general

[wicked talon]

Also how would you rate apple playground for image making

[ocean vortex]

Well yeah like the main AI sub. Since their limits are very reasonable and their platform is very feature rich

[wicked talon]

It uses different models but how would you rate it stilc

[wicked talon]

Hmm true

I wonder when apple will integrate Gemini

Also s26 agentic ai

Seems like a security flaw

[ocean vortex]

Mostly just a thing to screw around with lmao. I have 16Pro but I barely ever use this playground

not extremely useful

[willow seal]

Will take a hell lot of time i guess

[wicked talon]

Yuh

What iOS are we on

iOS 26 probably will come out on iOS 27 ish

[ocean vortex]

Apple's thing isn't anywhere near as good as the best alternatives. It can forward requests to chatgpt but at that point might as well just use chatgpt app

[wicked talon]

True also it's bad how it asks do you want to ask chatgpt

Why couldn't they try integrate it to just do it anyway

[willow seal]

Basically it's aesthetic cover phone

[ocean vortex]

The phone itself is not bad at all. Does everything very well and it just works. But Apple Intelligence was a fiasco for sure

[wicked talon]

Apple intelligence was the worst ai I think on record

[ocean vortex]

You almost never have issues with apps not working on an iPhone, or glitching out, or having to deal with bloatware. Everything just works from the first try properly

[willow seal]

Camera wise in my opinion even samsung doesn't compare to it. Smooth af also a well used IOS will be way smoother then android. But the way it basically have same features just different fonts is money milking

[wicked talon]

I'm surprised Samsung allowed Google to partner with apple like that

[ocean vortex]

I mean... They didn't have a choice lol. Google owns the entire OS that Samsung relies on, they did not have a say.

[wicked talon]

Alot of the main galaxy ai features use Gemini nano/ ai core

[wicked talon]

True but I don't think Google can restrict Samsung

As it's open source

All Google can do is probably not allow Google pre installed apps

I think they did that with Huawei

[ocean vortex]

Ironically this seems to be changing for the worse lately:
https://arstechnica.com/gadgets/2026/03/with-developer-verification-googles-apple-envy-threatens-to-dismantle-androids-open-legacy/

Which would annul their main advantage. Not too great. Wouldn't want for them to do this even though I'm not using Android. It is supposed to be open OS...

[wicked talon]

Google started restricting side loading with there advanced protection

A account security program was restricting side loading

I would switch to grapheneos if they block side loading

Or install surveillance

Which UK government is trying to push

[wicked talon]

Also yes android main feature is open source

And side loading

[paper vortex]

Hello

[loud verge]

Why have you kept scanning on?

I keep it off.

[wicked talon]

Why?

[loud verge]

Because it always flags mod apps.

Always.

[wicked talon]

I've never been given a false flag in my life on pixel

And I've installed many modded apps

[loud verge]

Hmm...

[wicked talon]

Can't talk about piracy :/

[loud verge]

Noted.

[wicked talon]

Yuh

Against discord tos

I've only ever got flagged on my old s9

Which gave fakes almost every app

[rich spruce]

Hi, does anyone have a Sora 2 code? I already have an account. I'm looking for a code

[stray aspen]

no

go to openai server

[rich spruce]

I can't get in, does anyone have a solution? I've correctly set up my surfshark VPN

[stray aspen]

i would give you one

but last time i did i got banned for 'scamming'

[loud verge]

😭

[wicked talon]

VPN is probably blocking it

From you entering it by discords anti ddos

[spring kelp]

What's the best ai for coding?

[rich spruce]

Are you on WhatsApp? Or instagram

[stray aspen]

nah

i sent the code here

and they banned me

lmao

@rich spruce

[wicked talon]

Btw I don't think you can join there server due to the VPN

You would have to turn it off

[rich spruce]

Yes, it's blocked after the code. You can send it to me on Telegram; I don't know if it blocks it or not.

[rich spruce]

Yes, it's the VPN, they know.

I'm looking for a code; I came across a YouTube video on this Discord link

[rich spruce]

Claude 4.6 amazing

[spring kelp]

Thanks i will try it

[stray aspen]

claudius

[quiet skiff]

They added the direct mode battles back, right? Haha

[stray aspen]

really?

[ashen mauve]

anyone got any ideas for a good rp?

[hollow ivy]

roleplaying game with an AI as GM?
I can recommend a sandbox game:
You discover a hidden, ancient alien spaceship, while traveling/exploring antarctica. The spaceship has an AGI/ASI in it, which offers you to become its new (biological) pilot.
Was fun to play with Gemini.

Or you play an adventure, where you inherit a time-machine from your reclusive, genius uncle.

Claude also is really good in these games.

A third game: your character discovers an anomalously large spider in your house (attic, bathroom, or whatever)

-# third scenario not recommended for arachnophobic people ^^

[golden ocean]

https://cdn.discordapp.com/attachments/1384341249137446992/1432602472245100544/attachment.gif

[sinful thorn]

Everything but Video in direct chat 💔

[wheat onyx]

5.4?

[gleaming roost]

This is the third time Codex has reset my weekly quota to 100%. 🤔
Should I be worried?

Oh, now it makes sense.

[wicked talon]

Why is 900 people watching the itv logo bro 😭

[lucid geyser]

gpt 5.4 on arena but it sucks

[fierce kelp]

You shocked?

[fickle venture]

And free users...

[frosty lava]

where you get this info from

i don't see it on arena

[lucid geyser]

galapagos

[echo aurora]

Yes, Battles in Direct experiment was added back. There is now a Skip Button added. cc @stray aspen

[frosty lava]

yes and how do you know its gpt

5.4

[lucid geyser]

it was on design arena a few days ago and it matched gpt style, and everything

and on lm arena it says its openai

its like the instant or low thinking version though

so it sucks

[frosty lava]

okay i think you can be right but it might not be gpt at all too

wait for the real one

better to make a real opinion

[stray aspen]

thank god

[lucid geyser]

very very likely

to be gpt

[frosty lava]

yes but like you said if its a spark version or an instant version or a non thinking version then it's normal

if it did worse than gpt 5.3 codex then it can't be the thinking one

[lucid geyser]

huh

its pretty bad, just try it yourself

[frosty lava]

till now we never saw any new model being worse in capabilities than the previous version :

cause it wouldn't make sense at all

i told you what it can be, but anyway we don't know since its an hidden one

[fervent plank]

It is extremely direct and kinda funny. No fluff

But knowledge cutoff in july seems wrong

Should be something like december

[frosty lava]

we don't have gpt 5.3 yet we only have the instant version so it might be the normal 5.3 without thinking ?

or directly 5.4 but without thinking

[short sequoia]

https://sovra-mhce-lambda-lexicon.vercel.app/

[sly raven]

Quick question: why are certain models (like Gemini 3.1) not receiving new votes anymore? It’s been a long time since I've seen the numbers move. Is this intentional?

[tired crypt]

hey i wanna know something, how do people get this notification?

[thorny schooner]

Is anyone else having this issue now where it just gives me a error message repeatedly even after I locked out and everything and no I don't have a VPN on I already turn it off

[languid thunder]

Same. Was about to ask.
Also, how can we use that model? On the website i tried to direct talk and search, but it didnt show up

[vernal storm]

Do you need developer?

[ashen mauve]

well gang the A/B testing in direct chat is broke whatever you do DONT click skip or your chat is going to get bricked

[lofty frigate]

Okay look I don't know if I'm the only one having this issue when I'm using Image generation and using the nano banana pro model, And let's say I passed image of myself and make a prompt to include another figure next to me when it finishes generating it indeed creates another figure but then For some reason it doesn't generates me it generates someone who is wearing the same clothes as me but not me, Is anyone else experiencing this issue, Now if I hit try after a couple of tries it does generate me but many times it generates a random person who is wearing the same clothes as me

[vital cove]

Do you need developer?

[lofty river]

anyone know why its not letting me reset password ?

[hardy lion]

the video generation is now on the website, not in discerd. check it out at https://arena.ai/video

[lucid geyser]

Bro your in the discord

[sinful thorn]

Everything but video in direct chat😭😭🙏🙏

[rigid holly]

Ayo is it just me or does it demand for you to log in your account to use it?

[crisp mauve]

/voice

[rigid holly]

You'd think they'd make an announcement or something if they made it so you had to be logged in to use arena

[tawny brook]

i forgot the prompt, but i ran this again and look what it gave me LMAO

[bright shard]

It's required; they should add that it will let you do some things even if you don't log in. They should leave it like it was before.

[keen beacon]

Why I can't send prompt in lmarena incognito

[keen beacon]

@echo aurora

[echo aurora]

Hmm what do you mean?

Why not

[keen beacon]

i'll show you

why i cant do this

thats incognito btw

[echo aurora]

Have Video Arena selected?

[keen beacon]

no

why?

im doing a text

[echo aurora]

I’m not sure. Was able to make it work without having to sign in. Can you post a bug in #1343291835845578853 ? Add the relevant info. I can’t take a look now, but will later.

[glacial swan]

Guys, does anyone know why in Claude 4.6 I can’t upload files? Screenshots and things like that would be useful, but for some reason it doesn’t work in LMArena.

[undone saffron]

#1372230675914031105
-# Before creating a post with that proposal, check first to see if someone has already created a post with your idea

[undone saffron]

Why is it that at a certain hour, when using AI to create a video, the captcha system is bug and cannot be overcome?
Then it won't let me write on code-arena or direct chat because the captcha ruined my previous captcha token

There comes a point when the same problem [becomes annoying](#general message)

[native folio]

Im not sure they'd have necessarily advanced cutoff date between dot one models, most likely trained over the same normalized data with superior inference

[light sleet]

finally u can stop responses, seen in canary arena

[rare veldt]

hello

[runic rain]

Hi

[inland quest]

Omg

Absolute Cinema

Stop button

[inland fossil]

Hello

[whole sundial]

<@&1349916362595635286>

crypto scam

[hollow mulch]

My main account got stop genration button, but my onther account still don't have that button:(

[wicked sage]

THEY ADDED A STOP BUTTON???

:LFG:

[hollow mulch]

Yes im see it today

Well no more infnite genrate yay 😄

Gpt 5.3 code vs opus 6 who is better guy?

[manic moss]

How long time do we have to wait to have new token for claudAI free ? Or its « one shot » mb I’m new

[shrewd citrus]

Is opus 4-6-search broken?

[hollow mulch]

Now can you guy fix issue this like model sometime can't use is report 'something went wrong with this message, please try again'

[distant idol]

@echo aurora when yall making the video arena a direct use or even side by side, its been months since i asked this 😭

[hollow mulch]

@echo aurora Is he developer or discord mod?

[distant idol]

inbetween ✌

king

[bright shard]

It's a bug; if you delete the cookies and try again, the login message doesn't appear. It only appears sometimes, so I imagine it's a bug.

[fervent plank]

I am pretty sure galapagos is just the upcoming open source model

from openai

[rich spruce]

Lneduo2en ?

[ocean vortex]

how come the UI is half Russian for you?

[frosty lava]

galapagos is doing much worse than any open ai last thinking model so it can't be a thinking model:

that's probably an instant or normal model

so it will be for free user too

for sure

[fervent plank]

That what I meant, it is going to be the upcoming open source model that you can use locally on your own computer

[fervent plank]

Either that or 5.4 instant, but I it does not feel like they are releasing thatt

[frosty lava]

honestly all i care about is thinking model capabilities i don't need another instant model or low cost

[fervent plank]

Yeah I am waiting on 5.4 pro until I renew my pro subscription.

[light scroll]

Arena has updated something? Previous I was able to generate 3 images per day with no account

[ocean vortex]

it could be 5.3 Instant no? That one is still not on a leaderboard

[tiny jolt]

.txt support for when?

[humble ether]

now generate image need login first?

[tame haven]

I can't chat anymore without login. Is it time to leave lmarena?

[humble ether]

maybe

[fervent plank]

Nah 5.3 instant uses 1 emojii per sentence. This is way more concise than 5.3 instant

[wicked talon]

I like how qwen can image create anything except 18+ content

[wicked talon]

To prevent people clearing there cookie + cache to get more message allowance

Nvm I hate it it doesn't make Donald trump in a I hate America shirt

[burnt sinew]

I dont know but id rather just pick opus 4.6 every time

And to have router show me what model its using

[tired mantle]

so annoying

[ocean vortex]

it behaves differently on API vs chatgpt though. Chatgpt has lots of instructions they are feeding it.

Here are 10 API responses of it I generated for someone else earlier today (just a silly argument about whether it can find and output release date of itself accurately lol):

[tired mantle]

it's broken

[ocean vortex]

As you can see not a single emoji anywhere

No added sys/dev instructions here, default settings and a simple question with OpenAI's search enabled.

[ocean vortex]

You have no more code_execution tool calls remaining this turn

[errant inlet]

https://www.linkedin.com/company/mhn-solutions-2026/

[fickle venture]

Make a new chat, but I am pretty sure Gemini nano banana is the only one having this issue

[tired mantle]

Only this new version have this issue. 3. works fine

[fickle venture]

Well yeah you can use it on Google flow for free

[tired mantle]

With Gemini watermark?

[fickle venture]

Uhh I dont think it haves it

[fickle venture]

https://labs.google/fx/tools/whisk

Here it is

[sharp mirage]

Yo

Hi chat

[languid kernel]

hi new here

[sharp mirage]

Hi

Wsg

The chat died

Yoooo. Chat

[languid kernel]

yo folks, where is video arena channel? pls anyone can guide me to it

[unreal tide]

holy plot twist

[steep blaze]

Can't find it either

[mortal vale]

@steep blaze Note that Video Arena has been removed from the server. More information can be found in this #announcements . You can still generate videos on the website.

[vital lake]

GPT 5.4 LAUNCHED
EVERYONE

EVERYONE

5.4 LAUNCHED

[frosty lava]

yesssss its real

[vital lake]

Someone addd ittttt

[cloud zinc]

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

[vital lake]

thx

[frosty lava]

okay so they are saying its better than 5.3 codex in coding too so now i can't wait for 5.4 codex to see even more improvement in coding task

[vital lake]

It doesnt seem like they focused on intelligence this time

Like they did with 5.2

[frosty lava]

Yes this one should be as good or better but not by alot than 5.3 codex in term of coding task or intelligence that's why its not 5.4 codex

but for sure they'll release the codex version if what you mean by intelligence is coding capabilities

[vital lake]

No, general intelligence

Not coding

I value a smarter assistant more

hopefully 5.4 isnt a disapointment

[frosty lava]

oh then, your right it seems more cost efficient and should be more intelligent but they didn't focused as much as for previous model

[shell pewter]

its not on arena yet right?

[frosty lava]

still should be smarter

[vital lake]

It is

[shell pewter]

thx man

[frosty lava]

seeing the benchmark it should be still a good improvement in intelligence from 5.2

[crisp anvil]

Mine isn't working

[vital lake]

Refresh?

Or open new tab

[shell pewter]

did you (or anyone else) who has used a lot ai and tested gpt 5.4/high? how is it compared to other tier 1 AI?

[frosty lava]

knowledge cutoff seems to be august 2025

[vital lake]

I mean it just released

[shell pewter]

1 saw from X, open ai is experiencing error originally before the launch (lol)
2 openai just released it so their own server must be experiencing a high volume -> slow
3 same thing for arena itself

[frosty lava]

actually every model fail at this answer due to knowledge cutoff and what they've been trained most on

ask gemini to guess it will do same and same for anthropic

[stray aspen]

gpt 5.4 is out

[vital lake]

Well they are given like zero data to determine

So a guess like GPT 5 era is good

[vital lake]

We know

[frosty lava]

yes its good that they are saying gpt 5

its not that old

[vital lake]

It talks so natural and smart

[frosty lava]

they said that they improved the personality to be better again so

i guess its good since everyone complained about it

[loud verge]

Guys

[compact flame]

Well gpt 5.4

Is it any good?

[loud verge]

Was there no announcement for gpt 5.4 on arena?

[frosty lava]

its made for general purpose task and supposed to be good at coding even tho its not a codex model

[stray aspen]

why does gpt 5.4 not answer

[loud verge]

Chad model

[wind ember]

same performance as gpt 5.3 but cheaper

[frosty lava]

should be better or as good as 5.3 codex on coding task

[wind ember]

still same coding slops

[vital lake]

Huge difference from GPT 5.2

Is more ways then one

[wind ember]

well its def better than gpt 5.2

but not that much diff with gpt 5/3

[compact flame]

I wonder if we should expect gpt 5.5 sometime later

[frosty lava]

the pro version is beating gemini deep think right ?

[stray aspen]

gpt 5.4 sucks

we need gpt 6

[fathom apex]

yea

gpt is sucks

[frosty lava]

So now openai are doing monthly release instead of two month ?

[inner relic]

Is gpt 5.4 good in roleplay

[stray aspen]

no

it sucks

[vital lake]

They release monthly so probably around mid or late apirl.

[frosty lava]

that's crazy speeding up by 2x ?

[vital lake]

Dont listen to him its been around for like 20 minuites no one could get a good understanding of the model within that time frame.

[stray aspen]

gpt 5.4 is literally so ass

[honest verge]

Gpt 5.5 is going to come out April 5

[stray aspen]

ill just go back to claude 4.6 for the 50th time this month

[frosty lava]

me when im hating without even testing

[vital lake]

Yup, to keep up with Opus 5

[royal sail]

the model just dropped broski

lol

[stray aspen]

wdym i just tested it

[honest verge]

And gpt 5.6 may 4

[inner relic]

same guy who called glm-5 suck.

bruh

[frosty lava]

okay bro

[stray aspen]

glm-5 is actually decent

[light sleet]

Gpt 5.4 is good in lua coding 💀

[royal sail]

The model is quite better conversationally

[honest verge]

and maybe gpt 5.7 early-mid june

[royal sail]

It sounds a lot less AI-ish

[inner relic]

alright.

[frosty lava]

it talk in a better way

[honest verge]

When gpt 6 will come out

[stray aspen]

lets get gpt 6 already

no more gpt 5.x crap

[honest verge]

Because at this point it's only 5.x

[royal sail]

they're scared to call anything gpt 6 right now lol

I doubt they have anything worth calling gpt 6

[frosty lava]

so you are basing your opinion on the name they give to the model not on the actual performance, nice

knowing every other companies are doing the same

[stray aspen]

when did i say that

[tawny canyon]

what do u think ,will gpt 5.4 take the first place in arena?

[stray aspen]

i literally just said that i tested it

bro what the hell is this

even glm is better

glory to anthropic

[royal sail]

5.4's coding skills are pretty much nearly identical to 5.3 codex

especially with frontend

[honest verge]

High version is out

[vital lake]

Not everyone values frontend only.

[royal sail]

I pretty much just imagine this model to be the general-use version of 5.3 codex

[stray aspen]

claude is way better

[frosty lava]

this one is made for general purpose task.

[inner relic]

Let's test creativity writting

[honest verge]

But no xhigh...

[frosty lava]

mainly but still good at coding

[inner relic]

since chatgpt messed up eariler

[honest verge]

I wonder if arena will ever have xhigh

[compact flame]

Well if people can't afford Claude it's mid

[honest verge]

Or it's too expensive

[compact flame]

Too expensive

[vital lake]

It seems like it would shine there

[royal sail]

based on what?

[stray aspen]

no

[royal sail]

the model has barely been out for an hour lol

[stray aspen]

on my tests

[royal sail]

on your 3 prompts?

[stray aspen]

yes

[vital lake]

He has only done basic frontend tests btw

[royal sail]

congrats

we should hire you for benchmarking!

[stray aspen]

i test each model for the stuff i do

if its not good then its trash for me cause i wont give any other use

[honest verge]

Waiting for extreme thinking

[inner relic]

All models are not perfect yet

that's reality

[honest verge]

But I think it's not true

Cuz it's still not there

[frosty lava]

yes you are doing front end task and testing it on a model made for general purpose task

[vital lake]

Oh yeah

[stray aspen]

nah

[vital lake]

But I think extreme thinking is a fake leak

[royal sail]

If you form an opinion on a model after 3 tests, then I don't really have anything else to tell you

[vital lake]

Frontend should be the least

[stray aspen]

gemini 3 deepthink on arena when

[frosty lava]

atleast before judging coding capabilities wait for the coding model lol

[royal sail]

Most benchmarks run hundreds of prompts

You ran 3.

[frosty lava]

never since its from a pro subscription

[stray aspen]

well that sucks

[frosty lava]

just like we ain't ever getting gpt 5.4 pro on arena

[stray aspen]

yupp has them models tho

[compact flame]

Isn't it for ultra?

[stray aspen]

they had a 5.2 pro

i guess it was fake

[honest verge]

And also all leaks about open source gpt are fake ...

[frosty lava]

yes my bad

i mean its from the highest subscription

[honest verge]

Really hoped openai will release gpt oss 2 or something

[compact flame]

Yeh

[frosty lava]

so it can't be on arena

[royal sail]

i doubt they're going to touch oss again any time soon

[stray aspen]

for what

[royal sail]

especially with how they're doing in the public eye right now

[stray aspen]

people want deepseek 4

[tawny canyon]

i did some tests and in general texting gpt 5.4 is not better compared to claude opus 4.6 or even gemini 3.1

[royal sail]

did you use 5.4 high?

[tawny canyon]

they are sleeping

[tawny canyon]

yes

[vital lake]

They just gonna distill why would we be excited

[honest verge]

Is o3 pro better than gpt 5.4?💀

[royal sail]

oh hell naw

[stray aspen]

no

o3 pro sucks

[vital lake]

Ummmm source? 😭 🙏

[stray aspen]

gpt 5.4 is miles better

[royal sail]

bro talkin bout ancient technologies in the big 26

o3 is the past

[vital lake]

The early pro models like o1 and o3 pro were sloppy.

Since post RL was new

[honest verge]

Then why openai keeps it

[vital lake]

Its still a decent model

[royal sail]

Some labs and companies still use it

[honest verge]

o3 pro and o3 are still live

[royal sail]

not saying it's a bad model

but there are much better options now

[vital lake]

Fo sure

[royal sail]

and cheaper too

[honest verge]

And openai won't remove them until 2027

[vital lake]

Like confirmed?

[royal sail]

there's no real reason to use o3 right now unless you REALLY like the model for your use case

[honest verge]

Because no official deprecation date

[royal sail]

there are more effective options

[honest verge]

They are deprecating gpt 5.1 but not o3

[vital lake]

Something about o3 was different idk why

[honest verge]

Like why removing 5.1

But not o3

[royal sail]

It definitely marked a huge leap in reasoning

One of my personal favorite models

just far too expensive

also hallucinated everything lol

[vital lake]

I tested it in chess and it was the first model ever to not hallucinate.

[stray aspen]

the ridiculous pricing

[honest verge]

Especially pro version

Btw what's the point of using o3 pro

Like is it better?

[vital lake]

Atleast it was cheaper then o1

[royal sail]

it just thinks wayy longer for higher reasoning capabilities

[vital lake]

Who remembers o1 pro pricing? 😭

[vital lake]

Barley higher

[royal sail]

yeah lol

[vital lake]

I never saw any big improvements, but maybe I wasnt using it right

[frosty lava]

that mean every month we are getting a new gpt its insane

[tawny canyon]

will gpt5.4 search mode be added to arena?

[vital lake]

Obviously

[royal sail]

I feel like OpenAI is gonna have to release something extraordinary to get back into the race

feels like they're falling off frontier

[stray aspen]

glory to anthropic

[frosty lava]

i don't think its the real capabilities of their model that are judge by people but the fact that they removed 4o lol

[royal sail]

well

[tawny canyon]

gemini will take 1st place in may

[royal sail]

4o put thousands of people into psychosis so

[frosty lava]

exactly

and people want it back

[stray aspen]

then they wil nerf the model

[honest verge]

I used it very much but it's not very impressive now

But in early 2025 it was peak

[thorny schooner]

I still hate they for some reason added battle mode directly to direct mode

[honest verge]

There was no better ai writing

[royal sail]

I will agree that there was something unique about 4o in the sense that it just had less refusals and felt less corporate

[stray aspen]

pineapple said theres a skipp button now

[frosty lava]

people just want a model that will make them go into psychosis and agree with everything they say

[royal sail]

but it was just too sycophantic

[frosty lava]

they want the model to agree with everything they say that's it

[royal sail]

yeah pretty much

[honest verge]

Hey 4o is still live

[royal sail]

because if a model agrees with everything i say then it must be the best model

/s

[honest verge]

Not in chatgpt but it's live

[stray aspen]

yes

[frosty lava]

yeah i know

[honest verge]

So what's the point of #keep4o

[royal sail]

the people who were going into psychosis can NOT afford that API pricing 😭 🙏

that's why they still want it in the app

[frosty lava]

i don't wanna be in a world where 80% of people are in psychosis due to an ai

please

use ai for coding task or general purpose

[distant idol]

any news on seedance 2 api guys

[stray aspen]

delayed

[light sleet]

Gpt 5.4 is the best at lua code?

[distant idol]

damn, to when thoo

[mossy girder]

Why UI is scrolling down forcefully after response is done? Before it wasn't there, it's irritating.

[stray aspen]

no lmao

its horrible

claude is better

[light sleet]

what's the best for lua?

[stray aspen]

claudius

[light sleet]

k

[royal sail]

bro

[stray aspen]

@daring rock

[distant idol]

what are u doing nephew

[honest verge]

He will use ai for this 100%

What's the point of paying someone 5$

When you can just use ai

[royal sail]

I swear these guys just ask a model "how can i make money fast with ai" and do whatever it says

[stray aspen]

its a scam bot wdym

[honest verge]

Ik

But still funny

[distant idol]

how to make money fast

[honest verge]

You have to sleep less

Sleeping for 30 minutes is enough

[stray aspen]

work

[honest verge]

No you have to sell your money

So you will make money

Then sell it again

[distant idol]

real shi 💥💥

[honest verge]

And you will make money again

Do this forever

[distant idol]

im an officer, do i need to quit

[honest verge]

And you will get new money

[honest verge]

You have to sleep less

What's this...

Concord grape?

Mammoth newt 0226

[frosty lava]

where did you find this

[honest verge]

Yupp ai

[stray aspen]

why is it so expensive

[honest verge]

Idk maybe new Claude?

[vital lake]

5.4 seems smart

[frosty lava]

personal opinion but i love how gpt 5.4 talk

[distant idol]

nephew what did u even tell it

[vital lake]

Debate about god

[honest verge]

Wait mercury 2 is out?

[distant idol]

gus fring is something

[honest verge]

And it's after gpt 5.4

[frosty lava]

what's mercury 2

[vital lake]

A really fast model, I forgot what type

Somethin new idk

[distant idol]

feb and march is filled with crazy models

[honest verge]

Feb is my favourite

[meager harbor]

gpt 5.4 worst than 5.2 latest

on arena

[frosty lava]

actually no way

[honest verge]

Also qwen image 2.0 and pro is out!

[distant idol]

people when its comes to gpt 5.4 either they are glazing or hating no inbetween 😭

[honest verge]

Crazy

[inner relic]

They are focused on coding skill not creativity

[vital lake]

Not true, its a balance

[frosty lava]

for coding skill it would be a codex version

this one is for general purpose

[inner relic]

Ok

Anyways.. My one question is

[honest verge]

Gpt 5.4 pro pricing is crazy

[inner relic]

Chatgpt 5.4 is good at writting and creative?

[honest verge]

Like it's too expensive

It's unusable for normal peoples

[frosty lava]

yes but i'd say same for gemini ultra or claude max

[vital lake]

Thats why it has pro in the name

[vital lake]

This lowers my expectations on 5.4...

[royal sail]

need that jelly

[honest verge]

Then what's the point of releasing pro version in app

[distant idol]

wharr

[honest verge]

No one will use it

[royal sail]

"arc-agi-2 is reaching saturation"

yeah okay

[vital lake]

For people with money

[frosty lava]

some people are buying it actually

[distant idol]

are u rich gus

[vital lake]

I wish

[inner relic]

not bad not bad.

[honest verge]

Arc agi 1 even existed?

Or every model just scores 100% now?

In it

[royal sail]

yes

it did exist but it was saturated

[distant idol]

😭😭😭😭 he learnt from the memes

[royal sail]

I don't really like this test, because it's just absuing how tokenization works

[honest verge]

Arc agi 3 coming soon

[royal sail]

🤦‍♂️

[distant idol]

😭😭😭😭

folk

[compact flame]

Ai is Future they said

[royal sail]

Like, counting letters in a word is something us humans do fine because we have a method of doing it

LLMs don't work letter-by-letter

[inner relic]

lol

[distant idol]

i thought they fixed the hallucination of this model

[royal sail]

Not a hallucination problem

It's just how tokenization works

the model doesn't see every letter in the word

it just sees the probability of that specific token happening

[meager harbor]

just look at the arena score

[meager harbor]

look at arena score

[royal sail]

I mean, the model just came out

I'd give the score time to settle

[frosty lava]

preliminary test mean its not definitive score

you didn't saw the warning ?

[distant idol]

it js came out

calm down saul goddman

[honest verge]

WHERE IS DEEPSEEK V4

Deepseek is already miles away

[meager harbor]

don't expect the score to change much, it won't win 50 elos point magically

[meager harbor]

don't expect the score to change much, it won't win 50 elos point magically

[honest verge]

Dementia?

[frosty lava]

you are ragebaiting

[distant idol]

nephew take a breath and relax

[royal sail]

it's releasing in 1 hour and 23 minutes

trust

my source is me

[meager harbor]

why should I relax ?

[honest verge]

Are you ....

LOL

lol

[royal sail]

model just dropped

[inner relic]

Yep

[sleek phoenix]

5.8 in the next month

[frosty lava]

let's ignore the fact that it just released and the warning "preliminary"

[honest verge]

WHEN GPT 6 IS COMING

[royal sail]

The current leaderboard position is irrelevant solely because of the fact that the model has only 2,000 votes as of now.

[honest verge]

IM TIRED OF 5.X

[sleek phoenix]

next month and 15 days

[royal sail]

That's not nearly enough to make a conclusion

[honest verge]

I want next generation

[meager harbor]

score won't change that much.... opus 4.6 will still be ahead by far, I mean are you new to this ?

[royal sail]

The model only has 2,000 votes

Opus 4.5 has 30,000+

[sleek phoenix]

4.5*

[meager harbor]

can you read ? it doesn't matter, there wilkl be max 10 elo diff

[royal sail]

mb lol

[royal sail]

how can you be so sure?

[honest verge]

gpt 5.4 is o5

[distant idol]

damn 💔

[royal sail]

Models can do this pretty easily if you just give it terminal access

they'll just use a command

[frosty lava]

i don't think this one is as dangerous as 4o

[royal sail]

Same reason why LLMs can't do long form arithmetic without a calculator

[meager harbor]

experience, have been looking at the ranking for just one week ? and also next to the score there is the +-12

[royal sail]

My fault

Didn't know you had experience

[meager harbor]

arena champion bro

[distant idol]

experience 💥

[honest verge]

Wait why gpt 5.4 is only available for text?

No coding?

[royal sail]

there are 500 other arena champions

you are not special

[distant idol]

dude he is the chosen one

[meager harbor]

yeah well you not being arena champion speak a lot for itself

no it's just that you don't understand simple thing is ragebaiting me

[honest verge]

How do you even become arena champion?

[royal sail]

I have a bachelor's in computer science and I've fine-tuned LLMs myself, and I also run my own benchmarks.

I'd rather have that than "arena champion"

[distant idol]

calm done uncle

down*

[meager harbor]

ok Noam Shazeer

[ocean vortex]

yeah but it's reasonable to expect for the chat model to be tuned more for user preference than the thinking one. Very often was the case with OpenAI. So wouldn't be too surprised if this doesn't climb much. Doesn't mean that it isn't better

[royal sail]

I mainly bring it up because I think we're jumping to the conclusion that this model is "worse than 5.2" way too quickly

Once the score settles, it's probably fair to determine if the model is good or not

[tawny canyon]

guys when will deepseek v4 be released?

[honest verge]

WHERE IS DEEPSEEK V4

[inner relic]

I dont know what whale is doing

[royal sail]

today in exactly 22 minutes

[honest verge]

IM TIRED

[inner relic]

They're not afraid of gemini,claude and openai

[honest verge]

I NEED DEEPSEEK V4

[frosty lava]

deepseek waited for gpt release to beat them lol

joke

[distant idol]

they always does this

[tawny canyon]

source?

[meager harbor]

tbf the model is on par with 5.2 but it is not the big leap everyone was talking about

[ocean vortex]

It's fundamentally incorrect to align arena score with better/worse in the first place

[royal sail]

nobody was saying it was a big leap?

[frosty lava]

yes but no opensource model have beaten one of top 3 like anthropic, gpt gemini

[royal sail]

if anything people were saying it was incremental lol

[royal sail]

Of course, it just gives some insight on where the model may sit

[meager harbor]

you should see Scam altman hyping it up

[distant idol]

opensource models are 1 year back in the race

[royal sail]

well argue with him then lol

nobody here thought it was a huge leap

[honest verge]

It's almost time for them to release something

Or they are getting destroyed

[frosty lava]

but they are progressing very fast honestly

[ocean vortex]

Dunno how long you've been following this, but Anthropic used to suck on lmarena. That was kind of ironic. But it proves that top spots there can only be occupied by the models explicitly trained to do well in this specific environment

[distant idol]

im waiting on bytedance models tbh, they are cookinggg ngl

[royal sail]

Oh yeah I'm aware that LMArena has had some misleading scores in the past

hence why i take the scores with a grain of salt and try to supplement my opinion with other benchmarks as well

[honest verge]

WE NEED A PIECE OF OPEN SOURCE!

[ocean vortex]

Not really misleading, but people need some context to properly read them. Arena is one metric out of dozens of them. And it isn't really raw performance or capability metric

[royal sail]

Eh, I'd say misleading in some cases. Llama 4 topped the leaderboard at some point.

[honest verge]

Btw where is llama?

[royal sail]

dead in the back of an alley

[honest verge]

I haven't heard anything about it

[ocean vortex]

yeah because they fine-tuned on arena datasets lol

[honest verge]

They just discounted it?

[distant idol]

when is arena-video gonna be a direct chat 💔

[honest verge]

Actually I remember it was

For some hours

But it got deleted

[distant idol]

fuhhhhhh

[honest verge]

And now it's only for battle arena

[distant idol]

ive been waiting for months

its the site can really be 10/10 if they js did that

[frosty lava]

my dream would be deepseek v4 #1 in capabilities from all existing model till now but it's impossible right ?

[ocean vortex]

Many of it is just style that is technically meaningless and easily changeable, some of it are also just the mere patterns... Patterns from the most active users on arena, what prompts they use and how they are voting

[honest verge]

Gpt 5.4 is very bad for coding in arena

Like it's so bad

[royal sail]

highly unlikely, yeah

[honest verge]

I can't

Even Gemini 3 beats it

It's worse than gpt 5.1

It can't do anything

[royal sail]

literally can't be possible

It's supposed to be 5.3 Codex level

[distant idol]

patience nephew 🙏

[frosty lava]

why in the coding i only see the medium version ?

on arena

[honest verge]

It's not even opus 4.5 lvl

[ocean vortex]

there are certain coding areas it should be really be really good at tbh

[frosty lava]

front end on gpt is really bad

don't do front-end using gpt

[royal sail]

hate how openai conveniently leaves 5.3 Codex out of so many benchmarks

[honest verge]

Opus 4.6 is still my king

I'm waiting for opus 5 so hard

[frosty lava]

how can we explain opus 4.6 have good "taste" in front-end but not other companies ?

[honest verge]

Because it actually does max effort

[frosty lava]

for example when i try gpt 5.3 codex on coding task that's not front-end it do very good job and maybe better than opus 4.6

so how do you explain it

[ocean vortex]

I don't mind it since I prefer general purpose model and codex is this odd one out for me lmao. But I understand where you are coming from, I didn't even notice at first that they included 5.2-codex rather than 5.3-codex in that graph 🗿

[frosty lava]

the same gpt 5.3 codex on front end is bad

[honest verge]

Opus 4.6 is built for frontend

It's capable of ambitious work

[frosty lava]

yeah i guess that make sense

[honest verge]

As anthropic says

[frosty lava]

so openai need to work on this

its not about coding capabilities honestly but about the taste it have

[royal sail]

I feel like GPT models have always been pretty terrible at frontend

never really got good UI outputs from GPT models

[frosty lava]

yes i hope they focus on it for the next one

[honest verge]

UI from gpt 5.3 just feels too plastic

While opus 4.6 somehow makes it alive and beautiful

[ocean vortex]

to have good visuals you ideally need a bigger model. So Opus and Google gonna have natural advantage there. And if OpenAI made it bigger they wouldn't be able to sustain current caps on chatgpt. It's a reasonable trade-off. They used to be struggling considerably more with this before gpt4.1 and subsequent models

[honest verge]

What's the secret

[frosty lava]

its capable of implementing features and doing very great things like that but don't have any good visuals

[frosty lava]

i think they can train it especially on front end task to improve it honestly

[royal sail]

I wouldn't say this is necessarily true, since some small models can push out some great UI with proper instructions

[ocean vortex]

you can but only to a certain extent before degrading other things

[vital lake]

Guys 5.4 is so good at creative writing

[royal sail]

Size is only really important for world knowledge and superposition (relating novel concepts to each other)

[vital lake]

Like way better then Gemini and Opus

[hushed gyro]

WTF is 5.4???

[devout vault]

openai will never make good models. again

[ocean vortex]

"with proper instructions". But if you adopted similar approach with bigger model chances are it would do better than it did before as well. Can only compare identical setups

[honest verge]

Anthropic was always ahead

[crude lagoon]

They reached the peak at 4o

[frosty lava]

but the reason why i use gpt 5.3 codex instead of opus 4.6 is that its doing better job on every other thing than front end

[royal sail]

GPT models with design skills still do pretty horrible

Theo.t3 made a video on the whole thing

[honest verge]

And because 5.3 is very cheap

[round forge]

What kind of creative writing do you mean?

[honest verge]

While opus 4.6 requires max subscription

Pro isn't enough

[frosty lava]

yes too

[inner relic]

2nd

[inner relic]

2nd in creative writting

[vital lake]

Roleplay, story board, all of it.

[ocean vortex]

Wouldn't say it's horrible. With gpt4o and before o3/gpt4.1, that's when it was really bad lol. They sucked on just about every single benchmark or test that touched on visuals

[vital lake]

Its creepily natural.

OpenAI 100% focused in this aspect more

[royal sail]

i mean, if you enjoy AI slop blue-purple gradients then sure 😭

5.3 codex loves doing it

5.3 codex (left), 5.4 (right)

[honest verge]

Was gpt o models ever really good?

[frosty lava]

but i love the 1m context

[distant idol]

guys which model has the most context

[royal sail]

Pretty good general purpose reasoning models

[zealous sparrow]

Same lame model

[royal sail]

Not amazing at coding

But they were frontier

[honest verge]

And where's o2?

[zealous sparrow]

Its bad at simplebench

[royal sail]

skipped lol

[vital lake]

Yes

[zealous sparrow]

I did my tests

[vital lake]

Copyright issues

[royal sail]

oh really?

[ocean vortex]

on webdev arena, arc-agi and related things they do get respectable scores nowadays. They also perform decently on svg tests

[vital lake]

Yeah some company name

[frosty lava]

Chat gpt is bad at making good visuals

[weak dagger]

@alpine pasture i like cherrys

[royal sail]

was not aware lol

just basing off my usage and general consensus back then

[distant idol]

i js ate chicken

[honest verge]

Missing

[honest verge]

Or it was so good

[weak dagger]

did you like my brother?

[distant idol]

yea, im still hungry too, come here nephew

[ocean vortex]

no 5.4 on chatgpt?

😠

[weak dagger]

RIP chicken

[ocean vortex]

this rollout thing is kinda slow recently

used to be near instant

[royal sail]

Sure, but arc-agi is more about reasoning, and svg tests are just a capability benchmark rather than measuring design skills

[ocean vortex]

Didn't even have 5.3 until now

[royal sail]

i do remember gpt 5 taking some time to roll out

[honest verge]

Gemini namings are so done 🥀

[royal sail]

even though it was supposed to be huge release

[frosty lava]

because they only released 5.3 instant

there is no normal 5.3

[royal sail]

i actually hate it

every company is terrible at naming models except kimi and minimax

and deepseek ngl

DeepSeek's naming feels the most straightforward

[distant idol]

never seen an ai model with a good name

[proud bobcat]

lmao gpt 5.4

gpt 5.3 hadnt even fully released yet

😭

[frosty lava]

they won't release a normal 5.3 it wouldn't make sense

only the instant version

[ocean vortex]

svg is measuring spatial awareness and design/visual skills directly. That's as good of a test as website design IMO. It forces the model to draw things manually

[weak dagger]

Qwens video generator is just so ahh 😭💔 its like veo 3

[proud bobcat]

IM CRYING HOW DID IT FALL BEHIND 5.2 CHAT

IM CRINE

[frosty lava]

preliminary test + its only one point behind

[distant idol]

it js came out

[inner relic]

dude

just test

[proud bobcat]

imagine its like gonna be a point ahead

😭

[royal sail]

I can agree that it's good for measuring spatial awareness, but drawing out SVGs is a completely different skill in comparison to interface design

[proud bobcat]

lets see

[hazy marlin]

honestly claude is OP when it comes to coding

[frosty lava]

the writing is better if you care

[inner relic]

yep

[distant idol]

agree

[hazy marlin]

the only reason I use it and not other ones because it actually edits the code not rewrite it entirely everytime

[ocean vortex]

arena score is not indicative of model capability #general message

[frosty lava]

i'd say claude is the best for front-end

[proud bobcat]

nevermind

solid

all gpt models get choked up on this

[royal sail]

This is a pretty bad way to test a model tbf

LLMs don't work in letters

[inner relic]

Alright test it, If it's good at roleplay

[proud bobcat]

lmao

[honest verge]

Gemini 3.1 0326 flash lite Omni pro preview high flash TTS image veo

[proud bobcat]

hold on i gotta see ts

[hazy marlin]

yeah, I will only be impressed if they were able to 1 shot an entire HTML game in 1 prompt

[proud bobcat]

i wonder about its coding

hold on

i wonder if they fixed the schizo code of gpt 5.2 and 5.3

[distant idol]

do i need to tell him to explain or thats awkward

[inner relic]

👍

[honest verge]

TECHNOLOGIA

TECHNOLOGIA

[frosty lava]

when do you think they'll release the codex version

5.4 codex

[distant idol]

yes techologia

[royal sail]

in 2 hours

source: idk

[distant idol]

source

[honest verge]

April 1

Source :Slam Altman

[royal sail]

do not slam the altman

[proud bobcat]

scam altman

[distant idol]

double kick him

[honest verge]

FINALLY

This is what I needed

MISTRAL VIBE CLI

FINALLY

I WAITED FOR THIS MY ENTIRE LIFE

[proud bobcat]

ah yes

mistral cli

i needed this

truly

(mistral sucks)

[distant idol]

he will murder u

[honest verge]

Wonder if this is even better than Gemini 2.5

I'll test it

[proud bobcat]

probably

[frosty lava]

why only gpt 5.4 medium on coding arena

i want the high version atleast

[signal pelican]

lol.... 5.4 is behind 5.2?

[proud bobcat]

yeah im confused too

[proud bobcat]

a bit surprising, even if its still early testing

[frosty lava]

preliminary test and its only 1 point behind

[distant idol]

give it a day or more

[frosty lava]

no honestly everyone saying its worse than gpt 5.2 is lying

that's just not true at all

[signal pelican]

yeah... I've been watching the leaderboard for a while... I don't think 5.4 will catch up that much.

[frosty lava]

the leaderboard doesn't refresh every second

[signal pelican]

oh I know

[distant idol]

its not a big leap tho

[fierce kelp]

You can click to see what the router is using

[burnt sinew]

Where

[frosty lava]

yes but its not a coding version and it's not worse like people are saying

but your right definitly not a huge leap in capabilities

[signal pelican]

I'm sure it will beat 5.2 someday... but I'm talking about others... lol

I think they are actually falling behind Gemini and Claude

[frosty lava]

let's see

[wraith wren]

Quick question. Even though gpt 5.4 was out for maybe an hour? Do you think gpt 5.4 is better than opus 4.6 at coding?

[fierce kelp]

Click Max above the response and it will say "Response provided by ..."

[cinder nexus]

gpt 5.2 still stands strong

[frosty lava]

they clearly have focused on creative writing too this time

[frosty lava]

not front end, but for example gpt 5.3 codex was better for everything that does not involve good looking

so 5.4 should do same as 5.3 codex

[wraith wren]

Interesting. So for back end gpt 5.4/5.3 codex is better?

[distant idol]

le gpt

[frosty lava]

Yes definitly imo

for good looking go to opus

gpt is bad at making thing look good

[honest verge]

LOL THIS IS GPT 5.3 INSTANT

LOL

Vs mistral

[wraith wren]

Oki. Ty

[honest verge]

MISTRAL is better

[royal sail]

This looks horrendous

[burnt sinew]

Terrible

[honest verge]

But gpt is too simple

[frosty lava]

yes i was never able to make a gpt model to do something good looking

[cinder nexus]

the font wants me to kms

[royal sail]

for a website, i would rather have GPT's lol

[distant idol]

💔

[royal sail]

The font for mistral is terrible

[honest verge]

Still good for the worst model ever

[distant idol]

the font of doom

[cinder nexus]

fr

[distant idol]

guys stop bullying him

[slim gorge]

why is gpt-5.4 already out??? gpt-5.3 just came out like wtf 😭

[distant idol]

oh

👀

[honest verge]

Left is gpt and the right is mistral I can't decide which design for repair station is better

(idk what repair station even means)

[royal sail]

Well these two designs are completely different lol

not really fair to compare

depends what you need

[proud bobcat]

this is awful

also gpt 5.4 is awful

like wow

its not

impressive at all

[honest verge]

Over hyped model

Everybody were like

[distant idol]

fr

[honest verge]

WOW IT'S GOING TO DESTROY EVERYTHING BEST MODEL EVER!

But we got nothing

It's not better at all

[royal sail]

overhyped by nobody except scam altman

[proud bobcat]

i think openai is just "RANDOM SHI GO"

[graceful vortex]

guys I have to ask one thing can anyone that have a knowledge with arena.ai answer me?

[proud bobcat]

what is it brinks truck sniper

[thorny schooner]

Is it just me where thos battle mode in direct mode increase the frequency of mistakes even if using skip like I keep I getting stuck my getting into a freeze area where it just tells me a error was made despite no matter what I do

[distant idol]

nice pfp

[honest verge]

Openai rushed 5.4

I think it's clear

[vital lake]

Bru how people only care about coding 😭

[graceful vortex]

ty yours too

[royal sail]

5.4 does feel great for writing unironically

[vital lake]

Wait for 5.4 codex before judging on coding

[vital lake]

Mhm

[royal sail]

not sure why they didn't talk about how good it is at writing

feels a lot better

[graceful vortex]

why is arena.ai free. I mean how isnt the API's are so expensive?

[honest verge]

But it was supposed for complex tasks...

[vital lake]

Its creepy how good it is

[inner relic]

Price is questionable

[honest verge]

While 5.3 was supposed for writing

[inner relic]

Roleplayers doesnt want a expensive model

[proud bobcat]

backed by the ai companies to get free testing on their models

[thorny schooner]

😭

[royal sail]

well, most models that are great at writing are expensive

[proud bobcat]

its a worthwhile investment

[distant idol]

yes

[royal sail]

Top writing model is Opus 4.6 lol

[graceful vortex]

so the companies giving it for free trial ?

[proud bobcat]

basically

[graceful vortex]

to arena.ai test it

[thorny schooner]

Basically money is replaced with data as the cost

[proud bobcat]

arena gives the rankings and companies get the data

[graceful vortex]

and people see the models : "oh wow its a good model"

[graceful vortex]

and company won

[proud bobcat]

basically yeah

[proud bobcat]

wheres the meh option

its defo better than gpt 5.2

[honest verge]

Yes

[thorny schooner]

Basically all three of that since we all kind of answer at the same time lmoa

[proud bobcat]

but its not

[vital lake]

Oops

[royal sail]

This question doesn't really make sense

[proud bobcat]

really great

[royal sail]

like

in comparison to what?

[honest verge]

But not for what it was made

[vital lake]

Overall bruh

Like its pretty clear

[honest verge]

Gpt was supposed for coding and agentic tasks