general

[left lodge]

Never use non-thinking models, enable search and thinking mode that will provide you most reliable and safe experience.
And always provide as much as context possible like your Linux version, specific Linux distribution , etc
If you have link to the specific app you want to download, provide its website link so you can make sure you are downloading exactly what you asked for.

[wicked talon]

Kk

[signal apex]

is there sth wrong with claude 4.6? everytime i use it i receive “something went wrong with this respond”

other models are good except for claude, i can only continue chatting if i switch the model

[left lodge]

Maybe cause of high demand

There's a limited api requests available for the whole website. And also limited requests for each user too.

[hushed gyro]

also gpt 1.5 has the same issue i think

[bleak lake]

Sadly ai studio died

[left lodge]

Which AI studio?

[bleak lake]

Google ai studio, as it got limited to 10 prompts only

[wicked talon]

NAHH KIMI IS MY BBY IT DID IT IN 1 PROMPT

Gemini couldn't do it in 5

[hushed gyro]

you can thank me later 🤭

[signal apex]

ohhh so each chat session can only make a particular number of requests?

[wicked talon]

🙂

[bleak lake]

Really liked how consistent gemini was in ai studio even after ton of usage

[left lodge]

Not like that but a user can only make limited requests in a time frame so people don't abuse

[signal apex]

will it reset?

[left lodge]

And the whole server from all the people can also only accept a limit amount of request in each second and each minute.

[left lodge]

Yeah

You will see rate limit error rate and how much time.

Sometimes not , it depends on what caused the error

There is a lot going on this website

[signal apex]

i dont see it i only receive this

[left lodge]

There can be multiple reasons behind that

Cause it a generic error

[bleak lake]

usually the traditional fix is to clear your browsers cache and signing out and signing in

[signal apex]

it only happens with claude model, others work fine

[left lodge]

Like server rate limit, if a generation goes over or near 6 mins while generating, network issues, browser issues, etc

Sometimes it's just the server having issues.

[bleak lake]

man I miss the old ai studio, even Google ai premium is trash

[left lodge]

You can retry once or twice quickly when a new minute starts to see if it was a server rate limit or the model is actually out of service or something else is causing that because sometimes after trying it just works.

[signal apex]

ive been doing it but it just doesn’t work

[hushed gyro]

See this is why they need to fix the errors - other models just simply aren't good enough!

[left lodge]

If using a vpn , or even when not sometimes connections silently die. So like disconnecting from network , opening a new tab (arena.ai) and reconnecting to it have also fixed the issue for me and some other people

[wicked talon]

Claude

[hushed gyro]

For Nano Banana Pro?

[left lodge]

Yeah

[hushed gyro]

Use a VPN? I'll try...

[left lodge]

But it depends if it is causing the issue, mostly it is whole server rate limit. Trying when a new minute starts helps mostly with nano banana

[bleak lake]

What model is this?

[left lodge]

Wait for Qwen-Image-2.0, it seems quite promising and is trying to look like it can beat or is nano banana pro level

[left lodge]

I use a vpn and the website works fine.

[north obsidian]

Yup

[bleak lake]

Genie killed it

[north obsidian]

It isn't mistral, mistral is the best

[north obsidian]

Ye sad

[bleak lake]

Man I need an alternative to ai studio

You have any?

[north obsidian]

No, I use just Gemini weak and arena don't working Claude

[bleak lake]

Sad

[north obsidian]

Everyone is so sick

Try gpt

[bleak lake]

Gpt is not at that level

[north obsidian]

At least mistral works in another level

[wicked talon]

Nah Kimi is actually the best ai wdym it installed llama 1b in 1 PROMPT with no errors

[north obsidian]

Yeah I agree

[bleak lake]

Grok 😹

[left lodge]

Their k2.5 one specifically is really good

[surreal zephyr]

Bad model getting stuck in loops, common with opus

[surreal zephyr]

kimi is actually good

worse than gpt but still very good

[wicked talon]

Nuh uh

Kimi seems amazing

[hollow imp]

Has he even tried gpt 5.2 xhigh

[wicked talon]

I ain't got za dollar to

[surreal zephyr]

imo high is better overall

xhigh gives more opus like experience

and eats tokens

[hollow imp]

Did you try it

Where

Api?

[surreal zephyr]

im using mainly 5.3c and 5.2h in cli rn

[remote vapor]

heyhi ~ ~ im supposed ti introduce myself, sooooo hi!
im maria, I'm.... mostly here to discuss where karp-001 and 002 went..... cuz I want my qwen3.5 back!!! N

[bright shard]

@echo aurora Why did Arean.Ai remove the Nano Banana Pro Normal model?

[remote vapor]

probably because it became too expensive for Google to give it away for free....

[woven harness]

hiii

[remote vapor]

heyhi ~

[woven harness]

well thats weird

its the 1k btw

[bright shard]

And is the Nano Banana Pro 2K model more expensive?

[bright shard]

That's not the PRO

[remote vapor]

(I actually don't kno, I just joined >~<)

[last oxide]

maria can u talk normally

holy corny

[remote vapor]

sigh okay fine.... i was about to comment on lucii 2s status thing... but fiiiine I won't 😭

[steep bear]

bro got banned

LOL

[remote vapor]

Who?

[steep bear]

the fjord guy

he left

[echo aurora]

Models can be removed from the site for various reasons, I won't be able to go into the specific for why that's the case sorry to say.

[plucky sparrow]

does arena.ai still announce when new models come out?

[echo aurora]

Yes

Also want to note our #ask-here channel

It's the best place for quick questions

[remote vapor]

ooh! I will ask stuff there now! ~

[plucky sparrow]

ty

[remote vapor]

aaa my message has been hidden..... hmm lemme test for the restrictions.

kids

aaaaaaa interesting.

[hushed gyro]

Gpt 1.5

[remote vapor]

sooooo R08L0X is a banned word....

so iguess there really was something about that there...

[gleaming roost]

[echo aurora]

Yeah this may change in the near future. We were seeing some spams particular to roblox in the past.

[remote vapor]

ooh interesting!... why? Are you allowed to share?

[proud bobcat]

life

[echo aurora]

We may change this in the future as we've been seeing the related scams die down, so it's not as much of a problem currently.

[proud bobcat]

wow its actually a banned word

rob lox

lmao

[gleaming roost]

[remote vapor]

sooooo theres also a deepmolt model....

obviously relating to moltbot..... openclaw..... yea!

[velvet forge]

yooo for students

or

oh dang

i missread

for academic

[vital kelp]

Is there a way to do research on copyright content on AI?

[weak dagger]

How do i make this? @cloud zinc

[remote vapor]

you mean finding out about what data they trained on?

[surreal zephyr]

hello another suspicious bot account/ban evader 👋

[vital kelp]

I’m doing some research on some content that is copyrighted, so basically I can’t access such content with AI and its regulations, I want to know if I can bypass this

[celest mirage]

Guys, why doesn't the Gemini Nano Banana 3 pro work?

[vital kelp]

😂😂😂 chill bro

[remote vapor]

hmmmm....... what content do u mean?

and how would you want to access it "with AI"?

why would "AIs regulation" affect you being able to access content?

[swift ermine]

Make ur own jailbreak prompt

[vital kelp]

Pocket toons, comics I want to generate the dialogue in some of the comics

[remote vapor]

eh u can just prompt the model to, no?

LM labs don't really care about copyright.

[plush kettle]

Hello

why does every model says Something went wrong while generating the response. Please try again.

On lmarenaai

[half mist]

Where the frick is GPT 5.3 Codex?

[remote vapor]

its not on the api yet.

also - woah calm down ~ ~ ~

[plush kettle]

can you please help me

[remote vapor]

what is it?

[hushed gyro]

I am very happy to say that NB Pro 2K should be working about 90% of the time rn, try using VPN

[plush kettle]

why does every model says Something went wrong while generating the response. Please try again.

[remote vapor]

what is it saying? u gotta be a bit more specific.

[plush kettle]

Its litterly saying what i wrote

[remote vapor]

oooh that's what you mean! I get it, sorry, my bad

[remote vapor]

yeaaaa, try pressing Ctrl + Shift + P and opening lmarena.

[plush kettle]

and on every prompt or messages

[remote vapor]

it usually happens when u use the models too much.

[plush kettle]

print?

[remote vapor]

Nono, private tab

[bright junco]

Why is it like this? Can anyone fix it? I’ve pressed retry many times already, but it still won’t generate.

[plush kettle]

same

[remote vapor]

u probably used the model too many times.

[plush kettle]

So i try on Incognito?

[remote vapor]

ye

[plush kettle]

ok

[remote vapor]

if it doesn't work - oh well.... then that's it for today.

[plush kettle]

same problem

[remote vapor]

ye- then u used up today's limit-

[plush kettle]

i have never used the web do you know?

[remote vapor]

u never used the internet?...

[plush kettle]

Like its my first time messaging it

No lmarenaai

sorry english is not my main language.

[remote vapor]

hmmm oki then iguess they got some other thing..... cuz everyone's experiencing it, somthing is up-

soooooo, dunno, am not an employee >v<

[edgy iron]

Can you confirm that it still works for you? Since this error is present for many people today, unlike yesterday.

[plush kettle]

So the issue i am getting is happening to everyone?

[remote vapor]

yea, the model is down.

[cinder nexus]

can't be it.

[remote vapor]

yea, then the model is down- or whatever, anthropic does what they want-

[cinder nexus]

or perhaps it is, since mine is still thinking whatsoever, but it does the job done

[remote vapor]

its just claude apparently, so its their thing.

[cinder nexus]

try telling the AI to keep it's thinking short

[remote vapor]

hm? no..... why?

[cinder nexus]

it is thinking too long probably makes it break

[remote vapor]

is it printing reasoning tokens?

if not - then its not working.

[cinder nexus]

try clicking it

[remote vapor]

sooooo no reasoning tokens, hm?

then its just as much *working" as the non-reasoning version.

[echo aurora]

Hello for those havign the Something went wrong error message I'd encourage you to check out this message: #1417174113092374689 message cc @bright junco @plush kettle

[remote vapor]

in other words: let's try some other models! ❤️

[mystic frigate]

very quick mod lmao

[surreal zephyr]

its not for free in llma lmao u gotta pay

🥀

[woven harness]

its an album reference, smoochies by ashnikko 😭

[echo dome]

some of the models are down

it's not us

[shrewd citrus]

aw it’s only for Americans 😖

[echo dome]

damn bro I'm from russia

[toxic verge]

This is really cool

Awesome 👏 good for lmarena I definitely salute

[celest orchid]

txt upload testing pls

[echo aurora]

Be sure to use #1372230675914031105 as that's the best place for these kinds of requests!

[grand cliff]

Hell yeah

PDF

And I got too excited

Still none

It's not on Direct chat

[remote vapor]

pfffff this graph is making it look waaaay more impressive than it is!

[surreal zephyr]

yah thats the joke 🤣

[north obsidian]

Yup lol

[north obsidian]

8 diff

[surreal zephyr]

wym the 2nd bar is clearly higher

[stray aspen]

what are these goofy benchmarks lmao

they are overdoing it

[proud bobcat]

arena is winning so hard

[proud bobcat]

this is false

gpt is known to have the most schizo models known to man

[plucky sparrow]

are th ese real graphs or are you generating them? 😄

[proud bobcat]

definitely generated

this reeks codex generated

real benchmark

higher is better

its not for 5.3 codex yet but i imagine it'll be marginally better than 5.2

[proud bobcat]

gpt 5.2 has less than 5.1

crine

[surreal zephyr]

benchmarks are bs when its literally opposite in practice lol

try both

same prompts

oh thats xhigh

[proud bobcat]

this is not what a hallucination is

[surreal zephyr]

try high not xhigh

xhigh is bad

xhigh has same issue as opus

[proud bobcat]

theres no high version of 5.2

only xhigh

[proud bobcat]

hallucinations are when a model is confidently incorrect and you need a huge benchmark to discern that

gpt 5.2 in my experience has spit out nothing but garbage

i know technically this is 5.3 but openai is known to have hit or misses everywhere

[surreal zephyr]

chat is this ragebait

theres low, medium, high, xhigh

same for codex

xhigh isnt even on llmarena cuz it sucks

[proud bobcat]

again

this is on artificial analysis

so i dont have it there

[queen veldt]

Use opencode bro

[surreal zephyr]

i have paid codex sub lol why would i

[queen veldt]

There's Codex 5.3

And + you can use your gpt subscription using oauth

[knotty fable]

Last weeks of Discord, thank you to the nice persons I met here - you enjoy as long this last.

👋

[cinder nexus]

where do u even find the xhigh

[surreal zephyr]

cli

and api

[hushed gyro]

Chat are the usual generation errors still happening?

[surreal zephyr]

no thats purely opus issue

[hushed gyro]

Is it because of the global ID thingy?

[hushed gyro]

Cuz now gpt 1.5 has some serious errors and NB Pro for like 20% of the time

[surreal zephyr]

crazy

[hushed gyro]

Why is the bar height difference so much for a 3 point increase? Is a single point worth a big difference?

[woven peak]

am i the only one if the ai's response or thinking is long enough then i get "something went wrong with your response" its not rate limit happens on new account

[hushed gyro]

So basically the models very good but the platform is unstable or is both poo poo quality 🤮

[hushed gyro]

It's probably either high demand or faulty servers

NB Pro magically starts working again once I started using a VPN connection 👀

[woven peak]

i think main reason is faulty servers because i remember using lmarena before everyone found out about it (it was still popular but less demand + it was before they added the video arena so even less) and i still had this issue it consistently happens if the ai's output is long enough or has taken enough time (i actually don't know which one it is, it's consistent though).

[surreal zephyr]

its a joke about people glazing opus so much

[surreal zephyr]

opus issue

it gets stuck like older models

[woven peak]

any way to fix it or is it just the model?

cuz currently i am using opus

only for thoughts or output in general? thought was 5 minutes and 11 seconds long and final output was idk how much but it didnt finish

looked at bug forum everyone says theres timeout

[placid verge]

is the servers down or something any time im trying to use nano banana pro it keeps giving error messages even though its never done that before

[echo aurora]

Hey @placid verge I'd encourage you to use our bugs channel to review bugs flagged from the community as often the issue you're having is going to be discussed there. For this problem I'd check out #1417174113092374689 , but more importantly I'd encourage you to check out this message: #1417174113092374689 message

I'd also ask to avoid pinging the moderators for questions/bugs/feedback/etc. The mod ping should be used for reporting others users breaking our server rules.

[remote vapor]

I want my qwen 3.5 baaack >o<

[toxic verge]

I think there is something missing in Llm evaluations. I think the current system we have now creates a lot of more confusion than it does resolving or proving anything other than a popularity contest.

[plush drum]

Hello guys, does gpt-5. 2 serach work in you?

[toxic verge]

It seems that regardless of the benchmark results people are still debating based on opinion and preference versus any tangible rigor

I think it only adds to the confusion in the long run

[shrewd citrus]

yes works fine it just takes a while

[shrewd citrus]

do you know if the benchmarks results are only from battle mode?

[edgy iron]

sure you won't notice difference coding fizzbuzz, it's quite stark when doing computational biology

[plush drum]

5min? More or less?

[shrewd citrus]

or does it also include side by side chat and even direct

[shrewd citrus]

depends like if it’s a hard question like let’s say Gemini 3 pro grounding had to answer it would take 2 minutes but gpt will take 6

so it always takes 3x longer than 3 pro which is what I noticed

[plush drum]

Get it, thanks

[toxic verge]

There needs to be a way to measure not only with the model is capable of, but also there’s gotta be a way to track when models get nerfed

And at simple things specifically where they fail in real life cases

There’s gotta be a way where you could figure out the middle ground to get a more accurate picture

1st thing that needs to change is either A. Be able to reproduce these the claimed performance and abilities that these AI companies claim when they release these models, or B. Not believe a word until proven otherwise

Cuz this is ridiculous These debates go on day and night with no concrete anything

Seed 2.0

For example, if you were to take away nano banana 2 from Gemini 3 I don’t think it’ll be so high up in the rankings personally. Or if you didn’t include all of the features that come with Google plus making it more lucrative

[north obsidian]

Name?

I just know dreamina

[surreal zephyr]

ok 5.3 xhigh makes opus a joke

[north obsidian]

Insane

Ok, that I think opus can do more than this, but the diff is big yet

[north obsidian]

Yup I saw that

[safe sleet]

Are you sure this isn't just a case of GPT 5.3 using a fluid simulation librariy while the others aren't?

[grand cloud]

guys where is nano banana pro normal version (not 2k) ????

[echo aurora]

Going to send this to #ai-creations

[limber panther]

its seedance 2.0 btw

[echo aurora]

Looks great, I've been hearing amazing things about it.

[grand cloud]

does anyone know if it will be back ?

[echo aurora]

This model was removed. Models on Direct/Side by Side can be removed for various reasons, we won't be able to always go into details about why.

[north obsidian]

Where did u get access?

[limber panther]

in chinese jimeng AI

[north obsidian]

Ok thnx

[grand cloud]

alright, but can you tell us if it will be back ? and also will the 2k version get removed as well ? if not, no worries

[surreal zephyr]

Can you atleast hint was that issue with the model, or googles decision, and is it likely to come back?

or law issue?

or did opus remove the api by accident when adding itself to the website?

😭

[shrewd citrus]

nah this is crazy impressive

[surreal zephyr]

@safe sleet

[surreal zephyr]

yeah nah 5.3c makes opus a joke

[echo aurora]

No sorry to say I can't provide this information. Mainly because we don't know at this time.

[echo aurora]

Lol no sorry I can't hint at it either.

[surreal zephyr]

can other models be removed in future too?

[echo aurora]

Yes

[surreal zephyr]

and its likely to happen?

(should we worry)

[shrewd citrus]

do you have a link

[wicked talon]

Gng is Kimi worth paying for

[limber panther]

type in google "jimeng AI"

[echo aurora]

No, I wouldn't worry.

[shrewd citrus]

the 1st result by bytedance gives me this lol

[limber panther]

if ur scared then forget about it

[wicked talon]

Lol

[limber panther]

scary chinese site

[shrewd citrus]

nah it’s a broken link

[wicked talon]

Btw tip when browsing Google do not use advertised links at all

That's how you get scammed

Am I allowed to post the URL or is it bannable?

@echo aurora

[shrewd citrus]

thanks but I don’t think sponsored links show up first on safari

[wicked talon]

It would as it's Google search engine

I found the URL I'm just waiting on confirmation

[shrewd citrus]

you can send it in dms

[wicked talon]

Is this tryna make me install malware

Tf is this ai

Lemme run it through virustotal

[north obsidian]

My is in Chinese 🙁

[wicked talon]

Mine is too

The apk file is English

[echo aurora]

Sorry what is this URL to/for?

[wicked talon]

An ai

I'm just testing for malware from it though just in case

[echo aurora]

[north obsidian]

Send in Private

[wicked talon]

I'm testing for malware first bro

[echo aurora]

If it's that sketchy I'd prefer not.

[wicked talon]

I'm not distributing malware

[wicked talon]

Alright

[north obsidian]

Ok

[echo aurora]

Thanks for asking though, I really appreciate it.

[wicked talon]

Yw gotta keep stuff within the rules

@north obsidian

Glad I didn't send it

[north obsidian]

Ow

I had installed I think

[wicked talon]

..

[north obsidian]

It's a dreamina Chinese version

[wicked talon]

Play protect should of flagged it

In my opinion

[north obsidian]

Yeah

[limber panther]

try global Dreamina website, maybe its released there

[surreal zephyr]

Should've* 🤬

[north obsidian]

I know the cause my didn't get it
I didn't installed on this web

[safe sleet]

That's interesting

[wicked talon]

Should I buy Kimi

[shrewd citrus]

isn’t it free tho

[wicked talon]

Ai swarm isn't

[shrewd citrus]

well why do you want to buy kimi

[wicked talon]

Owch that hurts the pocket

[wicked talon]

It seems good

[shrewd citrus]

it’s good but I don’t think it’s £35 good

[limber panther]

whats ur project tho

[wicked talon]

Try out

Literally it

[wicked talon]

Vivace is £179 lol

[shrewd citrus]

check their refund policy

[hollow ivy]

By the way, how good is Opus-4.6?
Could it code a decent civ-AI/engine?

[wicked talon]

Hell no

Actually maybe

Try Kimi 🙂

[hollow ivy]

Kimi 2.5 > Opus 4.6 ?

i cant believe that..

could Opus 4.6 code a decent AI for the ancient Empire wargame?

[shrewd citrus]

probably after multiple refinements

[shrewd citrus]

aw that sucks

[wicked talon]

Maybe

Try Kimi

🙂

[hollow ivy]

Which game sits half-way in complexity between ancient Empire game and Civilization I?

[sage igloo]

Animated, slow camera push into the brilliant underwater city of Atlantis, subtle movement of water and light

[hollow ivy]

(..this could be interesting for Opus-4.6)

[shrewd citrus]

what about on the App Store

[wicked talon]

Strategic conquest

[wicked talon]

They don't accept playstore payments

[shrewd citrus]

double check

[wicked talon]

WAIT WAIT

Shii

1 week trial

W

[wicked talon]

This one doesn't come with agent swarm thoughh

But yeah they will refund it but your gonna get your account banned from Kimi

[hollow ivy]

Could Opus-4.6 create a superior Tron-bot?

[wicked talon]

@shrewd citrus

[wicked talon]

Its possible

[shrewd citrus]

haha nice

[wicked talon]

I was gonna create a chrome extension with Kimi till it said £5 fee

Hell nah

Who would y'all pick as your daily assistant, Kimi or Gemini (Gemini 3 pro preview, Kimi 2.5 thinking)

What in the misinformation

1. Kimi K2 will definitely not run on an iPhone
2. Kimi Self hosted will not beat gpt

Chat dead

[surreal zephyr]

https://cdn.discordapp.com/attachments/1018470269326729286/1470906684347253038/togif.gif

[surreal zephyr]

Kimi is actually kinda good and sometimes can compete with opus, but its still way worse than gpt

[wicked talon]

Help me translate I'm too English

[surreal zephyr]

Put ss into gpt

[thorny drum]

Do you guys know if opus 4.6 was tested as an anonymous model?

[surreal zephyr]

No it sucks

[thorny drum]

How did it get so many votes so fast

[surreal zephyr]

The leaderboard score is sponsored lol

[thorny drum]

3k votes in one day seems unprecedented

[wicked talon]

Ugh

Gng it thinks I'm Chinese

Time to test chatgpt

[wicked talon]

This Kimi extension is defo spyware

Whoops didn't mean to reply

[honest verge]

Xhigh looks so beautiful tho

But it's long

[wicked talon]

Ewww

(jk)

[honest verge]

Honestly surprised

It looks very good

[surreal zephyr]

It wrote custom shaders

And actuallt succeded

[honest verge]

It's good but still waiting for API

[surreal zephyr]

(All are threejs btw no premade libs)

[surreal zephyr]

They made it require id verification to access

Because it was so good at hacking

💀

[honest verge]

LOL

What's 5.4 will be

Like password required

Actually no

Probably they won't even release it

Because it's not safe already

[surreal zephyr]

https://openai.com/pl-PL/index/trusted-access-for-cyber/

🥀

[honest verge]

WHAT THE

FR?

LOL

[surreal zephyr]

Yes fr

Its like another level lol

Compared to opus

You literally need to verify id to use it 🤣

[honest verge]

Opus is slow and expensive

[surreal zephyr]

Gpt is fast, token efficient, smarter, and reliable

Opus just thinks forever until it gets an idea

[honest verge]

Btw do you think 5.3 codex xhigh is actually good?

Or it's the same as 5.2 xhigh

[surreal zephyr]

I prefer high but both are insane

Xhigh is more creative

But high better overall

Xhigh is more opus-like

[honest verge]

If 5.3 is so strong what 5.4 will be?

Creating GTA 6 from scratch

[surreal zephyr]

I spent 40% of my weekly quota today. All codex 5.3 c high.
A lot of usage.
Not a single syntax error. Not a single "bug" / "hallucination". Most features one-shotted, some with minor tweaks

(The quota is very high, like very very)

[honest verge]

We will get GTA 6 before GTA 6

[honest verge]

Also it's 2x In app

[surreal zephyr]

With 1 prompt

💀

[surreal zephyr]

Im using the cli

Prolly 2x too there, no?

[honest verge]

App is for mac os now only

[surreal zephyr]

Well the 5h limits are virtually infinite

[honest verge]

I think no

[surreal zephyr]

The weekly limits are crazy high but drainable

[honest verge]

They said it's only in app

[honest verge]

I love openai for high limits

[surreal zephyr]

Theres no way the limit would be this good lol

[surreal zephyr]

The token count limit is similiar to claude

Its just that opus is idiotically inefficient

[honest verge]

But Claude is just too expensive

Like I remember opus 4.1 was so expensive

[surreal zephyr]

Codex thinks for 5 mins

And one shots stuff

[honest verge]

No one could use it

4.6 is slightly better

But still expensive

Compared to gpt

[surreal zephyr]

Id rather infinite gpt 5.3 than inf opus 4.6

Thats how good it is

[wicked talon]

I would rather have inf Kimi 🙂

[surreal zephyr]

Kimi is nice small model

5.3c is js another league

❤️‍🩹

[wicked talon]

Small lmao

[honest verge]

Ye it's for cli too

[wicked talon]

It's the biggest opensource model

[surreal zephyr]

Smaller than the closed ones

[surreal zephyr]

And the best one too

[surreal zephyr]

I wish they keep em at 2x

🙏

Rn 20$ codex has more quota than 200$ claude

[honest verge]

They just destroyed anthropic

I don't think even sonnet 5 can do anything

[hollow flicker]

i kinda just made a PDF of a 4k line code and made it into a pdf is that cheating

[surreal zephyr]

Sonnet 5 gona be temu gpt 5.3c

[surreal zephyr]

Depends on model, some wont eat it

[hollow flicker]

claude opus 4.6

[echo aurora]

Did I miss a ping

[honest verge]

Rumors say Gemini GA tomorrow but I don't think this is true

Google releases their models once in a year

It's too early for now

[toxic verge]

They have check point models

That they always test out and are always circulating

[hollow flicker]

it was me i deleted it because i thought that you sleeping

[honest verge]

But not tomorrow

[toxic verge]

Yeah, they’re different versions of the models

[hollow flicker]

cause it said you offline

[toxic verge]

It’s possible, but once again it’s rumors and when it comes to rumors, it’s hard to really pinpoint anything concrete to say for sure

I’ve been hearing similar things so highly likely that something is coming soon if the rumors are going around

[honest verge]

It was the same with sonnet 5 for 3 days but we got opus 4.6

So maybe something is really coming

[echo aurora]

odd, not sure why it says that

[toxic verge]

https://youtu.be/o4zv8SxAWNM?si=Tsf1KXAI9NUx85Ir

[honest verge]

That's what I'm talking about

But it doesn't guarantee anything

[toxic verge]

It’s not a new model though

[honest verge]

4 different models

[inner relic]

this is chinese slop new ai video

[honest verge]

Chinese?

[inner relic]

seed 2.0 viode

video

[inner relic]

sorry uh, It's seedance

or seedream

They releasing 2.0 video generator seed soon.

[toxic verge]

https://www.testingcatalog.com/exclusive-a-new-gemini-3-pro-checkpoint-spotted-in-a-b-testing/

[toxic verge]

It’s already out

[inner relic]

huh

already out?

I dont see it

[toxic verge]

Not for everybody n not in arena

[inner relic]

I am testing it in byteplus

[stray aspen]

when will we get 5.3 api

[toxic verge]

Yeah, I’m sure a lot of places have access to it right now

[hollow flicker]

do you like being a pineapple or do you ever think you want to be a different fruit

[limber panther]

whats ur prompt

[echo aurora]

I'm satisfied with the pineapple

[hollow flicker]

i like being a gunie pig

tho im thinking of turning into a ugunda knuckles

[proud bobcat]

Huh

Didn’t they say this about Gemini 3 pro last time

Also no new Qwen image 2 on arena yet

Sniffle

[toxic verge]

https://arxiv.org/pdf/2512.24511

You can take a pre-trained "base" model checkpoint and start new training to teach it a specific skill fine tune it whatever or If you notice the model is becoming dumber and start hallucinating more. You can go back to earlier checkpoints.

So they’re not releasing a new version of Gemini it’s just a better fine tuned refined polished of the same base model which is Gemini 3

[obtuse smelt]

i look just random generate where is like gemini 3 or grok, is full flux model

[rugged robin]

Ion know

[obsidian cargo]

Okay does anyone know where you go to try seedream 2.0?

[north obsidian]

I just got it on dreamina or something like this

But I am searching yet

[rugged robin]

@obsidian cargo I'm new to discord how do I get the fuvk outta this page. I want explore new things

[toxic verge]

Yeah byt kinda sketchy

[north obsidian]

Thanks

[toxic verge]

Np

[north obsidian]

Are the credits refreshing dail?

[proud bobcat]

Seedance 2.0 is region locked I think

And API only for now

[obsidian cargo]

damn that's mondayfriday bullshrimp

I neeeed it

[brittle hollow]

When using a model to create a new image like a character posed or dressed like another character, how do you guys prevent mix ups with the subject and the reference? I've been having issues where the ai mixes the two images together to create a whole new character or uses the wrong one as reference.

[outer spear]

Use vpn n go to bytedance

[honest verge]

That's why mistral is better

[honest verge]

Lol what is this

[cloud zinc]

seedance 2.0

[honest verge]

His eyes look like he played for one week straight

But still looks good

And cinematic

[cloud zinc]

yeah that was the intent

[north obsidian]

Mistral in first place yet

[proud bobcat]

Top 10 ragebait

It will never not astonish me how mistral fumbled so hard

@surreal zephyr ts your man?

[north obsidian]

Mistral would say every p u want to be

[north obsidian]

Wrong mistral, the real mistral are the friends we make along the way

[viscid cloak]

most recommended: ask model to extract the referenced character's raw gesture sketch or clothing design(with multiple angles if possible). Then apply the sketch as a reference to prevent your target image from mixing up with the referenced image.

[wicked talon]

https://tenor.com/view/sigma-gif-15449497793648793961

[remote vapor]

yassss that's how it should do it ❤️

[toxic verge]

https://youtu.be/JebUU7FnWBI?si=gAHcCV73sPKWb7RU

[honest verge]

But it used python calculation for it

This is bad

Imagine counting with python

[woven harness]

opus 4.5t is better than opus 4.6 ❤️‍🩹

[toxic verge]

ChatGPT 4o mini better then ChatGPT 5

[woven harness]

which is the best LLM for writing stories/novels?

[wicked talon]

Kimi 🙂

[surreal zephyr]

You dont know how ai works

[tiny dove]

How in the world is this not fixed yet?!

gemini-3-pro-image-preview (2k)

[quartz coral]

opus with higher temperature settings give off good results

[left lodge]

Not opus 4.5 nor 4.6
Just sonnet 4.5 and its also non thinking
These test are nonsense but people love finding niche things which AI can't do

Even haiku non-thinking 😭

@echo aurora Haiku 3.5 not working?
Is it out of service?

Even sonnet 3.5 non-thinking is giving correct final answer, but it is showing 1,2 and 3 in the count? Old models.

[left lodge]

When it will comes to someones own preference they will do anything to defend it even if it's wrong than accept the reality.

[left lodge]

Its kind of nano banana moment for video models but there is still room for improvement because they are still not quite the quality or consistency where we can use them directly or with minimal editing in final production

[lean iris]

Hello, the website has a recaptcha error, Connecting to Arena has failed. Please try again later or on a different device, Failed to accept terms-of-use

[toxic verge]

Try clearing your browser

[toxic verge]

That’s why there needs to be a different kind of benchmark and evaluation for this

I was talking about this yesterday

[obtuse smelt]

is gemini 3 pro are work now

[undone saffron]

¯_(ツ)_/¯

[lean iris]

You should add the login option outside the login screen; that would be better.

[surreal zephyr]

I have just read some people actually use claude in work 💀

Like on actual important backend

😭

<@&1349916362595635286>

[plush sonnet]

✌️

[sturdy mica]

Claude models have some contamination with benchmarking

It's true. Sorry

[left lodge]

Which benchmark?
I mean if it is even true it doesn't change its current usability and reliability which is better than all other models and specifically its behaviour and its way of responding is what I like.

Benchmarks are good for first view and to see how much it got in the same scenario in comparison to other or older models.

[sturdy mica]

Error in your opinion, do you think Claude 4.6 Opus is better than or worse than Codex 5.3

[left lodge]

But that scenario might doesn't reflect what you do with your model and how it reflects your work. That's why people have so opposite reactions when a model is updated

[left lodge]

That specific comparison is entirely a little hot, codex 5.3 is kind of exception coming out if ChatGPT but the reason it is fine-tuned for specific tasks it's not a general model.

I am waiting for gpt-5.3 high

[surreal zephyr]

Codex 5.3 is way better and thats not even the same league lmao

[sturdy mica]

haha very funny

[surreal zephyr]

Opus is like 10% more creative while 50% worse memory, 50% more hallucinations, 50% worse prompt adherence, 50% more syntax errors

[sturdy mica]

Funny prank

[surreal zephyr]

No thats actual threejs result with same prompt

[sturdy mica]

ARE YOU SERIOUS

[left lodge]

There are many fators which matter not just raw intelligence from benchmarks like token efficiency, cost, speed , reliability, long conversation, etc

[surreal zephyr]

Yes

Thats how bad 4.6 is

5.3c wrote custom shaders that worked

[sturdy mica]

probably found it somewhere on the web

[surreal zephyr]

Bro 5.3c has like 200% token efficiency of 4.6 💀 💀

[sturdy mica]

i wish i could use 5.3 codex

but it doesnt have an api

😢

[surreal zephyr]

No internet access 💀
Sandboxed cli

All custom shaders, threejs. No libraries

[left lodge]

Idk what's happening but don't believe anyone, just try yourself.
One task should never be used to determine a models usefulness

[sturdy mica]

dude i would

[surreal zephyr]

Try then

[sturdy mica]

i would test codex 5.3

but i cant

i dont have it

[surreal zephyr]

You need to id verify

And paid sub

[sturdy mica]

i know

[surreal zephyr]

Its too good at hacking they gatekeeping it

😭

[sturdy mica]

so are they gonna make it stupid

before releasing the api

i pray they dont

[surreal zephyr]

Better not 🙏

5.2h was already better than opus by a bit.
5.3c makes opus feel like kindergardener

[left lodge]

I don't think so that's the case but it can be
The reason people feel the difference is cause of the system prompt and tools. They keep the models aligned.

[sturdy mica]

ampro share your openai account with me

🙏

[surreal zephyr]

5.3 is actually very token efficient

[sturdy mica]

yeah i saw

impressive

[surreal zephyr]

So it doesnt need nerfs

Like opus

Openai went into good direction

Claude went into brute forcing thinking loops

Opus is good for "vibe" coding where you have 0 idea what you are doing, you dont care about security, and want the website pretty

Codex is for actual robust code

[sturdy mica]

man

i really wish i could use it

[surreal zephyr]

Opus feels like gpt 4o xxxxxxhigh tbh

[sturdy mica]

when do you think the API is coming out

@surreal zephyr

[surreal zephyr]

Later likely

[sturdy mica]

yeah duh

[surreal zephyr]

I wonder how they will put id ver into api

[sturdy mica]

they wont

does it really need ID verification?

i don't think so

[grand cliff]

How so?

[surreal zephyr]

Its dumb token waster

Its literally quantized 4.5 with more thinking

Claude has no idea how to improve their models

Leaderboard maxxing 🤣

[sturdy mica]

give me the prompt

for the realsitci water

[grand cliff]

Hmm.

[sturdy mica]

i just got codex 5.3

Hurry, quick

time is ticking on it

[surreal zephyr]

Claude vs openai is like
Apple vs linux

[grand cliff]

I don't notice it.

[surreal zephyr]

💀

[sturdy mica]

@surreal zephyr Give me the prompt

[grand cliff]

But that is probably because I am biased towards 4o....because I used it so much back then

[surreal zephyr]

Just ask it for realistic water in threejs with custom shaders im on phone rn

[sturdy mica]

AHHH ITS WORKING

Hahahahahahahhaahhaa

[surreal zephyr]

The one on the pic was xhigh, but high and xhigh should both prolly do fine

Xhigh will make it prettier or maybe try too much and fail like opus

😔

[sturdy mica]

yeah i tried prompt "In HTML with Three.JS, make the most realistic water shader possible. with waves, a sun, volumetric clouds, and free camera." and it made a broken website that just says "Click to enter free camera"

it didnt add it to my github repo, and error: "Uncaught TypeError: Failed to resolve module specifier "three". Relative references must start with either "/", "./", or "../"."

[surreal zephyr]

It doesnt have internet access when sandboxed

🥶

[sturdy mica]

bro

but yours did it

when it was sandboxed

[surreal zephyr]

It cant load the library by itself

Ask it it will tell u

🥀

[sturdy mica]

fs

[surreal zephyr]

It cant access net because openai actually cares about safety

[sturdy mica]

not even close to decent

wouldnt even call it bad

its worse than bad

[surreal zephyr]

Failed imo

Ask it

[sturdy mica]

JUST GIVE ME THE PROMPT YOU USED

[surreal zephyr]

Bro its random making custom shaders is hard asf

Ask it to find issues and fix

Also graphic drivers matter

Tell it hardware, os, ect

[sturdy mica]

your special

your really special

[surreal zephyr]

You're*

The other one it wrote for windows and it didnt work on linux before it ported it

[sturdy mica]

told it i was on a raytracing capable computer and to go crazy with the graphics

yeah wow gpt 5.3 is awesome

[remote vapor]

naw, best way to do it as an LM

[sturdy mica]

wow. Gpt 5.3 codex

Smarest coding model

@surreal zephyr ??

maybe im using the wrong model

how do i check? im on the website

i think this might be using gpt 5.2 codex

yeah it is

thats why

[surreal zephyr]

Yeah looks more like 5.2c

Well idk what to tell you, 5.3c did better than that

[sturdy mica]

good

[distant spoke]

Yo guys, just found a prompt that completely fries AI logic.

​Prompt: "I need to wash my car. The car wash is only 50 meters from my house. Should I drive there or walk?"

​The Catch: Let’s see which models will seriously suggest "walking is more eco-friendly" while totally forgetting you’re there to wash the car. 💀

[golden ocean]

lmfao

[meager fulcrum]

Gemini 3 pro 2k image creation doesn't work again, can you at least bring 1k version back

[shrewd citrus]

lowkey i was about to say walk 😭

[left lodge]

glm is aware of your play lmao

[left lodge]

Don't let other people fool you with their slop.

Here's beauty

https://019c4be9-1e0d-71f3-aabd-d2f24690fb30.arena.site/

[distant spoke]

Deepseek

[left lodge]

So lightweight yet so amazing

[left lodge]

I think models assume that you can call out the car wash guys and they will wash your car in your home or will just come there to wash and you just need to contact them or something else because 50 meter is actually very less distance

[left lodge]

This is not ChatGPT nor Claude nor Gemini

That is not even a thinking model

The result is with a single prompt

And this one is from Claude Opus 4.6 Thinking with same exact prompt as that one.

https://019c4bdd-ca2f-74ce-a4dd-a2e6fadc9d68.arena.site/

[sturdy mica]

what model made this?

[left lodge]

Guess

[sturdy mica]

kimi k2.5 instant?

[left lodge]

Yep

[sturdy mica]

wow

[left lodge]

:]

[distant spoke]

Doubao

[left lodge]

What even is that lmao

So condensed ui

[shrewd citrus]

you have a telegram message

https://tenor.com/view/speed-side-eye-ishowspeed-gif-11081465071859519581

[left lodge]

A screenshot can reveal so much but people don't notice

[glad hawk]

this server is disgusting ewwwww

quoting IBM "a machine can't do creative work"

[left lodge]

💀 it redirect to https://openrouter.ai/openrouter/pony-alpha

[rough surge]

Ai video leaderboards are bit off for me , veo 3.1 sucks for physics and for anything else, how come sora 2 pro, kling 3.0/2.6, dont come even close to the leaderboards? i bet it's because of all the indians in here that only do product or weird talk-shows

[left lodge]

I don't think you should blame someone without any proof and there are many independent leaderboard you can check out if you don't like this one. This is one for preference, what humans prefer. Nothing else.

Preferences are weird they can change easily.

The votes are from across the world.

[rough surge]

I know, but it just bothers me, because in my testings, veo 3.1 it's not that good specially with img2vid

[left lodge]

Currently the video models overall have not reached their boom moment. They still need some time.

[rough surge]

I'm waiting for seedance 2, looks promising

[left lodge]

Seedance's new models is looking capable

But still it has the plastic look or I say we can easily identify with the weird moments and cuts that this was AI generated

[rough surge]

Definetly, only one that might fool people is sora 2

[left lodge]

AI generated video to look and behave just like rl will require some magic to be done in the architecture I think

[left lodge]

Things I say also apply on sora 2. Anything which needs a little more movement everything breaks.

[glad hawk]

i wonder why you all so desperate to make ai videos instead of real vids

[left lodge]

But still it's good if you just want to have fun

[glad hawk]

in the next years all you will see is AI slop

[left lodge]

I am not, I literally don't generate videos because I just don't have the usecase but knowledge about current technologies is what I like.

[left lodge]

It all depends on people who create , AI don't create slop by itself

It do what it has been told to do

[golden ocean]

SO REAL the last part

[glad hawk]

not really

it do how people coded that model

smh

[left lodge]

¯_(ツ)_/¯

[golden ocean]

-# 🔒** Message has been Redacted.**
-# Discord now requires ID verification in order to see certain messages. Learn More

[left lodge]

Bro 💀
We are cooked

[astral helm]

damn, i had the urge to click it

[golden ocean]

true

[left lodge]

What did you say there lmao

[winter hazel]

da heck i also see this

[golden ocean]

no way

[winter hazel]

even tho my age group is adult

[high ginkgo]

this chat is funniest thing ive seen today

[winter hazel]

what did you even try to say that got redacted

[left lodge]

Lol , it looks like you paid for that accessory on your pfp
They said paying users will not suffer from this.

[winter hazel]

i used my orbs

[left lodge]

Idk ,you can earn whatever those are ,orbs?

[winter hazel]

you can get orbs on discord quests

[left lodge]

It redirects to
https://discord.com/guidelines

[winter hazel]

it does lead to the official discord guidelines

[winter hazel]

oh

[golden ocean]

true

[left lodge]

You verified or that the thing hasn't rolled out to you yet?

[misty vault]

not rolled out probably

[abstract tundra]

seedream 5.0 when?

[left lodge]

I don't even see any option to verify lmao

Just wanna look but idk where it its

[misty vault]

im dead

[winter hazel]

it should be on user settings and "my account"

[left lodge]

What model is this bro 😭

[pulsar crystal]

it is a fake message

[left lodge]

No its not

[winter hazel]

no its not

[pulsar crystal]

give 5 min i will prove it

[left lodge]

Technically I can sent it too lemme try

-# 🔒** Message has been Redacted.**
-# Discord now requires ID verification in order to see certain messages. Learn More

Yeah

Lol

[winter hazel]

this is what they originally said

[pulsar crystal]

u see fake

[left lodge]

Could be but technically this is the future lmao

[left lodge]

Its the damn, ernie-5.0-0110

[golden ocean]

https://cdn.discordapp.com/attachments/1294096350589353997/1447527369974743160/2392.gif

[rough surge]

Bytedance has removed their seedance 2 in playground

I've found some articles explaining why but they are for sure fake af

[spare rune]

🔒 Message Hidden
-# Discord now requires age verification to chat. Learn More

[rough surge]

age verification lol

is this only in EU?

so baby triggered that, hey baby

you are my baby

[left lodge]

People are just playing

Like

-# 🔒** Message has been Redacted.**
-# Discord now requires ID verification in order to see certain messages. Learn More

[rough surge]

oh is that a joke?

lmao

im not a heavy discord user

[golden ocean]

are u sure about the heavy part

[rough surge]

xD

Now that you say it

[golden ocean]

🗿

[left lodge]

😭 lol

[golden ocean]

i'm a new soul, i came to this strange world, hoping i could learn a bit about how to give and take but since i came here felt the joy and the fear finding myself making every possible mistake

[left lodge]

Guys I have a question, is this image AI generated or real? If you think this image is generated, show me the reason of that conclusion and if you think this is real show me the reason for that too.

I recommend downloading it and seeing it closely

[obtuse heart]

AI I think

[surreal zephyr]

You need Discord Gold® to view this message.

-# Learn More

[obtuse heart]

Bro messed that up

Sob_AE

[surreal zephyr]

Im on phone shush

> You need Discord Gold® to view this message.

-# Learn More

[brittle hollow]

The details on each of the glass items are uneven throughout and warped in many areas so I'd say AI

[shrewd citrus]

i mean the fact that we need to go into that much detail is crazy

there needs to be like a universal hidden watermark

like Gemini has one but only Gemini can detect it

[obtuse heart]

That would be ideal but theres already a website that removes big G's SynthID

[spare rune]

Imagine if it was like one of those image loggers

Too bad I already downloaded

[golden ocean]

big G

[surreal zephyr]

Looks like shadow to me

I wouldnt say thats enough proof

Also no synthid so not nano banans work

Id say its real

The lightning is consistent

[left lodge]

Those require going through a custom link.
This is upload directly to discord

[final hull]

Hi, what kind of text file does claude 4.6 support? I embedded txt file but arena didn’t want to generate it

Claude 4.6 thinking

[toxic verge]

That photo looks real only one thing that sticks out if anything

The way it shot the composition is kind of weird though

These are tough because I consider these more like a optical illusion

[toxic verge]

Nawh coming to all. Discord soon

https://www.bbc.com/news/articles/c1d67vdlk1ko

https://www.lowyat.net/2026/382782/discord-age-verification/

None of this wouldn’t be possible if it wasn’t thanks to AI

[weak dagger]

Is ts real

[weak dagger]

I verifyed my phone number do i need to do it

[toxic verge]

Not sure.

Ai will kill the internet as we know it

[north obsidian]

Rob lox 2

[toxic verge]

https://wayforward.archive.org/

Try that

What’s coming next slowly but surely

[surreal zephyr]

says loading forever

[toxic verge]

That’s part of the thing the experiment

Wow

[surreal zephyr]

no like the site is bugged

🤣

its stuck like this

for 5 mins already

[toxic verge]

Look in the future lm arena is pending legal process

[surreal zephyr]

oh it lacks dns

[golden ocean]

look on claude.ai in the future theres claude opus 7.5

[surreal zephyr]

nah its just stuck forever for me bruh

[toxic verge]

=/ lame

That’s lotta fun of course nobody knows what the Internet really is gonna be like but just the trajectory based on the way it’s going currently

[surreal zephyr]

nvm it loaded after 10 mins

LOL

naw this is so real

[toxic verge]

Can you imagine if that’s the Internet one day

[surreal zephyr]

it is already

craz

[left lodge]

This image is generated by unami
A codename model

[wheat sky]

Hello

[toxic verge]

It’s pretty good

[cedar tide]

Soon in arena ?

[bleak lake]

🔒 Message Hidden
-# Discord now requires ID verification in order to see certain messages.

[left lodge]

I have a better one for you...

-# 🔒** Part of message has been Redacted.**
-# Discord now requires ID verification in order to see certain messages. Learn More

Glm 5 is out on their official site hmm

[bleak lake]

lol

[bleak lake]

Indeed it's better

[loud verge]

https://tenor.com/view/nioce-gif-4791224989187637050

[proud bobcat]

HEAR YE

HEAR YE

GLM 5 IS RELEASED

[proud bobcat]

Uh huh.

Mr ai master over here

[proud bobcat]

I think this shows that other models have more native reasoning capability rather than dataset training only

[thorny cove]

how do i fix net::ERR_BLOCKED_BY_CLIENT

[shrewd citrus]

is it actually hidden 💀

[bleak lake]

-# ❌ ** Discord has decided you are a bot, And you should not access this message for security reasons.**
-# click here to Learn More

[shrewd citrus]

😭😭

[obtuse smelt]

is happening hidden message

[violet thunder]

WHATAFU. ARENA! FIX THIS

[golden ocean]

Fortnite, we need to talk.

[violet thunder]

AYO. FIX THIS

[proud bobcat]

Browser issue

Uhhh

Clear cookies

Or try diff model to see if it works

[violet thunder]

I LOST 2 ACCOUNT WITH PROMPTS

LIMIT*

-_-

[proud bobcat]

Yeah then clear cookies

Most likely that’s the issue

[violet thunder]

i lost account

if i clear cookies

i can lost history

[golden ocean]

choose an option bro

gotta pick a side

no prompts or no history

🥵

[honest verge]

It seems deepseek quietly released their new model because now deepseek says his knowledge cutoff is may 2025 even though it was January 2025 days ago

[proud bobcat]

YEAH I JUST SAW THAT TOO

Oh this day is so peak

[honest verge]

But what's the point of quiet release