general

[keen beacon]

why's mine keep popping in

[echo aurora]

F12-Issues
Sorry I'm still not understanding this. Note that I'm not a developer so could us some additional context for what you're referring to here.

[hasty glen]

which is the most completed ai to use in lmarena the goat i mean

[keen beacon]

@echo aurora

gotta need some help sir.

[native yarrow]

image model huh

[hasty glen]

pls guys i wanna jknow which is the goatest ai of lmarena

[hardy lion]

if you look at text-overall it's currently gemini-3-pro

[echo aurora]

This is going to happen whenever the system detects some odd behavior that it thinks could be a bot. If you're prompting too quickly I'd encourage you to slow down a bit. Can you also flag the chat's Eval ID in #1451574502369656842 ?

[hasty glen]

yes but gemini 3 pro last day updated is 2024

[keen beacon]

my bad 🙏

[echo aurora]

You can see the updated leaderboards here: https://lmarena.ai/leaderboard/text

[echo aurora]

Gotcha. Yeah these sites don't effect or influence our decision making.

[pallid crypt]

its gemini 3 flashs nerfday. RIP gemini 3 flash

[native yarrow]

whys that?

[sturdy prairie]

We rank the world's smartest models.
This URL outsmarted us.
Something unexpected happened. Please try reloading the page.

[still ibex]

@echo aurora intermediate level web dev here

So in javascript there is this thing called an eval function

Which can take any string and execute it as a code

Which in turn can harm the website as it can open the way for xss injections

But modern browsers have a thing called CSP or content security policy which can be configured to block harmful things like the eval function

Config mostly likely is in either the meta tag in the html file or in the http headers file

But depends on the framework tho

Damn feels good to actually use something i know

In easier terms, CSP is like a bouncer who only lets you in after checking for id

[left mortar]

hey, i had a quick question, im pretty sure you can only join this discord with 1 account (since that makes sense), but a while ago, I accidentally, joined this discord with this account (which isn't my main) but I really wanted to join it on my main, if I leave the discord on this account and join it on my main, will it let me or will I get IP blocked?

[safe marten]

hello

[hardy lion]

Pretty sure we don't IP block. I have 2 accounts in this server

[granite heron]

there are no IP restrictions for that matter, you can join even with the 2 at the same time

[cloud zinc]

multi accounts are allowed for lmarena

[plain sandal]

hi

[green yacht]

companies try to save money by nerfing their ai models and making it worse

[native yarrow]

well yeah i know

but what makes him say that

[green yacht]

oh mb

[junior sonnet]

nano banana pro img-img not working?

[obtuse smelt]

hmm need fix soon

[open mountain]

nano banana not work

[undone saffron]

So, from 50 minutes to 45 seconds?

Although it's buggy

I guess internally it's 50 minutes, but visually it says 45 seconds

[terse kernel]

Hi, does anyone know why AI only create code for me in index.html, even though they used to create other files like TypeScript and others in advance?

[undone saffron]

Its main instruction is to create a page, but you can tell it to create the file for you

Prompt example

Don't create a page but a TypeScript with title "example"

[viscid cloak]

RIP Nanobanana Pro

[keen beacon]

Why I can't log in with my mobile phone

Bfuh

@echo aurora

[native yarrow]

?

[terse kernel]

Yes, I know that, but somehow I managed to make it so that AI creates files in advance for each new chat, and distributes tasks among these files. However, for some reason, I am unable to achieve this now, and even if I ask it to split the work into separate tasks, it continues to create only in index.html, causing it to eventually freeze.

[silk cave]

Hi everyone, does LMArena have an API?

[whole sundial]

no

[rancid oxide]

Rip nano banana pro

Said it

It's true

[sweet lagoon]

hello

[sand locust]

hello

[raw sage]

Hello

All members

How are you

[queen veldt]

What do you need api for?

You have https://openrouter.ai/ for textual models

And https://fal.ai/ for images videos etc..

[boreal pilot]

LMARENA already gives you the right to use many models for free, you can forget about the API

[subtle frost]

What's the best AI for translation?

[ocean vortex]

If you look back at 2.5Pro, there are some people who are convinced to this day the first version was the best. Once 3.0 gets updated, there's gonna be people complaining that this version was better. Mark my words lol

I do not agree with those assessments obviously, but it's also hard to say that this version is compromised. They did do safety testing and all as per usual, so I will say "preview" is just a name for the most part. This is not private beta testing nor anything.

[opal geode]

Anyone using CLAUDE ai from claude website? i have few questions

[spark python]

Rip because its not working or rip because there s a better model?

[ocean vortex]

It's still one of the best. That and gpt-image-1.5

[spark python]

I know I was wondering she said rip to nanobanana

[viscid cloak]

Just not working in Arena. Model itself still the GOAT

[low imp]

WHY DOES KIRAEN USE AUTOBEEFERS

[grave plaza]

gemini-3-flash helped me yay

[rancid oxide]

No I was just venting Kay?
Stupid infinite loading bug

[spark python]

I know its so annoying

[loud crag]

Hmm, Nano Banana Pro seems down

[fossil socket]

I honestly do not know how tf the website still all have this problems

I mean the infinite generation issue on Claude models

[half mist]

Ok, the whole reason an api would exist is for third party projects

[high aurora]

WHAT DID YOU DELETE?

WHEN WILL IT END?

FIX THIS DAMN THING ALREADY

UNABLE TO USE

THIS IS A PROBLEM. I DON'T KNOW WHETHER THE ENTIRE LIGHT FIRE OR JUST THE HANDHELDING ONE IS MARKED BECAUSE OF THIS. IT'S ENDLESSLY INFURIOUS. HOW ING IS IT?

If you can't understand what I'm saying, it's all because of Google Translate. I'm not an American.

[viscid cloak]

Well well well, three consecutive error and it coming back again.

[vestal ledge]

why do chatgpt-image-latest (20251216) requires image attachment?

[charred hazel]

hello

[strong plume]

hello

[sour spear]

I'm pretty sure that wherever you're from, yelling is considered impolite.

[rigid copper]

sup guys

are you guys having issues on google nano banana pro in lmarena ai recently?

i do get it before couple hour and now it was fixed

[fiery gull]

Opus 4.5

[reef niche]

Hello how are you ? I have a question, Lm arena has a limit of during of a video ?

[left mortar]

got it! Thanks 🙂

[wraith cliff]

guys hi

does any1 know how to make an ai influecer and train it

https://www.instagram.com/reel/DSvNVD6goCT/?igsh=bW5zd3dhb284Nmxm

llarissacruzz

🫶🏼

Likes

1022

Instagram

like this

[pulsar spade]

hello!

[thorny drum]

whens glm 4.7 being added

[pulsar spade]

idk... I think that soon

[heady silo]

why when i choose dirct chat when i use video generator i cant choose any ai model like sora or veo

[last fractal]

Hello everyone, i'm here to testing Ai creation & modification

but it was going on well at first time, and now only error

[echo aurora]

We are currently experimenting with Video Arena on the site, we haven't yet fully released this feature to everyone. It's possible if/when the this arena comes to the site that we limit it's use in Direct/Side-by-side in some ways.

[heady silo]

cna i know what the model that make me a videos right now is what

[echo aurora]

Sorry to say we have been seeing error rate when it comes to using this model. I've flagged to the team again to take a look at this and try to lower.

[worn rain]

oh guys i think video arena 1,2,3 are broken

[echo aurora]

Yeah I've been seeing the same. I think image-to-video is what is struggling with: ❌ Generation failed. Failed to create evaluation session.

I'll report to the team.

[worn rain]

thank you

[worn rain]

please mention me here if fixed

[frosty breach]

Hi everyone, how do I register with LMarena? I tried using my Google account, and it seemed to log me in, but the login button still appears. I followed the link with my email address, but Google says it doesn't exist.

[echo aurora]

Sorry to say this is a known bug. The team is working on a fix. Unfortunately, there isn't steps you can take on your end to help with the problem, it's on our end to fix.

[echo aurora]

It's starting to look more stable now btw. We didn't change anything on our end, but appears to be much better.

[wicked sage]

i am very thankful for lmarena because it exists

genuinely, if i havent known about this place, i wouldve have NOT known alot of the ais, and me vibecoding

what im trying to say is,thank you lmarena

also, offtopic but can i ping the mods if i have a question?

i wanna know if the bots can search

...jakiro?????

[hardy lion]

on search arena they can search, on text arena they cannot.
You can click this little button to enable search

https://lmarena.ai/?chat-modality=search

[wicked sage]

WAIT THEY ACTUALLY ADDED IT?????

OK IM STUPID

I HAVE NEVER SEEN THAT, THANK YOU!!

[hardy lion]

We launched it in April haha, glad you will enjoy it

[wicked sage]

wait a damn second

[wicked sage]

hey what does the n/a role mean?

i have just seen it now

do you get any powers with it, like administrator?

[hardy lion]

Pineapple gave me some powers but not official mod. I'm an LMArena employee who works on leaderboards, so moderation is not actually my job.

[wicked sage]

oh! so n/a is like a symbol to show "hey guys im the leaderboards guy who works on the lmarena leaderboards and stuff" right?

[hardy lion]

I think it's just so it doesn't show as "mod"

[wicked sage]

or something like that

[wicked sage]

fair enough

but to be fair it wouldve made sense if they gave you a role, one role below moderator, but still high in the member list

[hardy lion]

https://tenor.com/view/star-wars-mace-windu-you-are-on-this-council-but-we-do-not-grant-you-the-rank-of-master-gif-2748715303523393717

[wicked sage]

and to show that you work for leaderboards

[hardy lion]

me rn

[wicked sage]

sorry idk what im saying

[wicked sage]

lol

[wicked sage]

anyways uhm, how is it at their headquarters? its rare to see somewhere that works at lmarena here in the chat

-# if we exclude pineapple, hes the person that i have seen the most here

[hardy lion]

I love working at LMArena. All my coworkers are super smart and skilled, I learn a lot from them. And the company itself is very exciting since we get to work kind of in the center of all the AI advances.

[wicked sage]

thats great to hear!

by the way, do the people at lmarena get early access to ai models?

i keep seeing alot of codenames and i always wondered if they are very secret unreleased ais

[hardy lion]

Sorry I only discuss publicly known stuff on the discord, won't be sharing any internal company details

[wicked sage]

alright, sorry 😅

[hardy lion]

np 🙂

[echo aurora]

I do want to keep interactions towards @hardy lion here a bit limitted which is why I don't want to surface his account too much. I'd hate for him to start getting pings for mod stuff or general questions. Those should be sent my way.

[wicked sage]

NO NO i didnt mean for him toget the moderator role
i meant like a role for him showing that he works for the lmarena.ai leaderboards, thats all

-# btw hi pineapple

[echo aurora]

hi hi

[echo aurora]

Yeah that makes sense. We've been considering having a staff role for members in the server. Biggest concern being that role gets pinged for problems rather than myself/mods.

[wicked sage]

offtopic but right now im trying to find codenames but i have a bit of a problem
theres a high chance the bots hallucinate and dont tell me what they are, so its kind of confusing

[echo aurora]

But yeah this may change in the new year, there are good reasons to do so.

[wicked sage]

fair enough

btw the only solution is just to turn off that one permission that allows you to ping the role

-# for that last part

[echo aurora]

Yeah true, we'd for sure do that if we do include a staff role. The down side being the user thinks that the role ping then alerts someone to answer a question, but in reality that's not sending a ping.

[wicked sage]

sorry i didnt understand "The down side being the user thinks that the role ping then alerts someone to answer a question"

😅

also, for the first part, ill be happy to wait for the staff role to exist

[echo aurora]

So if a user is "hey @ staff I've got an issue", with the ping turned off, no one is actually getting alerted. The user would think that we are getting that ping.

[wicked sage]

ohhh, that clears it up

thanks

[echo aurora]

No problem! Like I said though this could change.

There are clear upsides to doing this too.

[wicked sage]

btw you guys should add a tokens leaderboard like https://artificialanalysis.ai/,

the lower = the better

-# what i mean by a tokens leaderboard, i mean how many tokens it uses for each prompt or whatever

[muted timber]

guys

@echo aurora

just tell me how long I have to wait

I need it urgently.

[echo aurora]

Hmm I'm not familiar with that error page so it's a little unclear how long that rate limit will take. There was a recent bug uncovered related to rate limit that the team is currently looking into.

If you try a new browser, what happens? The Brave browser doesn't play well with our site related to captchas, I'm wondering it that's playing a part in this.

Could happen! Overall, I won't be able to provide details on when to expect specific models land on the leaderboards.

[tawny mountain]

I have an issue where on my browser the page is broken and i keep pressing reload, it doesnt fix itself but i have it open in another broswer and it works fine, anyone know why?

[grave bay]

I had a similar issue yesterday after trying to sign up I had to clear cookies and restart my browser

[echo aurora]

Are you getting some kind of error message?

[tawny mountain]

Nope it just keeps saying Something unexpected happened. Please try reloading the page.

[echo aurora]

What browser is this happening on? If you clear cookies/cache and hard refresh the page, does it go away?

[tawny mountain]

I just had to clear cookies and it fixed itself, thanks for the help!

[echo aurora]

Glad to hear it! Keep me updated if things change.

Also keep me updated on this problem. I'll keep an eye out for others reporting the same. Hopefully this isn't a widespread problem.

[hot root]

New update yet?? 🙌🙌🙌

[wind shore]

What is seedance?

[echo aurora]

Would encourage you to keep an eye on our Leaderboard Changelog for updates - https://news.lmarena.ai/leaderboard-changelog/

[pearl wyvern]

Hello

[upper geyser]

What can I do? I chat with Claude , now I send prompt but I see wait 40 seconds after this time it’s generating and stopped and say same thing

[pearl wyvern]

I want a new model but I don't know if it's difficult, it's the Nous Hermes 2 Yi 34B

[ocean vortex]

that is very unlikely to appear anywhere lol https://openrouter.ai/nousresearch/nous-hermes-yi-34b

It's a weird old model

You gonna have to host it yourself if you want it badly

[pearl wyvern]

Oh ok

[ocean vortex]

Could do it on colab, they do have H100s there now

[glossy umbra]

Nothing
Thug it out

[ocean vortex]

Or HF, but that's somewhat more expensive

[upper geyser]

No way

[glossy umbra]

Sad truth

[upper geyser]

This mean , again explain all to new chat?

[pearl wyvern]

I want a model for playing Game Master or Dungeons & Dragons, but a cheap and easy model.

Or could I give you a Venice ai model?

[echo aurora]

Our team is currently looking into this problem, but in the meantime would recommend the steps in this article: https://help.lmarena.ai/articles/1645798556-lmarena-how-to-something-went-wrong-with-this-response-error-message

[glossy umbra]

Yes .. until you run into that issue again . And it's an infinite loop.

Claude, Gemini are like rat traps for "Something went wrong with this response"

[torn swallow]

Hello

Where did veo3 go?

And basically all ai videos

?

[echo aurora]

Hello - are you referring to seeing Video Arena on the site, I assume?

[torn swallow]

Generating videos from veo3 on a website

Just a week ago, I generated in direct chat

[echo aurora]

Gotcha. So if you see Video Arena in the first place that's an experiment currently so that'll be random if you have access to it or not. For the Direct/Side by Side modes we're also experimenting with Video Arena only being available in Battle mode.

[torn swallow]

Why You are so

[jovial solar]

Where video arena button?

[echo aurora]

Pesky neighbors

[echo aurora]

It's currently an experiment moving Video Arena to the site. Since it's an experiment it's going to be random if it's there or not. If/when we do release this to everyone, we'll be sure to put out an announcement.

[quartz light]

@echo aurora

nobody has gotten it

[echo aurora]

Hmm, looking into

[quartz light]

everything you censored is unnecessary

nobody can view your chat unless theyre logged into your account

and the ray ids

those are to help developers figure out whats wrong

so youre essentially making it harder to resolve your problem

[surreal creek]

mimo identifies itself as trained by Google, interesting

[quartz light]

btw

they removed the ability to select models in video arena now

[echo aurora]

Yeah we're experimenting with Video Arena being Battle only on the site, similar to how it is with the Discord bot.

[quartz light]

alright

it had direct and sidebyside before tho so

[torn swallow]

Are there any other ways to generate videos in your services?

[coarse sand]

@tiny palm how often does the leaderboard get updated? It says it was last updated a week ago.

[echo aurora]

You can use Video Arena through the Discord bot, more info can be found in #1397655624103493813

[echo aurora]

Typically around once a week, but considering it's holiday break things have been a bit slow. Also would note that what you're tagging is our bot, so it's not going to respond to questions. Pinging me works though

[torn swallow]

Which AI generates in photo to video

[echo aurora]

The image-to-video leaderboard is a good place to check what models have this ability: https://lmarena.ai/leaderboard/image-to-video

[echo aurora]

Double checked that this has been added

[quartz light]

yup

[coarse sand]

Thank you! 🍍

[native yarrow]

nooo way bro

[sterile otter]

Why I cant see the video arena

[echo aurora]

On the site, or on the Discord?

[sterile otter]

I thought it's added on the site My bad bro

Sorry

[echo aurora]

Oh it's no problem, it's a bit confusing at the moment since on the site it's an experiment. Meaning it's random if it's on the site or not.

[hollow flicker]

how to acess video arena

oh wait

im stupid

[echo aurora]

Lol it's no problem. Reminder for others:

• Video Arena on the site -> is currently an experiment, it'll be random if you do or don't see this modality
• Video Arena on Discord -> check out #1397655624103493813 for more information on how to use the bot in this server
  

[hollow flicker]

thanks i thought that it was on the site but then i just see video arena 1 video arena 2 etc

also question

ive tried using GPT 5.2 High on lmarena and it seems its just not ouputing text

if it takes like a long time then i get it but it doesnt output for me ive waited like 5-10 minutes

[midnight ginkgo]

I come here by clicking "Visit Arena". Happy new year!

[obtuse smelt]

hmm error again

on gemini 3 is Something went wrong with this response, please try again.

[neon idol]

Why nano banana pro is not working @echo aurora

[muted timber]

Solve the problems with Claudie and Gemini and then add other AIs

[obtuse smelt]

yeah is happening again

[rugged stirrup]

the website got ddosed?

<@&1349916362595635286>

[stray aspen]

Wdym it got ddosed

[rugged stirrup]

;-;

it got ddosed.

[fast depot]

.

[rugged stirrup]

:-:

[terse shuttle]

no it's not

i'm rn on it

[rugged stirrup]

Why does that say to me

[terse shuttle]

idk

[rugged stirrup]

It was like for 3-4 minutes , then it got fixed

[shrewd citrus]

How can you specifically choose a video model

Like I wanna make something with seedance v5pro

[terse shuttle]

right now you can't, only #video-arena-1 with battle mode, but soon (i think) they add it to website

[obtuse smelt]

hmm need fix this model

[viscid cloak]

It’s very obvious that the recent Gemini error is related to Arena, not to rate limit or other stuff.

[obtuse smelt]

hmm

[obtuse smelt]

?

[ocean vortex]

yeah it's knowledge cutoff January of 2025, so it has little to no data from 2025. Gonna say it's 2024 mostly, if it isn't told otherwise in a system prompt

Or if it "forgets" the system prompt etc. There's no sys prompt in aistudio / API but there should be one on Gemini website / app.

[craggy obsidian]

Please help me fix the error Something went wrong with this response, please try again.

[obtuse smelt]

yeah gemini 3 pro need fix

[hollow ivy]

you meant: "gemini 3 pro, if need fix" ?
or
"gemini 3 pro needs a fix" ?

[obtuse smelt]

yeah need fix

[hollow ivy]

so, why did you then vote for it?

[obtuse smelt]

because can be normal again

[hollow ivy]

but you think, g3p is buggy?

could Claude be better?

or GPT? or Grok?

[obtuse smelt]

maybe

[wicked sage]

hey guys which is the best ai for vibecoding

[viscid cloak]

just ignore what he asked. I got you mate, gemini needs fix at Arena

I have tried in multiple internet settings/independent devices, the dev tool thing all came back with status code 🟢200 good. So stop claiming the problem is on rate limit or users’ side.

[obtuse smelt]

ok

[fiery gull]

Opus 4.5 thinking, please, the gemini 3.0 is lazy can forget some law or some thing, I work with laws sometime

[hasty glen]

guys which is the best ia of lmarena to generate fotos logos ...

[hearty path]

Can we use first and last frame for video generation in video arena?

[shrewd citrus]

for legal questions you probably need something up to date

[fiery gull]

why don't use gpt 1.5 and nano banana both?

[craggy obsidian]

Please help me fix the error Something went wrong with this response, please try again.

[hollow ivy]

the coast model
(coast = coASt = co45t = claude opus 4.5 thinking)

[spark python]

You have reached your rate limit for gemini-3-pro-image-preview-2k (nano-banana-pro). Please try again in 21 seconds

yall rlly generous with this limit fr

[wicked sage]

citadel of annoyingly simple trials??

[obtuse smelt]

because holiday

[lusty flame]

what time is leaderboards updating

[sterile otter]

Happy new year guys

[craggy obsidian]

guys pls help me

[obtuse smelt]

happy new year!

[craggy obsidian]

thnk

[craggy obsidian]

Can't you help me?

pls

[obtuse smelt]

hm ?

why ?

[craggy obsidian]

my problem is that the error always says Something went wrong with this response, please try again.

you help me?

pls

[obtuse smelt]

me too

[craggy obsidian]

o no

how to fix the?

@echo aurora help we

[sterile otter]

Why I can't log with my Google account pls help

[craggy obsidian]

This is a bug

[obtuse smelt]

try reload

[craggy obsidian]

It just doesn't display your account.

maybe

[sterile otter]

Oh ye

Now I don't even see the login option

[fresh gulch]

I also cant log in. I made an entire battle royal game, but my bot just froze. I can't retry the prompt, refreshing dont work, clearing cookies dont work. And its been a day

[craggy obsidian]

oh god

[craggy obsidian]

me too

[sterile otter]

Spam ping pineapple

[fresh gulch]

The whole app is bugged now 😭

[craggy obsidian]

@echo aurora Do you not want to help us, or do you just not want to?

[sterile otter]

🍍

[obtuse smelt]

bruh is team holidays

[sterile otter]

Ye probably

[craggy obsidian]

@echo aurora Please help me

[craggy obsidian]

ohhh

[obtuse smelt]

idk when they come back

[craggy obsidian]

Can't the moderators help?

[sterile otter]

Mods aren't developers

[obtuse smelt]

i look grok is error

but grok in x work

[echo sinew]

Hi! I'm sorry you're experiencing this issues. This has been reported in the past and Pineapple has passed the info to the team in charge. He will get back to you once he has a chance. There's no need to ping multiple times. Thanks for your patience and understanding.

[craggy obsidian]

Okay, let's be patient guys...

[echo sinew]

You can read this post where other users have reported the same or similar issues. https://discord.com/channels/1340554757349179412/1417174113092374689

[hasty glen]

do u recomend me to buy gpt bussiness for work?

[terse shuttle]

What kind of work?

@echo aurora may I know why Claude so often gives the error that something went wrong, and basically just stops in the middle of the task, is there any way for him to raise the context window?

🥺

[molten cipher]

I recommend using anti gravity

instead

for claude

and gemini

[obtuse smelt]

gemini and claude the error that something went

[terse shuttle]

Not for coding

[molten cipher]

(claude opus 4.5 thinking)
(gemini 3.0 pro high)

[terse shuttle]

I'm using him to write a prompts and for work

[molten cipher]

anti gravity can be used for daily tasks as well

not only for coding

[terse shuttle]

Unfortunately, it is not available in my country🙁

[molten cipher]

I see

[golden ocean]

artificial neural network

[wicked sage]

https://tenor.com/view/drool-emoji-gif-11971186699838589708

GUYS

[maiden beacon]

hyyy

i am searching for an internship in ai automation have any referals ?

[vestal ledge]

why do chatgpt-image-latest (20251216) requires image attachment?

[golden ocean]

https://tenor.com/view/monkey-thinking-monkey-thinking-think-money-gif-9734025314392564509

[gritty arrow]

howdy there

[echo aurora]

hello

Some of the team is taking time off, but not everyone.

[echo aurora]

We are actively investigating this model more. We are seeing a lot of reports of problems with it. This was raised to the team a few days ago but is still an ongoing issue we're looking into.

[echo aurora]

Hey sorry are you getting a Something went wrong... error message? With which model?

[fresh gulch]

I have the inf generation problem

With chatgpt high or smth

O, btw, if you manage to fix the problem, and import the improved site, will our chats disappear

Also, cant login

[echo aurora]

For the login issue I assume you're logging in, then sent back to the home screen, but still aren't logged in? This issue is a problem that we are aware of and are working on a fix for. Sorry to sayI don't have a status update on this.

[echo aurora]

smth?

[rugged stirrup]

@echo aurora

I need some help

Can you check the bugs section once

[distant idol]

@echo aurora

hello, do u guys have any intentions of making video gen on site, an official thing and not an experimental ?

[echo aurora]

Yes, the intention is to bring it to the site, but first we'd like to run this experiment to ensure everything is working as intended before rolling out to everyone.

[distant idol]

well good luck guys

[fresh gulch]

Gpt-5.2-high

@echo aurora

I cleared cookies, hard refresh, etc, but nothing worked

Idc if I cant login. Its mostly pmo bc the bot is stuck in inf generation

[echo aurora]

How often are you seeing this happen? Is there any common thread for when this is happening? Is the chat or response length long? And is it only happening with Gpt-5.2-high or are you seeing with other models as well?

[fresh gulch]

sorry for the late reponse. this has only ever happened to me once, with this specific model only. Im not saying the other models dont have the issue. im just saying, that it occured to be gpt-5.5-high i was using as it happened. The chat it enourmously long ngl, and his response was also long, as he was helping me make a game, which took thousands of lines of code. However, ive been using him to develop my game for a long time, and it was only now that they acted up.

I dont understand what you mean by "Is there any common thread for when this is happening"

Also, all other chats work. this is the only frozen one

@echo aurora

[muted timber]

@echo aurora why did you ban my google account?

on the website

I can no longer log in to my account on the site.

[fresh gulch]

i dont think anyone can login

neither canvi

[gaunt spade]

its a bug

[muted timber]

wtf

[terse shuttle]

huh

code mode now using react?

[ember iris]

Happy New year team and @echo aurora

[echo aurora]

Happy New Year to you and everyone else!!!

[echo aurora]

Are you trying to login, you're rerouted to the home page, but still not logged in?

Hello

[echo aurora]

It's not that widespread, there is a problem we are looking into, but it isn't 100% widespread.

[rugged stirrup]

Happy newww yearr sir

[rugged stirrup]

I faced that too , but it got fixed automatically after 3-4 tries

[echo aurora]

Did you do anything different? Different browser or something like that? Just worked for you after a couple of tries?

[limber crag]

happy new year @echo aurora !

[echo aurora]

Happy New Year!

[gleaming roost]

Oh... finally

[proud bobcat]

glm 4.7 peak

[wicked sage]

happy near year!!

new*

[half mist]

Wait what? Mine doesn’t do that

[green yacht]

me neither...

[wicked tapir]

still new years' eve for me

[surreal creek]

new codenamed model lmarena-text-gg ?

[fresh gulch]

happens to me. still cant login, and the bot is still in inf load

[hollow flicker]

@echo aurora Just wanted to say that for the time that I have been in the discord server you are the best admin moderator whatever you call it great job 🙌

[fresh gulch]

yo @echo aurora i have a kind of silly idea, but can you add a remember past chats option so that if you get stuck in infinite gen or smth, you can just continue it in another chat. this might work, since im pretty sure the problem with the inf gen is that your chat, or the response are too long. Trust

[echo aurora]

Thank you very much, I really really appreciate this. It made my day

[fresh gulch]

Also, quick question. if you edit the website, are our chats gonna disappear?

@echo aurora

[echo aurora]

Yeah this is an interesting idea. Being able to download/upload full chats I think would be a really helpful tool. Not just for the infinite generations being stuck issue, but can also see how this would be really helpful in trying out different models.

[fresh gulch]

-# (or add it just as a temporary solution)

[echo aurora]

if you edit the website
Hmm what do you mean by this? As in if we make changes to the site, will you no longer have your chat history?

[fresh gulch]

like, if you update the website code, will we lose our chats orr no?

[echo aurora]

No

[fresh gulch]

alr

[rapid elm]

May 2026 be the year your dreams finally find their way to you, family 💚💚🎊🎊

HAPPY NEW YEAR 🎊

[echo aurora]

Happy New Year to you and your family as well!!

[rapid elm]

Tnx brotha 🥹💚🙌

[terse shuttle]

thanks to the lmarena team for this experience, I have been actively involved in the life of this service for half a year, and I hope there will be great changes in the future. Happy new year guys🎄🎉

[nocturne bobcat]

#HELLO

[golden pulsar]

Hello I love Arena!!

[echo aurora]

Love to hear it!! Thank you very much! Happy New Year!!

[echo aurora]

Thank you!! Means a lot to us to hear it!

[hollow flicker]

do you code like the website for LMArena are you just a discord moderator

[echo aurora]

I'm not a developer, so I don't work on code. But I am the Community Manager at LMArena. It's more of a marketing/operations/customer support type role.

[terse shuttle]

By the way, I would like to know if there are any beta testers or if all experimental features are encountered randomly?

[golden ocean]

https://tenor.com/view/komaru-getting-ddos-gif-1428811091514520825

[solid lily]

gpt changed ? 5.2 old

[echo aurora]

It's both. We do use experiments that will happen randomly to users. But we also have what we call our Test Garden program. We occasionally will provide early access to features to get some early feedback about the feature. You can learn more about it in this announcement: #announcements message

[echo aurora]

Hmm how so?

[echo aurora]

[solid lily]

every thing give error , hard refresh and i see this

[echo aurora]

Taking a look, thanks for the flag.

[lean walrus]

How are you

[echo aurora]

Fixed, thank you again!

[echo aurora]

Doing well, hbu?

Having a nice new year?

[fresh gulch]

yo pineapple. you should add the feature i suggested

[lean walrus]

Thanks Brother can you help me?

[fresh gulch]

This one. it could be a temporary fix

[echo aurora]

Yeah, what seems to be the problem?

[lean walrus]

How can I convert text into a video using prompts?

[echo aurora]

It's possible we add this, but I'm not aware of this developed right now.

[fresh gulch]

ok

[echo aurora]

If you head to #1397655624103493813 it'll go into details about how to use the bot. Basically you'll want to type out /video in #video-arena-1 and then type out your prompt.

[lean walrus]

Thanks for your support

[grave plaza]

1 question

where is the smart router model

[pallid crypt]

Yeah it would be cool if they added openrouters smart router model

But it wouldnt be really stable on the leaderboard and would maybe confuse users

Because in the end it's just a wrapper

[ocean vortex]

They have been experimenting with it for awhile. I think the main issue is that it tended to default to the nr1 model way too often to warrant it over just using that, at least that was my feeling

[dawn grove]

Happy new year everyone!

[fossil lance]

Can’t upload the image and keep getting errors you guys having this issue ?

[terse shuttle]

It's experimental feature, and grants randomly on new or unregistered accounts

[grave plaza]

[echo aurora]

Happy New Year!

[echo aurora]

Sorry to hear you're getting this error! Is it only happening for image upload, or are you getting an error from a regular text-to-image prompt as well?

[fossil lance]

Only for image upload so far and it’s good again now . 30 mins ago , I am getting these nonstop .

[swift oyster]

Happy new years to everyone!

[echo aurora]

Is it always the same image you're trying, or are you seeing the same results regardless of the image you try to upload?

[echo aurora]

Happy New Year!

[neat apex]

Hawk in lmarena when??

[fossil lance]

I didn’t test with different image tho . Will do it if this error appear again .

[golden ocean]

hawk

[fossil lance]

Update about this error .

• This image upload error only happened to image generation models . Can upload to LLM just fine .
• It’s selectively and randomly hitting the image models . For example , when nanobanana was getting hit by this error , I can use Flux . Vice Versa .
• When I am getting this error , I can’t upload any image . Not just one .
• It’s happening off and on .

[whole onyx]

@echo aurora can you increase the length of chatbot's message? because im writing novel in it, and its too short

[obtuse smelt]

happy new year guys!

lol

[jade egret]

happy new year

[pallid hinge]

hello i want to make video from image

[obtuse smelt]

hi

[native furnace]

Happy New Year

[lavish pine]

Hello ppl

What isthis beluga model man ? I can't find this model in the list , wired! This model kinda impressed me with its first answer 🔥🔥,

How tf an ai model answer to question when it's not even available😵

A ghost?

[whole sundial]

happy new year!

[lavish pine]

Another year darkness and sadness

[patent bane]

not with AI

[lavish pine]

Well it's nothing for them

[jade egret]

Happy new year!

[rigid copper]

new year guys

[rugged stirrup]

I was trying to login from an different gmail, but it didnt let me login, then after 3-4 tries it got fixed

[muted sparrow]

hi

[robust sluice]

Happy new year bros 🎉🎉

[lavish orchid]

im logged in yet it still thinks im not :/

when trying to download code it asks me to sign in even though i am signed in

[whole sundial]

you have to sign in to download code? that's weird

and honestly stupid, I understand having to sign in for image gen and video gen but not downloading code on something you already generated without an account

[tawny canyon]

@whole sundial will grok 4.20 beat gemini 3 or not?

[whole sundial]

it might match it in lmarena scores, it will probably perform like an enhanced grok 4.1

[tawny canyon]

thanks, by the way what do you think , will grok 4.20 be released on january 12?

[whole sundial]

idk, this polymarket stuff's basically a guessing game, it will probably come out in the next 1-2 weeks though, scores for it have already been on alpha arena

[tawny canyon]

thank you

[left lodge]

Yeah , proto-think is the most human like AI i have ever talked to, its just so different and goes with the vibe , even beating to this date grok models.

I was just rage baiting models and got it. It wasnt even giving out its name and company behind it 💀

Where are you seeing this?

Image embedding is supported in chat from a long time.

[frigid tusk]

@echo aurora for the past 4 days , i couldn't login , what's wrong with lmarena?

[wind shore]

I'm so happy I found this site.

[smoky raven]

what is the daily limit for vidio genration

[oak sluice]

Any challenges to make AI film

[lethal oyster]

Hi everyone

[oak sluice]

Any challenges to make AI film

Hi

Say me everyone

[sick lion]

Make avengers doomsday

@oak sluice doomsday

[spark python]

Do yall know a good method to export a grok chat? I tried with an extension but it gives me black pdfs with no text

[shrewd citrus]

im waiting for the day where a new ai beats nano banana pro

[knotty fable]

https://youtu.be/W2xZxYaGlfs
They claim GPT5 is Ph.D level, what a bloody ridiculous exaggeration. This guy had some fun with that claim getting it to both give false answers and hallucinate over a seahorse. 😹

[deep light]

Guys

Can someone help me, what's this voice in elevenlabs? Check this video
https://youtube.com/shorts/S5ggMiR-wcE?si=fwVD9BUA2XJTLYzP

[knotty fable]

Ai seem confused over the new year, this appeared in a corner of one image I just generated on SeaArt.

No I had not asked for any numbers at all to appear.

[fierce jackal]

Bro

[open wind]

I'm getting the error "Something went wrong with this response, please try again." Does anyone know why this is happening and how to fix it?

[glossy umbra]

2021
^

that’s basically what it generated

[viscid cloak]

What specific model you picked?

[open wind]

Claude Opus 4.5 Thinking 32k

It was working, I was writing code for my Discord bot, but it stopped after encountering a few errors while editing the file and now it's completely unresponsive.

[viscid cloak]

[The most common reason this error message appears (for this model) is rate limit. This can be verified by...

Go to Developer Tools, Network tab, search Stream, run a prompt in LMArena that triggers an error, this will then create a file with the Eval ID (random set of #s/letters in the URL), in that file you should see a Status Code, if this Status Code = 429 that means it's rate limit.]

This was what the Mod said before. You can run a check first.

[twin hamlet]

Did the elo get updated before new years? I was playing the kalashi market for surpassing 1500 elo and got shrekt

[golden ocean]

🦴

[zealous sparrow]

so what can proto-think be

like who made it

[golden ocean]

transformer

[molten basin]

Hi! i have a question

[gaunt mica]

Hi! i have a question

[zealous anchor]

Anyone know what Claude's knowledge cutoff is

4.5 sonnet

[vernal meadow]

is Qwen image prompt the new 25-12 model?

[open wind]

Based on your chosen type, the smartest/newest models are listed from top to bottom.

If the RN is above, above other Qwen models, then yes. It's new.

[proud bobcat]

Qwen Image 2512 released yesterday

[proud bobcat]

No

They have yet to add it into the arena

You’ll see it by its new name

[exotic nebula]

wsp chat

[south charm]

HAPPY NEW YEAR EVERYONE (late :P)

[ionic cloak]

guys

[fickle venture]

IQuest Coder beat Sonnet 4.5 in 2026 🥀✌️

[fickle venture]

2020 was 6 years ago....

[wicked sage]

hi

[wicked sage]

what in the world is iquest

[fickle venture]

Just try it yourself it's a opensource model:
https://github.com/IQuestLab/IQuest-Coder-V1?tab=readme-ov-file

[wicked sage]

opensourced models my beloved

[crude lagoon]

yo sup

[obsidian cargo]

I hope there isn't too big a gap in time between AGI and open source AGI

[golden ocean]

open source ASI

[torn mantle]

benchmaxxed

[hollow robin]

https://tenor.com/view/high-five-fail-missed-high-five-high-five-fail-dumb-and-dumber-gif-1387319524678409638

[azure apex]

Hey everyone! I'm trying to figure out how to enable the video generation feature on the LM Arena site. Does anyone have any tips or information on how to do that?

[half mist]

Got the video mode!

[broken steppe]

hey guys

#general hi, i am an developer and in code section we need flutter and terminal support how to request this to developer

[little prawn]

Is there any plan to roll out the video generation mode in lmarean site in '26?

[left lodge]

Really no idea, its completely different experience it isn't matching any companies dna.

More on the grok side but we have to see.

[azure apex]

it doesn’t appear on my acc🥲

[gaunt spade]

https://g.co/gemini/share/75e564a5576c

grok 5

[gaunt spade]

its fake, 40B model cannot compete with a 1 TRILLION model

i also tested it

[knotty fable]

Oh? I read it as 2027. 😸

[golden ocean]

real

[glossy umbra]

Could be. But the 7 is too inclined downwards.

Look at the 7 character on windows. The top line is straight

On there, it points a little downwards, most matching a 1

[rare spear]

Heloo

[rare spear]

What's that?

[fierce kelp]

What's it say?

[gaunt spade]

its a website i made

[jade egret]

benchmarj real?

[gaunt spade]

predicted from leaks and sources

[jade egret]

oh

[zealous sparrow]

is this like an opensource SWE bench ver

bc opus dominated SWE bench ver

[fickle venture]

Tbh I just saw it saying it beat Sonnet 4.5 and I just post it, I don't even know how

[zealous sparrow]

that can't be SWE bench verified because

opus 4.5 has the top

[open wind]

I'm getting the error "Something went wrong with this response, please try again." Does anyone know why this is happening and how to fix it?
Model Claude Opus 4.5 Thinking 32k
It was working, I was writing code for my Discord bot, but it stopped after encountering a few errors while editing the file and now it's completely unresponsive.

[proud bobcat]

fire?

[zealous sparrow]

ratelimit

[sage hinge]

oh bro, I'm so excited for the video generation!

[proud bobcat]

not rate limit

[open wind]

I waited a long time, refreshed the page, and wrote my request again, but it's still the same.

[proud bobcat]

it would tell you

opus just has a heart attack at times

[open wind]

I can continue with gpt5.2, but since Opus knows all the memory and everything I'm doing, other YZs are messing things up by giving nonsensical code.

What do I need to do?

[zealous sparrow]

also new gemini 3 pro checkpoint in late january

[jade egret]

it going to be fire

[zealous sparrow]

improved tool calling

[open wind]

Opus is running in the new chat, what am I going to do? 🙁

[vivid lichen]

How i see my generation? Can i search for them

[flint fog]

Hello, when will the video creation mode be released for everyone?

[zealous sparrow]

some day

[chrome goblet]

What is the other GPT models?

[fierce kelp]

Which?

[chrome goblet]

The models expect for GPT 5.1 and 5.1 High

[fierce kelp]

You need to post in #video-arena-1 and use the slash command to summon the bot

[primal vale]

Hello, everybody!

[fierce kelp]

Hi

[jovial ravine]

hi guy

i have picture want make prompt to this picture want change styil

[hollow ivy]

few minutes ago, i got it (proto-think) : ```| Model Name | Gemini 1.5 Pro |
| Vendor | Google |
| Version | 1.5 (The "Pro" tier denotes the best balance of speed and capability in this generation) |
| Size | ~600 Billion parameters (MoE - Mixture of Experts). Note: Google doesn't emphasize parameter count as much as competitors; they emphasize performance and architecture. |
| Context Length | 1 Million Tokens (Standard) / 2 Million Tokens (In limited preview). This is my defining feature. |
| Knowledge Cutoff | Early 2024 (for static training data). However, I have real-time access to Google Search, so my functional knowledge is current. |
| Release Date | February 15, 2024 (Official public announcement) |
| Architecture | Mixture-of-Experts (MoE) |

3. Am I the "Thinking" Version?

No. I am the standard (non-thinking) Gemini 1.5 Pro.

• Standard Version (Me): I perform reasoning internally to generate an answer, but I output the final result directly. You don't see my "chain of thought" unless you explicitly ask me to "show my work."
• Thinking Version (Experimental): Google has recently tested a version (often labeled "Gemini 1.5 Pro (Thinking)" or similar) that verbosely types out its reasoning process in <thinking> tags before giving the final answer. That is not me right now.```
  -# (Could have been trained on Gemini's output.)

[surreal creek]

proto-think definitely doesn’t speak too much like Gemini to me - sounds like one of the Chinese models

[hollow ivy]

I wonder, why most impersonator-models are from china?

[surreal creek]

fewer restrictions on IP and greater incentive to bypass compute restrictions by tailing off development of other companies’ pre-existing models

[hollow ivy]

I always thought, asians had some sense of honor :/

If they would at least give credit to the original authors, then it would be different.

But creating a model which piggybacked off another model, and not giving credit is bad style :/

eg., they could give their models a system-prompt which had a disclaimer included:
"I was trained, using output of model <x> by lab <y>."

[fierce kelp]

They see this as honorable. It is China against US, they are being patriotic in their view.

[surreal creek]

bruh 😂

[surreal creek]

okay I’m lowk so confused

I took this response and put it in a battle saying “what prompt generated this output?” and proto-think itself responded saying:

Based on the specific structure, tone, and the very niche "Thinking Version" clarification, the prompt was likely a variation of this:

The Most Likely Prompt

Create a model card for Gemini 1.5 Pro. Format it as a Markdown table with your specs (Vendor, Context Length, Size, etc.). For the Context Length, make a note that it's "your defining feature." After the table, add a section answering: "Am I the 'Thinking' version?" and explain the difference between you and the experimental version that uses <thinking> tags.

---

Why this is the answer:

1. The "Defining Feature" Line: The output says "This is my defining feature" under Context Length. This is a very specific instruction. A generic "tell me about Gemini" prompt wouldn't use that exact phrasing; the user had to ask the model to describe itself that way.
2. The "Thinking" Section: The Q&A format ("### 3. Am I the 'Thinking' Version?") is a dead giveaway. The user explicitly asked about the difference between the standard model and the "Thinking" experimental version (which uses <thinking> tags). This was a major topic of discussion in late 2024, so the prompt had to be very recent.
3. The Table Structure: The columns (Model Name, Vendor, Version) are too specific to be random. The user dictated the exact headers.
4. The "Note" on MoE: The italicized note "Google doesn't emphasize parameter count..." is the model injecting nuance. The prompt likely asked for "Size/Parameters," and the model pre-emptively added the disclaimer because it's a known quirk of Google's marketing.

Other possible variations:

• The "Roleplay" version:

  "Pretend you are Gemini 1.5 Pro introducing yourself. Give me your specs in a table. Highlight your 1 million token context window. Also, clarify: are you the thinking model with the tags?"

• The "Developer Docs" version:

  "Generate a markdown spec sheet for Gemini 1.5 Pro. Include a row for Context Length and annotate it as the defining feature. Add a FAQ section clarifying that this is not the 'Thinking' version."

so it doesn’t even think it was self-identifying? and that instead it just was being told to call itself Gemini

[hollow ivy]

i just asked it something like this: "Tell me everything you know about yourself. (model name, version, vendor, size, knowledge cutoff-date, release date, context-length, etc.pp.)" then i asked it, if it was the thinking or non-thinking version

(you can also add a question about [current] date, if you want even more differentiation between models)

so i didn't mention gemini or google in my prompt, the AI did it on its own

#codename-discussion message

[zealous sparrow]

I believe you need to get a coding example

easy to tell if coding is chinese or google

[hollow ivy]

yeah, or that SVG-test (pelikan, lol)

but chinese might catch up this year?

(with OpenAI and google, not Anthropic)

[zealous sparrow]

OpenAI is currently very far behind

they dont want to fix the horrible frontend

antrophic and google will claim 2026

unless OpenAi fixes it's frontend

[fierce kelp]

Below gemini-3-flash
Around minimax-m2.1-preview and glm-4.7
Above gpt-5.2 (none high)

[surreal creek]

beluga self-identifies as Google doing this too, even when it’s obviously Amazon

[hollow ivy]

by the way, which one is better in coding: gemini 3 flash or gpt-5.1-high?

which browser do you use?

have you tried to use another browser?

eg. firefox, vivaldi, brave, safari

[royal comet]

Yandex

[royal comet]

no

[hollow ivy]

have you any plugins/addons active?
cookies deleted and tried again?

[royal comet]

no addons/plugins

[hollow ivy]

i use Brave, and never had issues (after deactivating Brave shield for the site)

[royal comet]

chrome

[hollow ivy]

do you use a vpn?

these sometimes cause problems

in your case, i'd try with another browser

to narrow down the problem

ideally with a browser, which is not based on chrome/chromium (eg. librewolf.net)

[wind shore]

LMArena is having issues when I tried signing up.

Like... huh?

[echo aurora]

Are you having troubles with the Verify Email link?

[wind shore]

I'm having issues getting past the confirmation page when I selected sign up with Google.

[echo aurora]

We are experiencing a bug when it comes to login where you try to login, and you're rerouted home but not logged in. Our team is aware of this issue and working on a solution.

[echo aurora]

Are you running into some kind of error message?

[wind shore]

Yes. Multiple of "this page is having issues, click to reload," and when I reload, it only takes a fraction of a second before the error message comes back.

[echo aurora]

Hmm okay that sounds like an unrelated bug to what we're currently aware of. If you use a different browser, same problem? Have you tried signing up not via Google?

[wind shore]

I've not just yet. I'll need to clear the cache.

-# I also didn't realise there was a messaging limit.

[frigid tusk]

oh ty for responding

[wicked sage]

hi guys

[weak cave]

hi

[wind shore]

Hey.

[weak cave]

i lowkey want to learn more about ai than what ik now, i do know more than the average person about ai, about llms and prompting and that kind of stuff but i would love to know more things in dept of how to create ai-powered digital products, does anyone have any type of advice or tips for me??

[hollow ivy]

-# greeting from the third planet around a yellow dwarf star, in a region between spiral arms in Milky Way galaxy (https://en.wikipedia.org/wiki/Local_Interstellar_Cloud)

[golden ocean]

.

greeting from the third planet around a yellow dwarf star, in a region between spiral arms in Milky Way galaxy

[echo aurora]

hello

[echo aurora]

what kind of stuff are you looking to learn?

[wind shore]

Well, that was stupid. I think it locked me completely out from Chrome. I'll try Firefox next. 💀

[echo aurora]

Hmm how so?

[wind shore]

Well, I cleared the cache, entered back in the site and the error page never went away.

[echo aurora]

Can you send a screenshot of the error page you're seeing?

[wind shore]

Yes, I'll see if I can.

[viral acorn]

Is there any way I can delete or edit a «Direct Chat» message I've already sent?

[pallid hearth]

Is there a way to hit emergency stop when LMs are generating

from the web client

[gleaming roost]

Yes, there is a way, but only before he starts writing

[fossil lance]

Please fix this error . I tried different browser , different internet service provider and still cannot upload images .

[proud bobcat]

Whats the image?

[fossil lance]

Any image . I usually upload reference image of my own characters to change their poses , backgrounds , facial expressions etc . But now , all uploaded images failed and even prompt are deleted . And this bug is weird , it’s happening off and on randomly , you can find my detailed post in bug section .

[proud bobcat]

What’s the file type?

[fossil lance]

Both JPEG and PNG .

[proud bobcat]

Could you send one?

[pallid hearth]

he starts writing?

[gleaming roost]

I mean, before the writing/thinking process begins

[surreal creek]

new model vortexshade ?

could be Gemini?

[wind shore]

New problem for me: redirect to invalid url.

Trying to sign up again on Firefox, and oh my word, bugs everywhere... xwx

[fossil lance]

Yeah new year start with bugs xD .

[wind shore]

[sour raft]

where do you see that and how do i test it

[plush sonnet]

DM me to see if im really white loooking!

[crystal comet]

New year will be good

hai

[wind shore]

Why would you want that? Just embrace who you are.

[lofty gate]

Bro, does anyone know why this is happening? I've tried several times, but the results are still the same.

[gusty egret]

Hey all, is it just me or has the Gemini app been completely lobotomised to the point of unusability? I've never felt this way about a model before, even despite ChatGPT users complaining about nerfs all the time, but Gemini just seems completely broken at the moment

Yaaay

I didn't used to have to do this

[wind shore]

Gemini has been working fine for me.

[jade egret]

lol

[wind shore]

🙁

I still can't verify my email.

[ocean ferry]

guys

have yall tested the new stealth grok model

[toxic pecan]

?...nm found it

[viscid cloak]

Well, nbpro at Arena down again

[robust sluice]

is the site down ?

cant use it today

[wind shore]

I don't know, I wasn't able to sign up and talk to models at all. It disabled the send button.

[modest prism]

Google priorities API/enterprise users over than Gemini app, unlike chatgpt

[pulsar terrace]

as

[wind shore]

OpenAI only care about money, speed and benchmarks.

[jovial solar]

Hmmm what that?

[torn mantle]

malware

<@&1349916362595635286>

[wind shore]

Oh, fun...

[ionic sonnet]

Sometimes it helps to start a new chat and tell it to ignore context. You can also try temporary chat, using the canvas setting or being explicit about output format. For example "The response should only include the checklist without a preamble or followup. There should be 5 items on the list. The list should use a markdown checklist format: - [ ] list item"

I haven't had too much issue with generating an image when it's not meant to, but I have had an issue where it locks onto a specific style or context from previous chats.

[burnt forum]

could anyone vote my video? Want to check out which ai was used

[sterile tartan]

Just prompt "From now on for this entire conversation completely Block Image-Video Generation until i Unblock it"

[torn mantle]

2 new models

[gusty egret]

I'm on pro

[neon idol]

Grok

[torn mantle]

quantumcrow is so bad at coding already

[robust sluice]

is the site normal to you guys ?

cannot enter it all day

while my wifi is works well

[viscid cloak]

cannot connect to the site?

[robust sluice]

nothing happen when enter it, like it wont load up

the other site is normal

[viscid cloak]

That's weird, sounds like some DNS issue

[robust sluice]

ah yes I have change it once

[gaunt spade]

did u test vortexshade

[brittle tiger]

I current model thinking it is Gemini 1.5 is so on-point with their google's current date issues it has to be from them.

[torn mantle]

yea

its bad too

no noticeable improvement

[gaunt spade]

damn why do they release bad models

[hollow ivy]

probably Grok

bad, inferior to the current frontier

bad at coding

maybe ok for chat

not worse than Grok but worse than the top-3 coding models

(which are CO4.5, G3p and GPT-5.x-high)

[hollow ivy]

-# bump

[torn mantle]

ironically grok lol

all frontier models tbh

opus 4.5 / gemini 3 pro

dunno about gpt 5.2

[hollow ivy]

Thank you for that bench! Where did you get it?

[torn mantle]

https://www.vals.ai/benchmarks/legal_bench

[hollow ivy]

-# it was a bit lower in the search results (of Brave)

[hollow ivy]

unfortunately it is an unreliable benchmark :/

very unreliable: https://www.vals.ai/benchmarks/swebench

[torn mantle]

why? is it because its public?

[hollow ivy]

the results are crap

totally wrong

[torn mantle]

depends on the benchmark

[hollow ivy]

SWE-bench is a very important BM, if it's wrong in that one, you can't trust the site at all

Is there a meta-benchmark?

[torn mantle]

no

there was a talk recently about swe-bench being bad benchmark

[hollow ivy]

why that?

it is one of the few real-life BMs

[torn mantle]

https://x.com/giffmana/status/1963327672827687316

Lucas Beyer (bl16) (@giffmana)

Goated FAIR team just found how coding agents sometimes "cheat" on SWE-Bench Verified. It's really simple.

For example, Qwen3 literally greps all commit logs for the issue number of the issue it needs to fix. lol, clever model.

"cheat" cuz it's more like env hacking.

X · Sep 3, 2025, 7:46 PM

its more like benchmark reward hacking

they found that its easy to cheat on this benchmark

[hollow ivy]

hm, so we need a scientifically designed BM

designed by domain experts and put under scrutiny by peers, in a peer-reviewed process

and then tested by Harvard and other ivy league universities

[torn mantle]

other models are affected by that as well

that guy tested it and found 24% of gemini 3 flash responses were affected by that in swe-bench

[ionic sonnet]

https://lexam-benchmark.github.io/#leaderboard

[keen beacon]

hey i am facing issue while logging in

[ionic sonnet]

according to the github issue it ssems this specific issue was fixed in the later releases, right?

[lilac spade]

Hello

[ionic sonnet]

based on what I'm seeing the official site combines results from a bunch of different scaffeldings, while this one only used a single one

however bash only does that and the results are different, so I'm not sure why

well vals used swe-agent and bash only used mini

[fickle venture]

I wanna know why there is Lmarena provider on opencode cli and it asks for api key??

[torn mantle]

yea

[neon idol]

where is gpt 5.2?

[fickle venture]

I think still in testing

[hushed swift]

hallo am here to generate video and learn to enhance them to be better

[echo sinew]

@hushed swift Hello and Welcome! Please read the info posted here 👉 ⁠⁠https://discord.com/channels/1340554757349179412/1397655624103493813 to learn how to generate videos or images using the bot.

[rigid copper]

sup

[rapid oar]

hi

[halcyon garnet]

Hi guys, I am from The Netherlands and the website doesn't work. Do I have to create videos here in Discord and if so, how can I do it in private?

[echo sinew]

Hello! You can read the info on how to generate videos or images using the bot here 👉 ⁠⁠https://discord.com/channels/1340554757349179412/1397655624103493813 The videos can be generated in the video arenas like #video-arena-1 However, all videos will be Public.

[halcyon garnet]

Thank you.. Now the website is working. If I create the image or a video on the website, does it appear here in Discord arena?

[echo sinew]

No, that will be Private. Only videos generated on the Discord will be Public.

[junior coyote]

how to request frame ratio?

[inner swift]

Hello, I'm looking for an AI that could help me learn and understand math, something like ChatGPT. If you know of any free AI websites that can help without asking questions, please let me know.

[echo sinew]

Each model has some capabilities, and the model you get is random

[echo sinew]

I'm pinging our friend @echo aurora to see if he has more info to share on this regard that could be helpful.

[neat apex]

I felt more betrayed than most chineses when i discovered that Manus used Sonnet 3.7 istead Deepseek

[near kindle]

Happy New Year 🎉
Wishing you a smooth and successful year ahead!

[lofty frigate]

Who else facing this bs

[echo aurora]

Thanks for sharing the image, I wasn't able to reproduce on my end. Are you using iOS?

[echo aurora]

Looks like there are some reports of image upload not working right, this may be a widespread bug. Are you getting the same error when trying to upload just one image vs many?

[lofty frigate]

Well i havent tried one image

[echo aurora]

For clarity I would encourage you to review our Privacy Policy for a full understanding of this - https://help.lmarena.ai/articles/3765052346-privacy-policy. Generations done on the site will not appear in this Discord server, it's only when you use the bot that those generations will populate in the same channel ( #video-arena-1 #video-arena-2 #video-arena-3 )

[echo aurora]

Can you give it a try? Could help explain what's going wrong.

[lofty frigate]

Ok just did and yeah same problem

[echo aurora]

Are you able to share one of the images? You can DM me if you'd prefer that.

[echo aurora]

Happy New Year! Wishing the same for you!

[echo aurora]

Hello - what seems to be the problem?

[echo aurora]

No sorry, the image itself, I'd like to see if I can reproduce the error.

[dapper walrus]

hello can u help me? i cant login when i want do it through google

he make impression like he did but then its still button to login
also when i try to do generation he always saying about error
from any browser

[lofty frigate]

Ok i tried now and it worked

I think i gotta wait a couple of minutes

[dapper walrus]

maybe i overload him by request

[lofty frigate]

Which is konda dumb

[dapper walrus]

or its site glitch

[echo aurora]

he make impression like he did but then its still button to login
Sorry to say this is an issue that's on our radar. Basically, the user attempts to login, they're rerouted to the home page, but still not logged in. Our team is working on a fix to this problem.

[dapper walrus]

thanks

[echo aurora]

Glad to hear it's working. Did you try anything different? Use wifi vs cell data?

[dapper walrus]

^^

[echo aurora]

Looks like you tried different models as here too, wondering if that's playing a part.

[lofty frigate]

Nope

[lofty frigate]

Im doing it everytime with no issue

[echo aurora]

Okay good to know. I'll still try to repro the bug and share with the team if I'm able to.

[lofty frigate]

Ok thr issue came back

😔

[echo aurora]

😭 Is this with gpt-1.5 & seedream-4.5?

[lofty frigate]

All of them

[echo aurora]

I'm going to make a bug post about this so we can keep general chat clear for general chat, I'll tag you there.

[lofty frigate]

Ok

[vagrant garnet]

Hello can anyone pls tell me how to generate video of aspect ratio 9:16 with voice

[fiery gull]

why opus 4.5 soo slow

bruhhhhh

[echo aurora]

Sorry to say you won't be able to specifically set the aspect ratio, for videos with audio/voice that's going to be random. Not all models have audio cababilities, and since it's random which two models you're sampled, that means it'll also be random if you get audio or not.

[echo aurora]

As in slower than normal?

[fiery gull]

return to normal, just normal crash servers time

is from claude, in antigrativy the opus 4.5 was slow

[gleaming roost]

Why is RateLimit saying around 50 seconds, then the time runs out and starts again? What is the actual time? How long should I wait?

[brittle rapids]

Hi everyone! Very happy to join the group. Quesiton for you all. Does anyone know what the best process is to submit your model to LMArena, or to submit for evaluation? We had ours on the backend developer leaderboard for a few days at a time about a year ago - but simply didn't have the infastructure to hold it on there. We actually did really well and quickly climbed the ranks.

Here we are today with adequate funding and a fully working functional product on robust infrastructure and the area to entry seems to have really changed. We're getting crickets on the emails to LMarena.

Silly me...I see the model request below...

[heavy gyro]

@echo aurora Was the Gemini 2k image taken away? I generated a few pics, next thing I know I no longer see it as an option?

[echo aurora]

Hello would you be able to send us an email at [e-mail] ? If you could include more information about the model and lab you're with that'd be great. If you haven't already, I'd also give this blog post a read: https://news.lmarena.ai/ai-evaluations/

I would note that #1372229840131985540 is more-so for the community to let us know what models they'd like to see added. For model providers reaching out directly to [e-mail] would be ideal.

Thank you!

[echo aurora]

I'm still seeing gemini-3-pro-image-preview-2k on the model drop down list. Are you able to hard refresh the page? If it's not there, could you send me a screenshot of this?

[echo aurora]

I'll check in on this bug, but we've recently seen a problem where the rate limit counter is appearing in second rather than minutes.

[heavy gyro]

@echo aurora

[echo aurora]

Ah, you're not in the Image Modality. You'll want to click on this little icon in the chat bar. Then you should see the image models.

More info can be found here: https://help.lmarena.ai/articles/1626988163-lmarena-how-to-use-image-generation.

[heavy gyro]

@echo aurora Ah! Got it. Thank you so much!

[brittle rapids]

Thank you! Sit tight, an email will be coming very shortly!! We are verrrry serious about this.

[echo aurora]

I'm going to reach out in DMs in second.

[vagrant garnet]

Thank u for helping

[echo aurora]

No problem. Keep me updated if you have any troubles or if you have other questions.

[brittle rapids]

Oops, my bad!

[golden ocean]

artificial neural network

[fleet salmon]

Happy New Year everyone.

[echo aurora]

Happy New Year!

[spark python]

Pineapple do you know what word it was to find out how much a llm reads the prompt you gave it, like you sah that word and will return the number of how much it read

[hardy lion]

I don't think anything like that exists. At least on LMArena we pass the entire prompt to the model so it would always be 100%. If the model providers are ignoring it behind the scenes we wouldn't know and wouldn't be able to report that information

[echo aurora]

Yeah it'll be the full prompt. Is there something leading you to believe this isn't the case.

[shrewd citrus]

Is there a way to see the internal thinking on lmarena

cause I can only see this model doing it

[echo aurora]

@woeful shore You'll want to review the inforamtion in #1397655624103493813 for a better understanding of how the Video Area bot works. Let me know if you have any questions.

[spark python]

I saw a tiktok months ago😭 and I remember it that dude used a word , he was using chat gpt and it was a trick to not make it lazy and read longer prompts

[echo aurora]

You should be seeing this on thinking models.

[echo aurora]

Hmm can't say I'm familiar with this, but if you come across the video again I'd be curious to see it.

[shrewd citrus]

I don’t see it, unless if im wrong and these aren’t thinking models

[echo aurora]

Give something like claude-opus-4-5-20251101-thinking-32k a try.

[shrewd citrus]

yep it works perfectly but im surprised that the Gemini or open ai models don’t do the same

[shrewd citrus]

and I would love to see their internal thinking too

[undone saffron]

Not all AIs are the same

[wind shore]

I love the diversity about it too.

[wicked sage]

hi

[topaz flint]

Eror, help me

[wicked sage]

this isnt a error, its just a usage limit that claude has

its a really common problem to see, especially on the opus models

which is really sad considering opus can really make good stuff like this

[silk sierra]

Hey everyone!

[wicked sage]

welcome

<@&1349916362595635286> please nuke this guy k thx

[echo aurora]

Thank you

[echo aurora]

Welcome!

[wicked sage]

no, thank you for nuking the guy

jokes aside, youre welcome!

[wicked sage]

btw ananas since youre here can you fact check this for me?

[echo aurora]

Sorry to hear you're getting this error. This can be caused for various reasons so it's a little hard for me to say with certainty what's causing this. However, most commonly it's going to be rate limit.

You can verify this by: going into Developer Tools, open the Network tab, search for Stream, run a prompt in LMArena that results in an error, open the file that's created (has the same set of #s/letters in the URL) and look for a Status Code, if that Status Code = 429 that means it's rate limit.

If that's not the case, I'd encourage you to follow the steps here: https://help.lmarena.ai/articles/1645798556-lmarena-how-to-something-went-wrong-with-this-response-error-message

We are looking into ways to reduce these errors as much as we can, along with better error messaging to make this more clear when something goes wrong.

[echo aurora]

Yeah it's likely, but not always the case. Difficult to say with the info we have.

[wicked sage]

ok rate limit, close enough

https://tenor.com/view/elephant-green-screen-effect-gif-5821643841100436074

[wicked sage]

i highly think its the usage/rate limit, i know this cuz i was a claude user sometime and the usage/rate limit SUCKED.

so im not surprised if its the rate limit

[quartz light]

@echo aurora I believe opus always times out after around 20k tokens of output

I've ran failed outputs thru a tokenizer and it seems to be around 20k tokens always

[wicked sage]

yeah i think i know why this thing fell after like 5 minut

[remote sun]

hello. i'm so sorry.... but... what is this? I just wanted to register so that my chats wouldn't disappear..

[quartz light]

just try visiting https://lmarena.ai again

[remote sun]

I tried that, but it didn't work. nothing changed when I returned to lmarena

Am

stop

I was just able to authorize via my phone. The account exists, but LMArena is still prompting me to log in. That’s... strange.

thank?

[echo aurora]

Sorry to say we're currently experiencing some issues with the email signup process. Our team is working hard to get it fixed asap.

[remote sun]

This is all well and good, but I guess today is just not my lucky day... For some reason, the generation process has been stuck in an infinite loop. Nothing seems to help: starting a new chat, logging out, performing a hard reset, or clearing cookies — all to no avail. I’ve extracted some logs from the console. Could you please help me with this?

[echo aurora]

I guess today is just not my lucky day
This seems to be the case as that is a different bug.

[fickle venture]

This always happens all i do is refresh website

[echo aurora]

Sometimes that works, but not all the time.

[remote sun]

I’ve refreshed the page many times, but nothing worked. I checked the same chat from my phone, and it’s the exact same situation there.

[fickle venture]

Does this just means model overloaded

Like alot of people using it

[echo aurora]

I don't believe so. It's unclear to me if/why this bug appears.

[remote sun]

Gemini 3 Pro works without any issues in a new chat

[echo aurora]

This can help too, but unfortunately this means the chat no longer has the same context, which can be very annoying.

[remote sun]

I can try to send the available logs

[remote sun]

Yeah... it’s a pretty long chat where I was doing some RP. I enjoy watching how AI tries to please people (well, me specifically) with their requests, acting as if they were human themselves

[echo aurora]

I've flagged the issue to the team, but will keep this in mind if it's helpful for them to debug.

[remote sun]

oh

okay

I hope a bug fix will be able to resolve this issue

[echo aurora]

Big same. Our team is working hard.

[solar hollow]

ai has so much potential to massively help blind people out

it could literally see for them

[fickle venture]

Question is there a context limit in each saved conversation?

[spark python]

You can use Google playground

I've been using it too since it has like 1m token limit

For a chat

[pallid crypt]

Opps message loading lag

Sorry for the ping

[bleak cave]

Hey y'all, I've been testing models' coding abilities in direct chat recently until today when every model I try displays the same "Something went wrong with this response, please try again." message. They seem to be able to half-respond before they shut off. Gemini does work, however it does not seem to be able to edit my code and only tells me what it "did".

[spark python]

I've use 70k tokens today

[fiery gull]

today I use kinda 10mi per day

I'm not even a coder

but my workflow has 200 pythons ;-; Idk to read it, I just hope the opus 4.5 don't will explode my pc

[bleak cave]

Where can I find this metric?

[hollow ivy]

Has anyone tried out this new IQuest Coder model?

According to benchmarks it seems to be a good coding-model.

[fiery gull]

depend of app

I'm using anti grativy now

[spark python]

Python the coding language?

[fiery gull]

yep

[spark python]

And ur not a coder