some_function()
{
^
}

Putting the curly brace on the next line, you get a whole bunch of whitespace if the brackets are deeply nested. The agents have to read that whitespace, and they have to apply patches to it. Total waste of tokens. I've seen some measurable cost benefits to enforcing 2 spaces and making max line width more than 80 characters. Also if you disable code formatters completely, they never have to re-read that section of code after formatting before applying a patch again. Little optimizations like that can make the quota last way longer.

[white furnace]

something is off with codex because it drains the usage like crazy with one single prompt now

i know x2 rate is gone but its ridiculous rn. i could use it much more before

[young locust]

many people tweeted the same

[deft gyro]

lolwhat

I use my subscription pretty aggressively

on 20 dollar

just allocate as many tasks to gpt 5.4 mini as possible

[boreal holly]

codex --version also what features out of this list do you have enabled?

[deft gyro]

when you allocate a task

use this chart

there is almost no reason to use gpt 5.4 xhigh

nor gpt 5.4 high

gpt 5.4 low is slightly worse than gpt 5.4 mini high so use mini high for those kinds of tasks

or mini on medium for a bit cheaper

of course nano on xhigh would do better, but that isnt available in default codex api

[deft gyro]

do you know if theres something on the desktop app to enable the llm to use the question tool during non planning mode

or what the config.toml would be for it

[boreal holly]

I'm not sure. If the question tool is a structured output response format then it likely only works in planning mode

I noticed that planning mode came out right around the time app-server began supporting structured outputs so I think that tool is probably inextricably linked to plan mode

[deft gyro]

although I have to say gpt 5.4 high is much better than medium if you have prompting skill issues

ah k

thx

It would just be super useful if it could just ask a question when my prompts are unclear

rather than me watching and having to steer

[white furnace]

only thing i used was skills creator., everything else is disabled by default. i use microsoft store app so it gets auto updated or smth because i dont specifically update the app myself but i see changes

[deft gyro]

god

[white furnace]

couple hours earlier it wasnt so bad rn its draining like crazy

[deft gyro]

the automod in this server is annoying

Idk why queue is the default rather than steer

[white furnace]

im saying like 7-10% on basic 1 prompt that is to do with coding but previously it used barely 1% with 6 similar prompts

[deft gyro]

do u have plus or pro

[white furnace]

plus

[deft gyro]

do u have fast mode on

and which model

[white furnace]

i dont use fast, 5.4 and high

[deft gyro]

do u have cc-usage

could you post your cc-usage for today

or just your token count

[white furnace]

cc?

[deft gyro]

if yk how to get that

[boreal holly]

Oh gotcha, sometimes the underlying https://github.com/openai/codex updates, and with it comes some undesirable side effects. For example before v0.116.0, with features.unified_exec=true, the agent would poll background terminals every 5 seconds and it would reprint the entire command output + new outputs since the last poll, so commands that ran for a long time would exponentially eat away at input tokens. So with the GUI you automatically get bugs like that when the underlying server updates.

0.117.0 had an undesirable side effect where suddenly MCP servers require approvals and stuff, which was really annoying because the workaround I came up with for the command output polling bug involved using a MCP server to make em wait for the command to finish, and suddenly every single "wait" call was requesting approval. They sometimes very annoyingly and drastically change the way it works under the hood which is why I think that's probably what you're dealing with

[white furnace]

1 prompt drained right now the following:
from 5h limit - before: 25%, now 9%
from 57% weekly limit to 53%
quickly tested it, not going to touch it rn anymore

[signal tapir]

😮

Maybe it's not just the 2x gone after all

[deft gyro]

1 prompt is a terrible metric

use

tokens

or mtokens

like

[turbid axle]

this is the 3rd I hear this now, 1 prompt eating all tokens

[deft gyro]

ive had 2 prompts drain 100% of 5 hourly

[turbid axle]

pretty sure its a thing

[signal tapir]

1 prompt is enough of a metric as a comparison, if the difference is huge.

[turbid axle]

my pro account was ded in 3 days

[deft gyro]

and that was before the halvening

I had

1 prompt

[turbid axle]

and my plus account dead in just 1 prompt

[deft gyro]

drain 60%

[turbid axle]

somthing borked

[deft gyro]

so

it would stand to reason

if 1 prompt drained 60% of 5 hourly

[turbid axle]

1 prompt was 20% of weekly plus tokens, that is insane

[deft gyro]

that it could drain 100% of 5 hourly after the halvening

[turbid axle]

15min of ai work and account is dead

[deft gyro]

is ur codex up to date

[signal tapir]

usually gpt-5.4-mini could be used without a thought of how much is left. Now it gets used up every 5h period, and the total in 2-3 days.

[deft gyro]

that is odd

I mean

im coding right now

I wont run out of 5 hourly at this pace

[signal tapir]

What account?

[deft gyro]

plus

mostly use mini

on high

or 5.4 on med

or 5.2 on xhigh

[signal tapir]

that's... interesting

[deft gyro]

but that instadrains my quota

whats ur config.toml

[signal tapir]

model = "gpt-5.4-mini"
model_reasoning_effort = "medium"
personality = "pragmatic"

[features]
multi_agent = true

holy crap! I just noticed that multi_agent

It was not there before

[deft gyro]

multiagent drains

disable

[turbid axle]

I was really suprised by the token use today, but I thought it was just me. now I see multiple people complain about the exact same issue

something is wrong

but oai is probably celebrating praise jebus atm

[signal tapir]

Where would that setting come from? An update that set it as default?

[signal tapir]

Nah, they are still losing money

[deft gyro]

responses_websockets = false
responses_websockets_v2 = false

disable these

also

[turbid axle]

thing is, chatgpt account is not static, they can throttle how they please, no actual token cost like with the api

[deft gyro]

check if its checking commands every 5 seconds

it can completely screw the context

[signal tapir]

doesn't websockets make it able to cache things?

[turbid axle]

so if they decide they need more compute for their coming agi release, they will just throttle us to bits

either borked, or on purpose, but it deffo changes

changed*

[deft gyro]

🤷 it seems to drain quota fast

[signal tapir]

Weird

[deft gyro]

I dont know why

or how

oai vibe coding their quota

[signal tapir]

"Hey chatgpt. Change the quota stuff so we lose less moneys"

[turbid axle]

tbh oai can most likely make tons if they just stop agimaxxing

[deft gyro]

wut

[signal tapir]

They are competing with other companies that do the same. Customers are not loyal.

[turbid axle]

they would get outmaxxed quickly, that is true

[deft gyro]

if oai stops agimaxxing

they will become nonexistant

and even just inference oai prolly loses money on codex subs still

[signal tapir]

I just wish I could get gpt5.4-mini capability offline. I would spend unnecessary money for that. 😛

[jolly lily]

Does the Plus plan actually help with the limits, like is there a good difference compared to the free plan?

[deft gyro]

free plan?

that still

[turbid axle]

thats true, but they can make bank before that happens. its a choice

[deft gyro]

is a thing on codex?

[jolly lily]

Yup

I've been using it like a wek or so

[deft gyro]

there is

[turbid axle]

you get 0.1 prompt with the current usage

[jolly lily]

So any idea? Like do you actually hit the limits fast?

[turbid axle]

plus is much better is every aspect

[deft gyro]

kimi k2.5 thinking is above gpt 5.4 mini on high

[jolly lily]

From your perspective, how are the limits?

[turbid axle]

much faster, more tokens, smarter, etc

[deft gyro]

how much do you code

[signal tapir]

Does it actually work as an agent in vs-code, or does it just provide code?

[jolly lily]

Like 2 hours a day as an indie game developer (student)

[turbid axle]

I used to be able to code all week, if I did regular human hours of work, single agent, etc

now I just have it run non stop essentially, so I need pro.

[deft gyro]

you will be fine on a plus sub if you are smart about it

imo it is worth

[turbid axle]

2 hours a day is totally fine with plus

[jolly lily]

Like it literally spends a few % of the weekly limit each time I write a prompt regardless of what it is

[deft gyro]

not totally fine

[jolly lily]

I've gotten sick of it lately

[turbid axle]

you can prob do 6 a day if you dont use multiagent fast hyper mode

[deft gyro]

wut

no u cant

as a plus user who allocates time efficiently

well

I suppose if you do 6 every weekday

just youll really be hitting that limit

[jolly lily]

I do not plan on fully-automating my workflow, I apparently am actually capable of writing smart prompts

[deft gyro]

then yes

you will be fine

[jolly lily]

https://tenor.com/view/discord-scam-users-obvious-phishing-gif-15812245174444424671

[turbid axle]

youll have to use medium probably etc

[jolly lily]

Thanks!

[deft gyro]

have to use medium?

dude

medium

is literally

worth

its the best price/perf

and its not even close

[jolly lily]

I've used medium forever and it works perfectly honestly

[turbid axle]

medium is great yes

its the best daily driver

[jolly lily]

So 2 hours a week is likely to be fine as a student, right?

[deft gyro]

refer to this graph

ye

2 hours a week

[jolly lily]

*2 hours a day 😭😭

[deft gyro]

if ur using 5.4 medium and 5.4 mini on high

2 hours a day is good

you might want to buy like 1 set of credits per year

[jolly lily]

I use 5.4 medium all the time so I assume I'm likely to be fine, thanks!

[cyan wing]

and mini is like way faster too

[turbid axle]

a day will be fine, you should have plenty tokens for that from plus

isnt there a student deal also?

[deft gyro]

no

[turbid axle]

thought so

[deft gyro]

you get 100 dollars of free credits if youre a student tho

[turbid axle]

thats a lot

[deft gyro]

yes

its a ton

[turbid axle]

anyways, if we have to believe strawberry man its agi in 2 weeks

[deft gyro]

wut

weve had agi

since opus 4.5

but

ok

[turbid axle]

agi+ then

lol

[signal tapir]

have they went and defined AGI now?

[deft gyro]

no

but

like

[turbid axle]

AGI is undefinable

[deft gyro]

opus 4.5 is generally intelligent

[turbid axle]

just like intelligence, its just, you are more smartz

[deft gyro]

some people think agi = asi

[jolly lily]

^

[turbid axle]

when ai starts going things we dont understand anymore, we will be like, wow you smarts, you agi

[deft gyro]

no

thats asi

[jolly lily]

ASI is what is assumed not to be possible in the near future no?

[deft gyro]

asi is like

not possible

imo

but very unclear rn

[signal tapir]

lol, I just looked at the requirements. At least a TB of vram. 😂

[jolly lily]

I agree on it being impossible

[deft gyro]

maybe tomorrow google drops a paper saying well we might have maybe developed asi

[deft gyro]

yes, but the gemma 4 models are highly capable

[turbid axle]

I think its the same thing essentially, buts thats the impossible definition bit.

imo agi needs to be able to learn on the spot, do actual human level research etc, but once you have that, it will essentially instantly self develop into asi

[deft gyro]

gemma models have always been insanely efficient

and the latest are a continuation in that trend

[signal tapir]

I tried using one as an agent, and it didn't work at all

it kept putting chain of thougts into the documents it was suposed to edit

[deft gyro]

right now

[deft gyro]

for gemma 4?

[signal tapir]

I think so, yes

[deft gyro]

I mean

there are people who are using it in opencode rn

idk abt u

or what u did

[signal tapir]

Maybe I did something wrong

[deft gyro]

or what they did

anyways

gemma 4 26b a4b is insanely capable

for a 4b activated

moe

[signal tapir]

I will check it out

[deft gyro]

its about equal to gpt 5.1 on high

[signal tapir]

thanks for the suggestion

[deft gyro]

and you can actually like

run it on your gpu

rather than having to have a $20,000 setup with 10 5090s

[signal tapir]

I have to settle for one 5090, until I have success with a game. 😛

actually, 5080

I'm investigating now. If I only need openai for the occasional heavier problem I'm happy doing the rest locally

[steady vigil]

in what? I found it completely useless in any agent coding... opencode, ollama

[deft gyro]

did u enable thinking

[deft gyro]

which quant and specific model did you use

[cedar skiff]

It just the usual hype, one a given task or bench it might be great, but its not even close at general work

[steady vigil]

b/c MLX is broken currently with this in LMSTUDIO I use
lmstudio-community/gemma-4-26b-a4b-it

enable thinking where?

[deft gyro]

it does well in llm arena

[steady vigil]

basically as soon as you ask it to scan your code base for anything it dies and goes into a loop others reporting similar things

apparently its only useful for chat currently. about what I expect from google

[deft gyro]

well 256k context window

and what quant

are you using

like

tool calling sucks if you arent using 6 or 8 bit quant

[steady vigil]

I tried with and without k,v cache

Q4_K_M

normal stuff

[deft gyro]

ok well like

then ofc it cant tool call

also

dont quantize the cache

ong

[steady vigil]

why can't it call tools? it does just badly

[steady vigil]

it doesnt matter. its the same with and without like I said

was only testing

[deft gyro]

right now the llama.cpp that lm studio uses is broken for gemma models.

[steady vigil]

yeah but its equally bad in ollama pi

will just wait a week, I think its just too new

[deft gyro]

also

moe models in general do worse at tool calling

recommend to push for the 31b if you can

[steady vigil]

31b didnt help . plus it was too slow

[blissful basin]

I notcied very strange behaviour with codex recently. It somehow looses "flow" and starts replaying to previous prompts instead to new one

[deft gyro]

this is an issue with gpt 5.4

you can prompt your way out of it

[blissful basin]

Yeah usually next prompt works correctly, it just didnt happen to me before (and i use it a lot :P)

[steady vigil]

yeah I see this all the time in 5.4

[deft gyro]

this has been seen since 5.4 released

if it really bothers u switch to 5.3 codex

[steady vigil]

it seems to forget the last one often just repeat it and its usually ok. or start fresh

you can even ask it "what did I just say?" it doesn't know

[small violet]

codex is amazing

it completed what gemini via antigravity struggled to complete for a week in less than an hour

[hexed slate]

Why no "Open in CLion" button in the drop down still? You have multiple other jetbrains IDEs already, this is annoying for doing C++ work.

[steady vigil]

looks like this https://github.com/ggml-org/llama.cpp/pull/21418

just gonna wait for next beta

[warm cedar]

sh8t limits though

[signal tapir]

One would think a business account should be at least able to work 8h a day for a few days a week. 😛

but damn, is it good

I've spent a few hours trying to get a local llm to help me instead, but I'm not succeeding even a bit

on the plus side, I could have spent the time reaching my limits.

[thorny cloud]

Is codex/plus worth it?

[signal tapir]

What would you use it for?

[thorny cloud]

Vibe coding

[signal tapir]

More specifically, what kind of stuff, and how much?

[exotic cave]

alright what is the consensus on this new rate card for the monthly plans: https://help.openai.com/en/articles/20001106-codex-rate-card

[thorny cloud]

Errr there's a ongoing GitHub project and I'm gonna use it a lot haha

But other than that.... I run my own llms

So hmm

Mostly try to fix the GitHub or add features

[signal tapir]

So I've not gotten local LLMs to work directly as agents, but gpt codex does a marvelous job. It's a bit limited though

The best you can do is probably get one months subscription and try it

[thorny cloud]

Hmmm

Cause I did try Gemini cli

It's okay? But I feel it can do better

[signal tapir]

From what I've understood, openai and anthropic are the best

[thorny cloud]

Atm deciding Claude or codex

Yea yea

The project is run on rust

I do wanna see if I can vibe code to handle and sdk

[signal tapir]

I've heard that anthropic is worse when it comes to price, so I guess how much use you get out of it before needing a larger subscription, but that's not confirmed

[thorny cloud]

How does openai work?

[signal tapir]

you can get a plugin for vscode, and use it directly there

you have a few different models to choose from, that have different limits and capabilities

[thorny cloud]

What about limit? Weekly? Or daily?

And it's not uncensored righht

I would still go back to local for that

[signal tapir]

I don't quite know how the limits work yet, so someone else has to chime in here. All I know is they recently got reduced.

No, not uncensored. But you probably don't need it to be. Content shouldn't be part of the code anyway, but rather read from files.

It will probably help you build jiggle physics, without you needing to talk about what is jiggling, for instance.

Speaking of limits, this is painful

I accidentally had it able to spawn sub agents, which quickened the pace this ran out

[thorny cloud]

It's weekly?

😭

[signal tapir]

there is a 5h limit, and a 7d limit

[thorny cloud]

So you constantly always use it up? 🤣

[orchid plume]

unless you're a free user, then it's just a weekly limit 😛

[signal tapir]

Yea, I do

[orchid plume]

limits though are pretty generous imo, well, certainly on Pro, you can get a lot done

[thorny cloud]

Pro cost like.

Gulp.

[signal tapir]

Pro is good if you make money from your work.

Easily worth it then

[orchid plume]

$200 or depends on your currency yeah, not cheap but worth it depending on your use case and effective profit I guess

next week I'm really hoping we see the prolite plan (Pro 5x, unless they change that when they release it), which currently is set to be $100/mo. That said, I had to renew Pro today so it's too late for me to consider that plan, on the bright side at least there's four weeks to get it done

[signal tapir]

I am making a game. Haven't seen income in over a year. 😂

[thorny cloud]

What game?

[signal tapir]

Don't think I'm allowed to talk about it here

Also a bit off topic

Feel free to PM me though

anyway, if it succeeds I suspect I will get pro. Doesn't need to be a huge hit for that.

[thorny cloud]

I see I see

[cedar skiff]

havent really used much this week so i get to give fast a run

[swift lotus]

Sign-in failed: failed to start login server: An attempt was made to access a socket in a way forbidden by its access permissions. (os error 10013)

Anyone know how to get past this error on the windows 11 codex app? Trying to login for the first time and getting this error

[fast iron]

I'm having the same issue trying to sign in

[elfin hollow]

does a business seat have the same usage as a plus seat for codex?

[signal tapir]

the fact that this isn't easy to ascertain is terrible

[orchid plume]

no, pretty sure it's less quota than Plus

[elfin hollow]

oh

[cedar skiff]

it sthe same as plus for usage

[orchid plume]

https://developers.openai.com/codex/pricing?codex-usage-limits=plus#frequently-asked-questions - compare Plus and Business messages there

Business has lower message count compared to Plus

[cedar skiff]

i could have sworn i got a message saying the git hub code review usage was being folded into normal usage buckets on april 2nd, but it does seem to have happened

[twin flame]

I remember talking to a guy that had 2 business subscriptions (obviously on trial) and some random plugin to automatically switch to accounts when the quota on the first one is gone so this is kinda fair tbh even though this is still possible to-do

[pale rose]

In a discussion about a python library written in C regarding data-round-trips between the c code and the python, codex started emitting chinese lottery spam at me out of the blue O_o

[orchid plume]

ah I see, but it's not very clear whether the new rate card is better value or not 🤔

[potent mason]

Well time to test out all of the changes, freaking miserable PR

[signal tapir]

that's... a lot

[potent mason]

Yeah I'm testing out how it would look in a different language, it isn't even the full migration just the concept of it

[signal tapir]

Could that be the type of task that should be handled via the API, using one of them batch things?

[potent mason]

Maybe but I'm still just using the pro sub

[signal tapir]

if it is enough.. shrugs

why not?

[cyan wing]

The mobile game gem economy applied to developer tooling. 👍

[potent mason]

Huh they changed usage?

[hard tulip]

going to change in the coming weeks

[potent mason]

Anyone analyzed what it means practically yet? (Like more usage, less usage, more targeted usage, etc)

[cedar skiff]

give it to codex to do

[cyan wing]

**Verify this yourself:**

1. Go to Codex settings → Usage → "Buy more messages." Note the price: 1,000 credits = $40. That's $0.04/credit.
2. Go to the Codex rate card (Business token-based table): GPT-5.4 output costs 375 credits/1M tokens.
3. Multiply: 375 × $0.04 = $15.00/1M output tokens.
4. Go to the public API pricing page (developers.openai.com/api/docs/pricing): GPT-5.4 output = $15.00/1M tokens.
5. They match. Repeat for input and cached input. They all match.

**Conclusion:** Codex credits are a 1:1 wrapper around API token pricing at $0.04/credit, sold in $40 bundles through a UI that never shows you the equivalence.

[potent mason]

Okok, and for the plus and pro subs? Will they be the same? I'll check it out if no one has done the analysis

[deft gyro]

they will

because

there is a rate card

and it shows

what

the rate is

of credits per mtoken