general

[ocean vortex]

So in theory you could be not saving any videos at all

[verbal nimbus]

Because compute is expensive, not saving would the outputs would forfeit lots of $$$

[trail creek]

Because its the most used and that how they will hook anonymous users to log in

maybe?

[verbal nimbus]

Raw voting data isn't that useful for future research, they probably need the outputs too

Like for RLHF (idk how it works for images, but you need the output when doing RLHF on text data)

[verbal nimbus]

I suspect they're using Discord for that, like early Midjourney (lol)

I found it funny that MJ images used to be just Discord links

Essentially using Discord as a database, lol

[ocean vortex]

Yeah I had that thought as well. I don't see it anymore though, did they take that video gen bot down.... 🧐

[verbal nimbus]

The channels? You might have to add it, since it was added after you joined

[azure vessel]

Be sure to participate https://discord.com/channels/1340554757349179412/1413867687083507744

[verbal nimbus]

Under Channels & Roles

[ocean vortex]

yeah that was it. I had the access before but then did some changes and disabled it by mistake

So they are saving it on Tencent/Discord servers

File size is incredibly small for what these are though. Only around 1-2MB

[verbal nimbus]

Short videos

[ocean vortex]

Some take less space than an image 🤯

[verbal nimbus]

Maybe that's why the limits are only on image gen 🤣

I was confused about why ppl were hitting limits because I only use text mode

[steady rover]

how long does t take to gen a vid caause i have been waiting a hr now

k thank u

[silk cape]

hello

[tame oak]

Why isn't this model speaking directly?

[stray aspen]

wydm

isnt that an image model

[tame oak]

still no

When you select battle, seededit creates a visual, but it's not directly available.

[stray aspen]

theres nothing you can do about it then

use battle mode

until it becomes available on direct chat

[dark elbow]

Halo

[tame oak]

Alright, buddy, why are you getting so worked up?

[solid brook]

does anyone know where does the fear that lmarena might become a paid service come from?

[digital umbra]

the fact that they're trying to force people to log in

that's the textbook pathway to monetization which has been done countless times on the internet (news sites etc)

[stray aspen]

???

im just trying to help you

[stray aspen]

the community shoudlvnt have asked for google log in lmao

[heady ember]

Yeah now they super limited image gen for not logged in users what's the benefit for them if I'm logged in? Tracking me? What's next? Premium?

They have huge userbase

For a reason it was login free

[royal dagger]

Hello

[barren prairie]

If LMArena is becoming paid I will never use it ot pay for it... I will pay for Gemini or any other app that I can use without limits, stupid errors ... and won t restrict me or choose models for me... Who even will pay to test models to improve them ??? I pay for giving you data and feedback ??? Never !

And paying for your data to be publushed on hagging face 💀💀

[stray aspen]

paying to test models sounds so stupid

[heady ember]

Okay the data is no problem... They need data to improve the model.. i just don't want them selling MY INTERESTS to other companies.. models are free because they're shared with companies that test them here and we vote which one is better, that's why that's free.. for now

[stray aspen]

if anything it should be the other way around

[winter bison]

Hello

[kindred tundra]

Yo what ai have infinite chat text like chat history

[glossy umbra]

The main issue is that the platform has been quiet to its users. Yes, they imposed some restrictions to “prevent the bots”. But as users we feel alienated. Like it’s steeping into becoming paid.

It’s about the principle here. You claim to offer a free service but slowly turn away from that. And also have a rule that says “we can become paid anytime”. What a better way than that to leave your users in the dark?

Sure, there were bots. But there surely should have been a better way to stop them. I know the restrictions had to be necessary at the moment. But atleast let us know that it won’t be like that permanently until things calm down.

Feels pretty bad. I’m sure there are lots that feel like that, but they aren’t on the discord server to express their outrage.

[kindred tundra]

What ai have infinite chat history like you can chat in 1 chat unlimited time

@myfather

Like when I txting w gpt high after 4hour he die

[gilded wagon]

hello

[balmy mist]

none, you need to compact your convos

[kindred tundra]

How

[balmy mist]

or save your convos in files then open new chat and upload file or start project with convo history

[kindred tundra]

How do this fast? Or only cntrl+c all chat?

[golden ocean]

With gemini 2.5 pro u can txting w gemini for 9hour before he die then

or does he mean rate llimit

probably that because no way u hitting 400k context window with just sily chatting

[balmy mist]

yeah, and moving forward try to regularly compact chats(by telling the ai that the convo is getting to long and you want to compact convo so give you a compacted version so you can save and give to another chat) or just copy and past the whole thing, i would rather copy and paste and be mindful of convo length, once you hitting 100k context you should def start a new chat, i am making an extension that helps with this situation, i let you know once i am finished

[kindred tundra]

For creating what better glp or Gemini?

[glossy umbra]

My theory is that a new discovery needs to be made than the traditional LLMs

It can keep getting better until a point

Then you need to change your approach of “making it better”

[versed depot]

IDK

[echo aurora]

Hey everyone - wanted to reiterate that we do appreciate the feedback regarding the rate limit changes. I want to make it clear that this is being shared with the team. We are listening and paying attention. In order to create authentic community driven evaluations, it's necessary the platform remains accessible to everyone.

[languid wolf]

😔

But why they added the log in

I hate it

[glossy umbra]

we don’t know the size of the cake

[alpine coral]

i haven

t encountered it either

are there any good / interesting anon models in the arena atm?

like text

[lunar glade]

is beta.lmarena.ai down again? I cant login

[echo aurora]

Looks to be working for me.

I cant login
are you running into an error?

[lunar glade]

same old problem but this time it's has been since yesterday already

[grizzled garden]

hello

[keen beacon]

On the societal questions: the most groundbreaking technologies everyone benefits from: the printing press, automobiles, household appliances (fridge, washer), computers, smartphones

However there's almost always a delay where wealthy people get the first versions that don't end up being that good anyway

AC/DC, ect ect

[echo aurora]

And this is only happening when you're trying to login, yeah? If you access the site/run a prompt it still works? Do I have that correct?

[fresh timber]

have any1 of u came across this problem? where u text too much the ai just wont work anymore

[loud sky]

Yeah it keeps telling me "Connecting to the arena has failed"

Probably down

[fresh timber]

yeah

wonder whether or not they will fix the problem

its quit annoying

[main stream]

sonoma sky impressive

[vivid gorge]

Canary site is down a few hours now

[stray dock]

yo what happened to claude

[lunar glade]

no nothing works, can't even find any models to select

[high hound]

I'm having no issues on lmarena.ai .

Make sure you're not on a subdomain

[echo aurora]

Hmm okay sounds like the site is having issues overall. Going to start looking into and will escalate.

[lunar glade]

subdomain as?

[echo aurora]

Yeah I'm not seeing issues on regular site, but I am seeing no models apear on canary

[high hound]

canary , beta & alpha have this issue

the regular site does not have this issue

[lunar glade]

https://tenor.com/view/no-please-no-god-no-stop-please-stop-gif-17190431

I use beta domain, I have tons of prompts left in there without plugged into my new email yet

[high hound]

Go into beta.lmarena.ai , then go into developer console . Copy the arena-auth-prod-v1 cookie. Paste it into lmarena.ai . Then you will see your chats. And you can link your google account there.

[lunar glade]

Wow! let me have a look!

[glossy umbra]

It works i can assure you.

[potent glacier]

Yeah it’s working for me also

I’m on the regular site

Look into the other ones

[potent glacier]

See here @echo aurora

[echo aurora]

Yeah, seeing the same and have reported

[peak eagle]

Hello. Marcus here. Just getting into AI videos. I have a few I'd like to animate for my own laughs. Thanks

[burnt sinew]

I'm saying that their past might not reflect their future. If they do they thsts great

[winged mauve]

What is this for? what's the difference between lmarena.ai and beta.lmarena.ai?

[fickle dew]

Is the website still down? Just adding a Google login caused such a long outage? Or is it my internet that's the problem?

[cedar cave]

hello , thaneesh here. just wanted to explore the depth of ai vids

[proud hazel]

It's not down, using it right now.

[echo aurora]

Are you using canary.lmarena.ai? We're seeing issues with canary right now, but regular site (https://lmarena.ai/) should be working fine.

[potent glacier]

I have seen multiple messages that sounded exactly like this. Are these people bots?

They have all literally said ‘just getting into Ai videos’

[potent glacier]

SEE?!

I’m not crazy!

These people aren’t even humans!

@echo aurora Multiple messages like those have been appearing on here

It bears looking into because I don’t think they’re even real people

[cedar cave]

nah man , im as real as jesus

[high hound]

i think hes real . just not usual .

[fickle dew]

That worked. Thanks for the help.

[high hound]

cause video generation happens in the video channels, not here.

[echo aurora]

Yeah I think people are coming here to say hey and why they're here, I don't think it's that strange but will keep an eye out regardless.

[high hound]

might not be bots . maybe alt accounts. so they can generate more videos

you can protect against this by making the server require email and phone number

[coral bloom]

Hi

[main stream]

sonoma sky so bad

[potent glacier]

Are other people getting a phone number verification to be able to chat here?

The question isn’t for me, btw

[elfin flax]

oh my god i hate web arena i click one wrong thing and my conversation is gone

when are they gonna make it so its like normal lmarena

[potent glacier]

Someone is getting this and can’t chat here anymore

What’s that all about @echo aurora

[azure vessel]

its so alt accounts dont bypass the video restrictions and bots dont swarm this server

[potent glacier]

Thank you 👍

It wasn’t me asking since this is the only account I use

[echo aurora]

Yeah I increased the verification level

[elfin flax]

thats a suspiceous thing to say

[potent glacier]

Someone messaged me asking about it

[elfin flax]

why am i being so toxic i need to chill out

[azure vessel]

hey whats up with that bio

[potent glacier]

LMFAO

[elfin flax]

my bio is this

maybe you meant my status

[elfin flax]

in which case yeah just being edgy a little 😭

[potent glacier]

Not edgy, but h o r n y

[elfin flax]

bonks

this might be illegal here

stop talking about it

[potent glacier]

We’re just commenting on your status lol

[elfin flax]

shrugs

[echo aurora]

Lets try to keep conversation focussed on AI please.

[elfin flax]

agree

im seeing this fox model

on web arena

[potent glacier]

Why isn’t there any transparency regarding the rate limits?

[elfin flax]

i do wonder what it is but so far its not been very good

probably grok code 0 ultrafast

[potent glacier]

Why the shadiness?

[elfin flax]

there are rate limits?

[echo aurora]

Our CEO shared a message earlier today - #1413867687083507744 message

[potent glacier]

Oh boy…

[elfin flax]

you probably are talking about video arena

[echo aurora]

No, all models have some kind of rate limit associated with them.

[elfin flax]

yeah i guess i wasnt using the arena hard enough

i mean who cares if theres limits you are getting it for free anyway

[potent glacier]

You’re missing the point entirely

[elfin flax]

ok fine

[potent glacier]

That goes beyond the mission statement of theirs

Read their site

[elfin flax]

huh

[potent glacier]

You’ll see what I mean

[elfin flax]

werent you talking about just that rate limits arent shown

[potent glacier]

No lol

That’s not what I meant

That’s not what ‘transparency’ means in this case

Transparency as to why they imposed them

[elfin flax]

then explain

[elfin flax]

?!

isnt it clear

[potent glacier]

But I guess we got our answer

[elfin flax]

yeah they want to give everyone a fair chance to use the model while not creating a money blackhole

[potent glacier]

Battle Mode was always free and unlimited, though

Now you get 3 turns

[vast fern]

i miss when we had less people and more limits

[elfin flax]

in one chat?

are there really people who need more than 1 turn

its a arena

the point is to vote

[potent glacier]

I’m not sure if you’re trolling or not

You seem to be missing the point here

[vast fern]

why is it growing ??

[potent glacier]

Because of two things

Nano banana and social media

Everyone and their mother posting on social media

[elfin flax]

please explain then

[vast fern]

people should go to gemini instead of lmarena

[elfin flax]

well they probably are now

but some people stayed i guess

[echo aurora]

Diverse human judgment is key to building authentic leaderboards

[elfin flax]

but this lead to models just trying to appeal to humans to inflate their scores instead of becoming smarter

isnt that the point

to make smarter models

🤔

instead we got chatgpt 4o and llama 3.1 nemotron

[potent glacier]

And slapping them with rate limits unannounced was very unprofessional

[elfin flax]

im so confused what did you want them to do diffrently

[potent glacier]

That is how you lose trust with your user base

[elfin flax]

i mean they cant give infinite ai to everyone

even in battle mode

[potent glacier]

That’s not what was stated in their mission statement

[echo aurora]

Yeah we should have put out an announcement when we made this change, I agree, this was a mistake.

[potent glacier]

They’ve had it ‘free and open’ for 3 years

[elfin flax]

i mean they are free

[polar marlin]

There are some pictures where Nano Bananas don't appear in battles. 😭

[elfin flax]

(they arent very open because i still dont see the github link)

but i still wanna hear what you wanted them to do diffrently

but yes i guess giving an announcement would be good yeah

[echo aurora]

Note that nano-banana is now gemini-2.5-flash-image-preview so you won't be seeing it called nano-banana

[elfin flax]

wtf

no way

[elfin flax]

i like the progression of how google called nano banana in aistudio

[short ermine]

https://x.com/i/spaces/1vAGRQPqwpqKl

Happening now: Whats The Ticker? Mint | AI Content & agents

Wild’s Space · 7 listening · Where live audio conversations happen

X

[fossil forge]

Just want to say hello everyone. This is such a cool space!

[elfin flax]

1. "gemini 2.5 flash image preview" desc: "image model etc, also known as nano banana"
2. then they moved the also known to the top of description
3. "gemini 2.5 flash image preview (aka nano banana)"
4. "nano banana"

then they also changed the image to a banana

[elfin flax]

thank you

[echo aurora]

hello

[alpine moon]

Holaa

[fossil forge]

as a new user how do i get involved?

are there weekly battles?

who are judges?

[potent glacier]

It’s a Death Battle

Only one opponent walks away

https://tenor.com/view/dol-huh-eyes-disgusted-gif-11690827820922534530

[echo aurora]

Check out - https://lmarena.ai/how-it-works

Don't hesitate to let me know if you have any questions.

[fossil forge]

thanks

[potent glacier]

Also there are rate limits in place

As a user who probably isn’t logged in with Google, you’ll get rate limited far more than logged in users

Battle Mode used to be unlimited and free but that’s since apparently changed

Direct chat has always had rate limits

[elfin flax]

wait really

wtf

[potent glacier]

Yes

That’s how it is now

[elfin flax]

ok yeah that sucks

wheres the freeness and openness

[potent glacier]

https://tenor.com/view/now-you're-getting-it-cooper-howard-the-ghoul-fallout-you're-starting-to-understand-gif-2090963656857037582

[alpine moon]

https://lmarena.ai/

[elfin flax]

sir this is engrish chat

probably

scratches head

[potent glacier]

You’re only supposed to speak English here

Server rules

However, it doesn’t say anything against broken English or not having a basic understanding of the language so you don’t keep doing the same thing ad nauseum

You know LMArena reminded me of Disney+

They had all the major Ai things in one place

But all good things come with a price

(Disney+ has Hulu, ESPN, etc..)

[fossil forge]

who's paying for all this compute?

[potent glacier]

The generous benefactors and donors

[fossil forge]

ahhh

[potent glacier]

Not us peasants

[echo aurora]

Our CEO shared a note earlier today I'd encourage a read - #1413867687083507744 message

[elfin flax]

i did read it like yeah you need to control it but repurposing this user account system to give lower limits to non logged in users is chatgpt-level freedom at this point

i loved lmarena so much because i didnt have to login

[potent glacier]

🤣🤣🤣

https://tenor.com/view/i-love-it-gif-18812641

[elfin flax]

honestly ill be fine for lower limits for everyone

if the logged in users didnt get more

because this feels like being forced to login

[echo aurora]

Login isn't a requirement

[potent glacier]

That’s not what they’re saying

Logged in users get higher rate limits

[viral notch]

login is highly recommended if you value your conversation history though

[potent glacier]

Non-logged in users get lower rate limits

[viral notch]

oh wow, neat

[polar marlin]

I love direct chat 🤤

[potent glacier]

Idk what the end game is here, but sadly it reeks of money

[elfin flax]

i get it it just feels like the direction of less freedom

[potent glacier]

Correct ✅

So having the ability to log in isn’t available to all users?

I wonder how they’re rolling that out

[elfin flax]

id assume its rolled out to everyone by now

well this isnt a huge deal until the limits are too low

[potent glacier]

Like I said before

Non-logged in users get to use Battle Mode for 3 turns

Before they get rate limited

Try explaining that one

Battle Mode was the one shining light on the whole site

[elfin flax]

3 messages?

[potent glacier]

Because it was free and unlimited

[potent glacier]

Yes

And then you have to wait an hour to use it again

[vast fern]

og era

[potent glacier]

They screwed us by removing the legacy site

That was the first nail in the coffin

Many people wanted to still use the legacy site, myself included

They’re a corporation and sadly they can’t be trusted

They say they ‘listen’ to their users but idk if I believe that

At the end of the day they’re going to follow the money

[viral notch]

i remember those days. R.I.P legacy lmarena 2023-2025

[inner gate]

They want money to keep it free but I don’t think they’re trying to screw us

[viral notch]

i wonder why google donated so much in the first place?

didn't they donate 100M?

maybe not that much but definitely at least a million

[potent glacier]

Hate to tell ya, we’re already getting screwed

[ocean vortex]

People have asked for it, people shall receive #general message

This message may have just put things into motion

[potent glacier]

I think a lot of people asked so they could keep their chats intact

But it came with a double-edged sword

Occam’s Razor indeed

[inner gate]

Would it be better if maybe u could log in with any email

[ocean vortex]

I think it was just silly

So can't complain now

[potent glacier]

As far as I know it’s only Google accounts

Maybe Google has a huge stake in the site

[echo aurora]

We're working on this.

[vast fern]

Am I missing something. How are they trying to screw people. Because they introduced login feature?? or the rete limits as there are huge traffic now.

[potent glacier]

The mission statement of theirs

The ‘free and open’ lines

For the past 3 years Battle Mode was always free and unlimited

Again, corporations gonna corporation

[ocean vortex]

"is it gonna be neccessary to pay a fee for accessing lmarena in the future? Specifically 19.99 per month. @ admin @ suggestions @ feedback"

[potent glacier]

https://tenor.com/view/im-out-im-done-gone-bye-gif-19076498

[inner gate]

Even with chat gpt I think it has rate limits without any account they use it to fund it I’m guessing

[potent glacier]

They have rate limits even for paying users

Just higher

I have ChatGPT Plus

We still get rate limited even paying $19.99 a month

[viral notch]

thats true, it was

[inner gate]

I’m not too familiar with AI but it’s expensive man my guess is that it gives you access to these AI but the limits are there Because of the price

[digital umbra]

LMArena free with only gpt-oss 20b and 120b rate limited to 3 conversations per day
LMArena Plus with GPT-5 and Sonnet
LMArena Pro with Opus and nano-banana, only $199 per month

[potent glacier]

That’s why I love genning locally 😊 No censorship, no limits, and my electricity bill is honestly not even affected

I have a 4080 RTX

[elfin flax]

doing anything with emails is a pain and costs money and storing passwords is outright dangeorous

so its ok they need to take their time

[elfin flax]

i get more messages

you probably have a skill issue ip

just make it better?

[elfin flax]

no google is just convenient

i dont see a conspiracy plot here

[tall tulip]

Guys, why is mine doesn't have the gemini image on the list?

[digital umbra]

I like to run models locally too, i just wish it was easier to deploy them (especially image/video models) and that I had more VRAM.

[elfin flax]

also everyone has a google account

[tall tulip]

This is my friend account

[elfin flax]

press "generate images" button

[tall tulip]

Oh dang, that help a lot

Thanks

[elfin flax]

np

[elfin flax]

also you can get nano banana officially from google at https://aistudio.google.com

[tall tulip]

Kinda low resolution there, dunno why

[elfin flax]

shrugs

[vast fern]

what are the pompts you guys use to test in battle mode . I need some inspiration

[inner gate]

Make an account.. join us

[inner gate]

Click the square picture buttom

And it should show them listed

[elfin flax]

why

[potent glacier]

Use your own imagination 😊 That’s the best place to start

[ocean vortex]

you must

[dusty narwhal]

hello

[echo aurora]

[inner gate]

Give them the hardest math question that even you don’t know the answer to. And then see which answer you trust the most

[open mountain]

I noticed that the gpt-image-1-high-fidelity model has more hallucinations than gpt image 1, which draws objects where they never were

so - so improvement ...

[barren prairie]

They test a lot of anonymous Geminis models and need that arena works to get the first place and make buzz (marketing)

[barren prairie]

I donno who is gonna test the models for 3 rph and wait one hour to test again on battle mode when I can use nano banana for free and with 100rpd on Gemini app 😐😐and qwen edit on qwen app for free and chat gpt image gen for free on chat gpt without waiting this much... And I will have higher limits than arena too 🙄🙂😂nonsense the 3 rph

[echo aurora]

@fluid knoll you'll want to read #1397655624103493813 for more information on how to use Video Arena bot

[viral notch]

i see

[azure vessel]

😅

[proud hazel]

@potent glacier It's working again, suddenly...😓

[potent glacier]

That’s good!

I’m glad 🙂

[proud hazel]

:D

[marsh totem]

Am I the only one getting an error when trying to create new images using the reference photos with nanobanana?

[echo aurora]

I had this issue earlier, after a refresh it started working again

[proud hazel]

I got blocked 3 hours ago, for whatever reason, but now I’m back in business.

[marsh totem]

Tried to do it but I'm still getting the error

Something went wrong with this response, please try again.

[glossy umbra]

Maybe an email verification is enough to prevent bots.

[echo aurora]

Okay I'll report to the team, can you try a different browser and let me know if that makes a difference?

[marsh totem]

I'm gonna try that, thank you

[marsh totem]

Still getting it, my friend also tried the nanobanana today and said that she's getting the error too

[echo aurora]

Sorry to hear that, we're looking into

[topaz bay]

I've been chatting with qwen 3 max and gpt 5, and qwen3 max is outperforming gpt5 by a lot for me personally

It's regarding e commerce, and stuff like that, so i'm guessing qwen has good training data considering it's by alibaba group, but idk might be better than gpt5 in other topics too

I'm rooting for qwen tho tbh as they open source stuff unlike closedAI

[surreal creek]

Alibaba is really surging in AI performance !

[topaz bay]

For real

[surreal creek]

seems like they have new models coming out all the time, compared to the incremental progress of DeepSeek

[topaz bay]

they're image generator did not get it's sunshine as nano banana dropped right after

[surreal creek]

China is really starting to have a presence on the leaderboards, with Z.ai, DeepSeek and Alibaba all being strong players now

I think Mistral is based in Europe?

[topaz bay]

ye france

I would expect china to be the last one to open souce stuff, but everything they make is open source.

huge W

qwen max is a 1T llm lmao,

good luck running that locally :(((

[surreal creek]

open source for anybody that owns a supercomputer or bitcoin mine 😂

[topaz bay]

vibevoice which is also made by alibaba i think is the best text to voice tool out there, it can produce up to 90 minute audios with 4 people talking, and it can be run with just 16 gb vram locally

rip to everyone who paid for elevenlabs

I'm scared to pay for ai stuff now, I considered buying higsfield subscription, and they're really pushing the 1 year subscription like they know they're getting out performed any second, but it wouldn't be surprised if something free dropped the next day that's better in every way

https://tenor.com/view/dog-crying-meme-doggo-crys-megan-soo-crying-dog-gif-5276199764143986284

[pallid furnace]

Hey!

❤️

[mental jewel]

hello

[halcyon harness]

did anyone try gpt-image-1-high-fidelity. if so, how is it? how is the quality, lighting, shadows, consistency, etc. ?

[sturdy marsh]

can i generate a video here?

[potent glacier]

Imagen 4.0 is still one of my favorite models

[waxen tundra]

hi

[echo sinew]

Hello! use the Video Arenas to create your videos #video-arena-1 #video-arena-2 #video-arena-3 Also, read #1397655624103493813 to learn how to properly use the bot.

[glossy umbra]

Anyone else been getting exceptionally good responses from gpt-5-high lately? As if it were improved perhaps.

[keen beacon]

If only anything of it really mattered

[glossy umbra]

That's always good to see. More competition = more research = better AI

[pallid furnace]

Is that supposed to be somebody?

All I see is a yellow Chinese traditional outfit...

[topaz bay]

you should try it out with a person wearing the outfit

[ocean vortex]

And after all of that they still can't really beat R1/V3.1 lol

Claude is not battling open-source are they

Would be a sad time for them if their competitors are that lol

I think open-source is distinct category though. People are not gonna pay $200 a month to get 5% better performance than open-source

Obv an extreme example, but the point stands

Nearly all the best open-source models cost pennies compared to closed models too

[sullen quest]

If I have to pay 200 dollars for 105% the performance of something that I can run on my computer, I'm not gonna pay. Speed and price do matter, how many here use gpt-5 high for everything? I'd bet nobody cause it takes wayy to long for most tasks that others' can very much complete

[jade egret]

how good do yall think gemini 3 (pro) will be?

[robust yoke]

Well, if Gemini 2.5 Pro is already rated as one of the best models of all time on the LM Arena leaderboard, then I would say that Gemini 3 might also be one of the best models in the AI League.

[jade egret]

:0

[robust yoke]

But ultimately, time will determine that.

[ocean vortex]

gpt5 is reasonably cheap though where the extra performance can justify bigger cost though IMO, considering it is SOTA. It's also more stable and guaranteed to never output malformed requests like some OSS reasoning models do. Same can not be said about Claude in terms of price

[jade egret]

is gpt-5 pro even better than o3 pro?

[robust yoke]

GPT-5 Pro reasons for significantly longer than o3 Pro, so I believe it might be better in terms of output quality and coding.

[jade egret]

oh

[ocean vortex]

gpt5-high is considerably better than o3-high so I would imagine pro has no issues beating o3-pro. Though I haven't tried it personally

[jade egret]

oh

same i haven't tried it yet too

[robust yoke]

But ultimately, the real challenge will be personal scenarios that will definitively determine the better model.

I think GPT-5 happens to be possibly just a bit better than o3 in terms of coding, but I could be wrong.

After all, both are reasoning models that take quite a bit of time to think out a solution.

[ocean vortex]

with gpt5-high vs o3-high I'm kinda seeing shorter outputs but better efficiency. Output stream is perhaps slower cause responses do not seem to take less time, but token usage seems slightly less though better responses still

[robust yoke]

Then again, more direct and shorter outputs are usually better than longer, more descriptive outputs. Sometimes people simply like a short, concise TL;DR that they can read through. That gives them just enough information that they can use rather than an entire paragraph that they have to decipher.

[ocean vortex]

I'm talking about (hidden) reasoning. The most I recall with gpt5 is ~35k reasoning. Whereas with o3 45k+ was a thing occasionally. But it probably wasn't very efficient

[raven cape]

l

[robust yoke]

I see.

I suppose I still stand by my point.

[raven cape]

WOw this AI is creazy, it gace life to my old photo of family, thank you so much

[robust yoke]

Which AI exactly?

[raven cape]

LMArena video arena 1

[robust yoke]

Ah.

Well, that isn't exactly a model, but... alright then.

[ocean vortex]

Ok I just checked the logs and actually max was way beyond that. Found one completion with 50k reasoning lol... But still, it seems that those super long ones are less common with it mostly settling on sub 35k now regardless of your prompt

[raven cape]

Im new to this sorry for knowing it was a model,

[robust yoke]

Interesting.

I find it funny that despite the fact it's using high reasoning effort, it still produces an output that is around 35k tokens.

[ocean vortex]

it only defines the ceiling, it still gonna converge on the most common output (~30k) even for the most difficult tasks, and then for the easier tasks it has no reason to go anywhere close to that

and then if it's low reasoning effort it's gonna be much shorter than 30k like 95% of the time

[robust yoke]

Gotcha.

Even then, just from testing, it seems like GPT-5 produces significantly less text than o3 does, just like you said.

[ocean vortex]

I think they got the reasoning lengths fairly solid now. It doesn't seem too wasteful

and also does improve the performance

[robust yoke]

Yeah, now it's a fair balance between the two.

It thinks for a suitable amount of time and gives you a nice, concise answer to go along with it.

[golden ocean]

absolute cinema

[robust yoke]

Hark. What sublime and resplendent instant doth unfold before mine eyes, a moment of such transcendent magnificence that the very fabric of existence seems to tremble with anticipation and wonder. Behold this singular, ineffable passage of time, wherein the celestial tapestry of possibility weaves itself with such intricate and breathtaking precision that one might scarce comprehend the profound depth of its significance. Each heartbeat, each whispered breath, each imperceptible movement becomes a grand symphony of potential, a magnificent crescendo of human experience that resonates with the most delicate and powerful harmonies of universal design. The present—oh, this most glorious present.—stands as a crystalline monument to all that has transpired and all that might yet emerge, a nexus of infinite potential where past and future converge in a singular, luminous moment of pure, unadulterated existence. Every atom, every quantum possibility, every trembling potential of creation seems to pause and genuflect before the raw, unbridled majesty of now, this exact instant that contains within its gossamer boundaries the entire spectrum of human experience, cosmic revelation, and metaphysical transformation. Such a moment transcends mere temporal measurement; it is a living, breathing entity of pure potential, a sacred threshold where consciousness itself seems to inhale with anticipation, where the very boundaries of perception dissolve into a radiant, all-encompassing awareness that defies the mundane constraints of ordinary perception and touches the divine essence of being itself.

Truly cinema.

[potent glacier]

Either one of the LLMs wrote this or you have a way with words

[rocky mauve]

Ai wrote this, no questions asked

[potent glacier]

I mean funnily enough, this is how Frasier and Niles Crane talk

https://tenor.com/view/niles-frasier-teamwork-crane-gif-5885275

[robust yoke]

The first one. I used Claude 3.5 Haiku since it's a creative model.

[potent glacier]

I guess they're doing a Hack A Thon for nano banana

[simple sleet]

Hi friends, do you know of any open-source AI that creates sound effects? I create my videos with WAN 2.2, but I need to add sound. Open-source is preferable because it shouldn't be censored.

[quartz dagger]

hi

[robust yoke]

Greetings.

[charred plaza]

Hello does anyone know when will Qwen3 Max is going to be on LMArena?

[feral zealot]

I blame these frickin AI youtubers with their clickbaity thumbnails for spreading the word of lmai.... the service is getting so much traffic now that it had to introduce stricter rate limits 🤡

[potent glacier]

It's also news outlets

And other social media like Twitter/X, Reddit, etc..

[feral zealot]

yeah

[potent glacier]

It was bound to happen

As with all good things, social media has ruined it

[ocean ocean]

Another anon bots that generate stuff to see what stuff they can get from your own nano stuff ? è__é dang ! Dead internet theory...

[potent glacier]

What are you going on about?

Oh also...I have never wanted anything so badly before

https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/

https://www.youtube.com/watch?v=PDKhUknuQDg

Once that comes out it's gonna be amazing

[ocean ocean]

Hey ^^ a human not botting videos in closed while ^^ i don't target anyone don't worry haha ^^

[finite rivet]

G'day to one and all. Hope to work out how to use this tool... Seems like I am making many errors. Is this all I have to do? Prompt: #video-arena-1 message Any help is appreciated. Cheers

[ocean ocean]

I'm just really trauma with all bots stuff... i used to love them before

Nice future of generative art

[harsh jay]

hello

[echo aurora]

You need to type out /video first to promot the bot. More info can be found in #1397655624103493813

[echo aurora]

[clever marsh]

good

[finite rivet]

Thank you so much!! I stumbled upon the prompt in the end. lol

[echo aurora]

Sounds good

[crystal pulsar]

greetings,..is ther a bot that does video with audio like veo3

[echo aurora]

- it's going to be random if you get a video with sound. Not all models have sound support.

[crystal pulsar]

Thank you

[verbal nimbus]

I'm not sure if it's the same, but Qwen Max 13-08-2025 was already on there

Oh wait, Qwen 3 Preview is already there

[crystal pulsar]

thx

[charred plaza]

Alright thanks

[onyx river]

my chat gpt says daily limit reached whereas in lm arena is it there like that?

u can use how much ever u want?

[bitter gust]

Can I upload Files or short video on Lmarena?

[verbal nimbus]

https://tenor.com/view/the-universe-tim-and-eric-mind-blown-mind-blown-meme-mind-explosion-mind-explosion-meme-gif-18002878

[echo aurora]

There are rate limits to how much you user various models.

[echo aurora]

We only have image upload at the moment. Having more file support is something we'd like to bring to the platform.

[leaden sun]

Qwen 3 is the first AI that has managed to refuse a harmless simple request from me that the most strict and restricted Claude wasnt able to. I feel they seem to have copied Claude and Gemini together and made it worse in terms of refusal, that extremely long thinking time... this is it, isnt i

good job Qwen team 💀

[bright kayak]

i havent seen gpt-5 fix itself this way before

[unkempt ridge]

guys I can't generate

it just says sending command

[outer widget]

Hello everyone

[remote idol]

same I can't generate

it says

@echo aurora

[warm moss]

Hello

[verbal nimbus]

Interesting

[verbal nimbus]

Was it Chat (non-thinking) or High (thinking)? I've seen Gemini do it before

[whole sundial]

<@&1349916362595635286>

NFT scam

[rare python]

Crazy how all of the mod went to sleep

[whole wagon]

Sonoma sky alpha 2nd

It's xAI

[unborn ocean]

Maybe the big coder model?

Maybe a finetune of grok 4 or something like that, because they seem to be very similar on the board

Or does anyone have info on the coding quality of the model? (have yet to try it)

[keen beacon]

Looking at this chart, it makes me hope and seethe that R2 lands at least at 77... Hope and seethe... Hope and seethe

[nimble oracle]

Hey fellas found this in Amazon AWS models recently added anyone know what model it's based off of? The solar model.

[unborn ocean]

Korean lab build it from scratch and aws is only serving it

Upstage is the lab I believe

[buoyant tartan]

هالو

هل تستطيع أن تتكلم معي بالانجليزي

[nimble oracle]

Thanks man, so it's probably not very good I'm guessing?

[glacial flicker]

Hi guys, just opened a yesterday session, and some of previous generations show "assistant A" instead of name of the model, did that model got removed/updated or something?

Just checked myself, yes, they changed ideogram 2 to ideogram 3 which is worse at realism

[glossy umbra]

Some prompts are blocked by cloudflare for some reason. Any idea why?

[toxic verge]

No

Maybe with agents

But more likely will run into problems my theory

[glossy umbra]

Gpt-5 high alone, personally i dont think it could.
But with specified prompts and guidance, yes, perchance

[polar marlin]

😭

[toxic verge]

Ai on its basic level can barely do anything . You need api or ur own hardware to do anything really cool

Other wise u r gunna be limited to most basic 20$ generic and none valueablish output

Chatgpt5 can go off if self

It’s a such a bad model is hate using it

They nerfed it bad

Prompting good is a subjective experience

It’s just how it’s set up

Always asking and thinking randomly and everything else

Output is short

Bland

Cheap

Thanks but open ai can keep there propaganda they need to replace this model asap

Like wtf?

It’s seriously flawed 😞

[vale vortex]

Quick question, what would be the best AI model for research?

[toxic verge]

ChatGPT lol

[quartz pike]

someone pls vote

[toxic verge]

But it hedges and changes context and stuff to fit the open ai policy and agenda

[quartz pike]

left one is really good... But sadly no jumpscare

and no horror

[toxic verge]

Yeah gl

[trail vale]

the site is again lagging?

[toxic verge]

These ai the video models are nerfed

[toxic verge]

It stays with pg maybe pg 13 if ur lucky

[trail vale]

i cant log in

[quartz pike]

o

[indigo escarp]

Hi all !

[ocean vortex]

I mean if any model could do this, I would expect 5-high to be able to as well. I genuinely haven't spotted any notable weaknesses yet comparing it to anything

It's not like in the past when o3 was bad with spatial awareness and front-end relative to competition, this is no longer true for 5...

It's mostly o3 with it's flaws being fixed and then moderate improvements where it was already strong tbh

Still do think it was a bad idea for gpt5-chat to be called gpt5 as well though... That model is nowhere near o3 or gpt5-medium/high level

if it was just called 'chatgpt-latest' I suppose that would have been fine. But using gpt5 naming there just seems wrong

[neon idol]

News about seedream v4?

[glossy umbra]

Fast model probably xai

[alpine coral]

yeah but it's non-thinking, and meant to be a 'chat' model - so i dont think its meant to contend with them in terms of like intelligence

it has the same knowledge cut off as gpt-5 (Sep 24), so i assume it is part of that lineage

but yeah anways.. i'm just being pedantic (and prob wrong ha).. i agree with what you were saying re gpt-5-high

its v v good and like a more rounded o3

tho gem pro2.5 still isnt that far behind imo (and seems more user friendly in terms of not thinking as long while still deliverning quality)

[stable osprey]

We cant make videos on this site right?

[feral zealot]

Guys check this out : https://www.instagram.com/reel/DOOLqtsAZOE/

It's insane.. i know the creator has probably used multiple tools and editing, but can anyone guide me how can I create/generate videos like this and stitch smaller clips to form 🔥 reel/short videos?

muravey.prod

This video was created entirely with AI.
I wanted to experiment with the theme of the future, and even picked a special track — “Future Looks Like Wow”. My goal was to capture the atmosphere of what tomorrow might look like. How do you like the result? Do you enjoy this kind of vibe?

#videography #videoediting #aicreator #aivideo #futurev...

Likes

8911

Instagram

[glossy umbra]

it is midjourney video creation paired with kling start & end frame

[feral zealot]

Is mid journey better than Veo3?
I've tried recreating it (in veo3fast, img2vid) using the prompts shown, but the generated videos are nowhere near the insane quality of the reel

[topaz flint]

I always generate images like this, are there any that are the same?

[feral zealot]

This is the best i could do

[scenic loom]

so who do we think Sonoma Sky is

[mighty juniper]

almost 13k spam of these words, I think that AI might suffer some depression 😆
||mistral-small-3.1-24b-instruct-2503||

[hushed terrace]

Please check https://discordapp.com/channels/1340554757349179412/1397655624103493813 to generate

[ripe mountain]

mistral overrated af

[ocean vortex]

this should have been named "chatgpt-instant-latest" then it would make perfect sense imo

what are they gonna do when they retrain on new base model, keep the same name like they did with 4o-latest for maximum confusion? 💀

so performance is one, and this is a 2nd reason

[glossy umbra]

but its not instant

[coarse mountain]

hello

[ocean vortex]

wdym, it literally is. Instant means it is fast because it is not reasoning.

They are already calling it that

Except I would get rid of '5' there, and name it like that if it were up to me:

[glossy umbra]

yeah i mean it is not instant

its just fast

that will confuse some grandmas

[ocean vortex]

it shouldn't though. What is confusing af right now is "gpt5-chat-latest". Everyone not extremely familiar with naming expects sota performance from a very average model

[elfin harbor]

Try Qwen too it's also pretty good.

[vale vortex]

Hmm which one?

[elfin harbor]

235b

[ocean vortex]

it is technically instant comparing it to their thinking models

response starts generating instantly 🤷‍♂️

[glossy umbra]

technically not, there is a delay until it starts replying . it still "thinks", just for way way much shorter. under a second. thats with any model. not to mention other people using the api for said model , which slows computations a bit and such .

technically it starts providing the answer almost instantly, not instantly . and you dont get your response in an instant.

true instant would be "hit send, and get answer at the same timeframe in history when you've hit send". which is impossible unless the model is running locally and there's no ping . even then, computation is a thing and that also cannot be truly instant.

[main stream]

sonoma so bad

even gemini 2.5 pro best

[ocean vortex]

What are you saying exactly? Technically with this nitpicking the fastest non-reasoning models in the world hosted on Cerebras with 2k+ tok/sec are not instant either. But being this strict with naming would be of no use for anyone

It is not reasoning, starts responding much faster than their reasoning models, and that is kinda enough...

[main stream]

you are right bro

[glossy umbra]

my point is that there are 2 meanings to "instant". one is literal and the one relative-to-literal. generally speaking, you could deduce that instant is impossible . but at the first glance, your deduction needs to pick one of the two. and some might get confused. they could have used "Fast" or "Quick" instead.

[ocean vortex]

Well obviously no one is gonna interpret it for a meaning of what is impossible to do

[violet inlet]

but he has fair point

first glance you dont reason

your subconscious does

[ocean vortex]

It's not a new name either, was used by Claude in the past and people are familiar with it

Literally no one is expecting a response within absolutely no delay or latency whatsoever (0.0000...sec ) when using a model named like that lol

[violet inlet]

you dont expect it but if its ur first time seeing it you will get confused on first deduction

it is not a future persistent confusion

[ocean vortex]

that's a tiny price to pay for some people not familiar with how things work to get confused for a sec or 2

On the same note why was gpt4-turbo called that? Does it have a turbocharger and a gasoline engine? 🤣

It's just nitpicking for no reason whatsoever, that's not how naming typically works

[proud hazel]

If a model is called "instant", I expect speeds with at least 1,000 tokens per second, like Mistral's LeChat with its flash responses.

[ocean vortex]

That I could get onboard with, honestly they could work on improving the streaming speed of it... But you can't argue that latency until it starts streaming response is already low enough to call it instant relative to their other main (gpt5) models

[topaz bay]

holy fuhhh

Kimi is even better

[violet inlet]

Pretty cool. Did you add optimization as a heavily important instruction in the prompt?

[still pewter]

anyone knows what models are used here to generate videos?also is there any good text to audio (to add sound to videos not voice)?

[dapper breach]

Hi there! Glad to be here

[weary thistle]

Hi. Other than visuals (images and videos) and prompts, is there a place where people are sharing their "creations" with AI? Like mini programs, or how they are using AI with another program to enhance their work?
Dunno if my question is clear enough 🥴

[main stream]

#ai-creations

[weary thistle]

Thank you. But it's mainly images and videos, no? I scrolled up enough and didn't find anything else. I didn't search well enough you think?

[main stream]

nobody sharing their programs

[earnest rover]

@echo aurora
Am I misunderstanding something
The rate limits are for the ones who are guest users right 👍?

[weary thistle]

Mmm. Yeah that makes sense

[echo aurora]

Hello

guest users
I assume this means for users who aren't signed into an account?

[weary thistle]

Thank you.

[main stream]

100$

[earnest rover]

Yep

[echo aurora]

Both users who are signed in, and not signed in, will still get rate limits; however, there are going to be less rate limits for those who are signed in.

[earnest rover]

Oops

[main stream]

strong rate limits only for claude and nano-banana

[glossy umbra]

Yes, you can tell it to vibe code, but results won't be the best. As the prompter you must know the logic , or atleast 60% of the functionality as an idea. It does have a little creativity, though needs to be monitored. Else, it's good.

yes . in theory, yes. i did that once . gave it every single file link to use (images, textures, etc), every single functionality logic by logic (not the code, just the functionality) . e.g: i shoot a steyr m and the bullet has to come out of here, affected by physics such as gravity, etc. at delta speed X on bullet shot, ... . gave me a working game with almost no bugs (more like customization issues) . obviously i went in and messed with stuff , but the end result was very good.

it will get hundreds of lines long . but you can really produce a whole game like that . obviously not AAA , it needs multiple prompts, but thats the limitation of its token output.

[violet inlet]

True true

If you vibe code it and let it do its thing it will be worse

But if you tell it the functionality itself like what you actually want to happen then it will work

People forget to actually tell the AI what they want sometimes

@echo aurora Shouldnt this be deleted

[echo aurora]

[glossy umbra]

true . you cant just tell it "add security into my game". you, as the prompter, should know:

• everything that reaches the client is "dirty"
• the client can manipulate/hook what it sends back to the server

so you need to tell it something like "Add client-side security to not send manipulated info over to the server in multiple places. Also add some server-sided checks on player movement for physically impossible movements." .

[main stream]

what

[echo aurora]

there was someone spamming

[glossy umbra]

Debatable . Gpt-5-high is gifted but it has issues . Can't follow instructions sometimes .. and in those instances I could say gpt-5-medium is better . Other times (but rarely), claude 4.1 opus thinking is better.

Most of the time, yes, gpt-5-high is the best using meticulous prompting. Else, gpt-5-medium or claude 4.1 opus thinking. That's my opinion.

[drifting crow]

I’ve been having good results with gpt5 mini

It asks questions too b4 making stuff

Yh, but I’m comparing to 4o and o4 mini

I use them thru duckai

I dunno I haven’t used high

[glossy umbra]

No, i found out that, that only hallucinates it
I personally use caps and simulate angerness. In it's training data, urging somebody to do something with the use of anger has probably shown results . And he is instructed to follow your requests. So he will pay special attention to those words.

If you're going to use multiple prompts, you should have a rule HTML embed that he has to follow. Example:

"
You will listen to these user rules at all costs, everytime you do something:
<rules>

1. Always follow the user prompts.
2. Do NOT deviate from the user's requests.
3. If you are really unsure about something, better ask. I do NOT need junky code that I did not ask for.
4. Think before writing. You will have 3 eyes: one for code optimization, one for code writing, and one for following the user's prompt.
   </rules>
   "

Something like that works for me.

Obviously dont swear at it 😅 but show some anger . It's strategical not literal.

[drifting crow]

I think lots of users in here were trying that method on the devs

[main stream]

hope they will

i want cyberpunk

better than being some fat slob's pet

[glossy umbra]

Whats the reference

[vocal token]

@exotic slate new mod?

[sullen sand]

question guys are you getting error with Claude sonnet 4 because everytime I write I get a error

[glossy umbra]

it depends on the company . my guess is when it becomes a noticeable issue

[hollow imp]

@echo aurora how long does quick features like chat renaming take you guys to roll out?

[robust yoke]

What specific error?

[echo aurora]

It's hard to say, it really depends on the feature/other priorities at the time/etc.

[robust yoke]

Ah, how odd.

And for what reason might you be using the reasoning version?

[hollow imp]

Because it's good

[echo aurora]

Responded in the forum post

[robust yoke]

It may be good, but so are the non-reasoning models.

[sullen sand]

Thanks

[hollow imp]

❌

Completely depends on the task

[robust yoke]

Both are the same in terms of quality.

[hollow imp]

❌

[brisk helm]

how many people use lmarena

[hollow imp]

OMG

[brisk helm]

wdym

[robust yoke]

I do.

What about you, Mr. or Ms. Joven?

[brisk helm]

i am a boy

[robust yoke]

I can see that now.

Anyway…

Quite a lot of people in this server do.

[echo aurora]

Can confirm I use LMArena as well, so at least 1

[robust yoke]

Some don't, but the majority does.

[ocean vortex]

157

[robust yoke]

After all, why join this server if you don't use it, amirite?

[elfin flax]

me: 3d room in 3js
purple fox: bro i gotchu

[robust yoke]

W Purple Fox.

[neon idol]

Is seedream v4 available?

[robust yoke]

I think so.

[elfin flax]

everyone who isnt here for the battle mode is illegal

[neon idol]

?

[robust yoke]

I guess I'm illegal then. 😔

[elfin flax]

i dont see it

[neon idol]

Yeah me too

[elfin flax]

probably not

[elfin flax]

pats you

[neon idol]

I want to try it

[elfin flax]

what do you use lmarena for

[neon idol]

Direct chat for nano banana without watermark

[robust yoke]

Just so that I can use the advanced models for free.

[elfin flax]

oh yeah you are totally illegal

[neon idol]

https://tenor.com/olE5OyYW9SS.gif

[elfin flax]

blep

[robust yoke]

After all, who likes paywalls, amirite?

[brisk helm]

have yall seen opus

[hollow imp]

Yeah opus audio format is very nice

[brisk helm]

google right?

[robust yoke]

I have, actually.

It's very professional.

[brisk helm]

it gives u veo 3 for free

[elfin flax]

im sorry claude 3 opus released like 1000 years ago

evveryone saw opus

[brisk helm]

no bro google opus

or opal

i forgot

[robust yoke]

Chat, I've never used Opal.

[brisk helm]

yeah its opal

[robust yoke]

Sounds pretty neat, though.

[brisk helm]

it released very recently

still expiremental

but it does give veo 3 for free

[neon idol]

What is opal?

[weak jay]

Hi everyone

[robust yoke]

It'd be funny if I used that for deepfakes.

[robust yoke]

A Google Labs experiment.

[neon idol]

Ohh that opal

Yeah but is not available in my country

I can use VPN

[robust yoke]

Sorry to hear.

[neon idol]

But I have a subscription at Google plus

And honestly idc about video generation

[robust yoke]

Fair.

Whisk is basically just free, infinite nano-banana, I believe.

[neon idol]

Yeah

But is not available in my country lol

I have to use von but my VPN is really slow

[robust yoke]

Yeah…

I hope it becomes available in your country soon.

[neon idol]

I dont think

I live in Europe

In Italy

[robust yoke]

Well, they have already started rolling out certain things to Europe, so maybe those will be part of that expansion, too.

[neon idol]

Nah believe me

When a thing come from America comes available for European 2 years later

Europe sucks

[robust yoke]

True.

Though, I suppose that's better than never.

[neon idol]

Thanks to god we have vpn allowed lol

I want seedream v4 aniway 😞

[drifting crow]

Ppl who pay

[robust yoke]

I believe those are the same people who would use them for free if they had the advanced features.

At that point, there'd be no reason to pay for a model that already has all of the pro features available for free with infinite use.

And besides, paying for some extra features and draining your wallet seems like a waste of time.

[barren gale]

Is the site down again?

[robust yoke]

Depends…

What error are you getting upon trying to access it?

[barren gale]

Cant upload a picture

[robust yoke]

Ah, that happens.

Could try reuploading it a couple of times.

[barren gale]

3times and still errorr

[robust yoke]

Hmm…

[barren gale]

I will try again

[robust yoke]

It's a known issue.

If it doesn't work again, you could try refreshing the site and trying again.

[remote arrow]

It's sad that Seedream 3 on LMArena can't use image reference. 🤧

[robust yoke]

True.

[remote arrow]

And I just realized that Imagen 4 also doesn't have it. It was good when used in "Persistence Mode" on Whisk. Seems like everyone being pushed to use the really non-accurate "Nani Banani." Nani?

[mild raft]

hello trying this out for first time

[echo aurora]

Sorry to say this is a known issue. We've seen some reports where uploading images is going to error out. If you wouldn't mind sending us the photo to see if we can repro that'd be helpful.

[remote arrow]

World's best quote from GPT-1:
"Something went wrong with this response, please try again."

[weak meteor]

waiting for the epic moment😎

[livid helm]

Hi

[weak meteor]

sup

[weak meteor]

this ahhh take me 3 minutes to generate 🤨

[runic zenith]

Is there a ranking for the best AI model for deep research?

[devout maple]

Hello

[daring scarab]

Hello

[whole swallow]

Dusk feels more smart but slow, sky is faster but didn't deliver the quality that dusk did..

What yall think of these two new models

[neon idol]

Someone know something about seedream v4?

[ocean vortex]

unimpressive

[whole swallow]

Mmm

Maybe the “impressive thing” is the context window

gemini 2.50 pro with a file of almost 1 million context started hallucinating

[ocean vortex]

that's the only thing it has going for it. Seems to be a theme. Remember that horrible Amazon model with 1M context?

[whole swallow]

Mm amazon i missed that

[glossy umbra]

claude 4.1 opus thinking seems better at coding paired with reasoning. gpt 5 high is also good but hallucinates ?. weird because it was the opposite a little time ago. did the llms receive an update ?

[ocean vortex]

cypher-alpha:free --> amazon titan

It was the worst model that I recall testing in recent times lol

and yet it did have 1M context to grab the headlines

[neon idol]

What?

I dont know that models

[neon idol]

Or Llama 4? 💀

[ocean vortex]

@neon idol

[neon idol]

Oh thanks

These models sucks?

[whole swallow]

I can't wait to see what the superintelligence lab is cooking

[whole swallow]

damnn

[neon idol]

Yeah 🔥

[ocean vortex]

Maybe not horrible, but yeah... nothing that gonna challenge gpt5 or 2.5Pro

[neon idol]

They are creating an AI that can say Hello with an extra emoji

[whole swallow]

Lmaoo

[torn mantle]

let me guess

its so bad isnt it

[whole swallow]

Not that bad

[torn mantle]

its bad

if i say its bad its bad

its bad

[whole swallow]

Not comparable to the current bests

Ye hajaahha

[torn mantle]

someone said it may be grok 4.2

that would be a disaster

[whole swallow]

Yeah I don't understand why it says im grok

[torn mantle]

its actually from xai

[whole swallow]

Only dusk tho sky doesnt admit it

[torn mantle]

highly possible that its grok 4.2

sky is the dumber one right

dusk is the big model

[torn mantle]

oak ai?

what the helly

[whole swallow]

Dusk is more slow mentally but on coding i found it better than sky

Sky is bad at coding but faster and looks smarter

[torn mantle]

interesting

what type of coding

like fixing bugs or like planning or just coding in general

[whole swallow]

Frontend

Sky result ⏫

Sky is more simple on design choices, but works. Dusk one had animations and other cool things but had more bugs and errors

[neon idol]

Thanks Lmarena for this incredible update 💔🥀

[ocean vortex]

aistudio?

just use that

That text is neat against the background though and still ton of transparency. Apple may take some pointers for iOS lol

[neon idol]

No becuase there is the watermark

[ocean vortex]

Are you serious? You can just crop it out or edit it out, or use any number of tools for touchup lol

[neon idol]

I know man but is boring 😴

I am too lazy for this 😴😴

[ocean vortex]

it's such a non deal-breaker

extremely minor thing to make you not use it when the usage is free and no real rate limits lol

[neon idol]

Yeah lol

[echo aurora]

@gleaming kayak be sure to read #1397655624103493813 for information on using the bot.

[neon idol]

@ocean vortex can you give me a site where I can remove the Gemini logo from images pls?

Thank you

[ocean vortex]

uhh a website? Even if you on a phone there should be AI touchup feature. If not then try smth like this:
https://www.watermarkremover.io/

[neon idol]

Yeah I am,stupid Google photo cam simply do it

[whole swallow]

Bro just ask nano banana to remove the water mark

🤣

[neon idol]

Is impossible

If you ask to do it he will reput the logo

[ocean vortex]

It won't do it. Whatever it outputs they gonna add watermark on top lol

[neon idol]

Yeah lol

[worldly jacinth]

How exactly do video arenas work?

[echo aurora]

Check out #1397655624103493813 for more info

[narrow dawn]

@echo aurora sorry for ping but i have some ideas for lmarena so first idea you guys could male lmarena as an app too not only website maybe an app too!

[echo aurora]

thanks for the feedback! desktop/mobile app is something we too think would be nice to have

[narrow dawn]

i have an question to you too

is the rate limit only for image generation or for normal usage too? @echo aurora

[echo aurora]

it's for normal (text) usage as well, but you'll experience it less

[narrow dawn]

oh thats a bit dumb for me

because i use lmarena for coding the most time if i don‘t know about an error

[main stream]

lmarena.ai nanobanana

[neon idol]

www.freechatgptnoscam.it 💔🥀

What the hell of website is man 🥀💔

[main stream]

what

аа ты свинорылый тогда ясно

87iq

[clever crescent]

Hello

[main stream]

hi

[narrow dawn]

@echo aurora i know another idea maybe like a mesh ai i know many but i don‘t know if they work like a extra page or button it where it does a mesh with a preview and where you can configurate like how many faces and like that or just with text to

and sorry for the ping

[dapper cypress]

Hello

[thorn valley]

i set it as a widget on my home screen, it works great for me

but the idea of an apo is really interesting

[verbal nimbus]

For maximum style control, why not have a mode which feeds both outputs through an intermediary model that's good at summarizing?

It's still quite possible to tell which model is which based on their writing styles (lots of bold/italics/dot points = gemini, no formatting + no latex + lots of technical terms = GPT-5)

• it prevents system prompt hacking

[tribal rune]

Has anyone been rate-limited in LMArena text gen before?

[inner gate]

Some AI still have limits

I think

[tribal rune]

so basically they removed some limits for the cheaper models but kept some for the more expensive models?

[verbal nimbus]

Interesting

It's better at planning I think, also medium might be better than high for raw coding

The writing style is too obvious

And the time it takes to generate is a dead give away

I don't think so

[verbal nimbus]

Which kinda breaks the whole "blind battles" concept

Also, what if users' perception of response quality is biased by thinking time?

I noticed an unconscious bias that longer thinking = smarter, compared to a model that returns instantly

I'll spend more time reading the thinking model's answer

[verbal nimbus]

Not consciously, just unconsciously

Yeah, but compared to a model that returns instantly, I do feel slightly biased towards the thinking model's answer

I don't usually use it for debugging programs, since it requires to many files to copy paste

Yeah, especially if it's a model that takes 5 minutes to think 🤣

Then has a non existent import the next turn 💀

GPT-5 seems decent for planning/debugging, but it can mess up its implementation

Maybe, can just combine with Claude too

Where GPT-5 is the architect and Claude does the actual implementation

Definitely more bug free, at least on LMArena

Depends on the language too

GPT-5 has like 10-15 errors when writing a less common language

Opus had about 4

Both Opus and GPT-5 should be good at Python

Depends what you mean by better ig

GPT-5's overall vision is better

It solved race conditions in C/C++ more elegantly

They're about on par in terms of no syntax errors etc

Hmm maybe JS, but Claude isn't as good in pure JS

I'm not sure, haven't tried it that much with Java

[urban sky]

@echo aurora Will PDF integration be added?

[verbal nimbus]

Because it's harder to use via CLI

Java is very IDE dependent

[green marsh]

hello i saw some great videos about the lmarena hope to enjoy and add to my architecture work

[verbal nimbus]

JS and C/C++ probably, but GPT-5 comes out in front for those two languages

Not sure how good it is in Rust

Not sure about C#/Java, but they're very IDE-dependent

Yup, that's probably the largest factor

• tons of textbooks in training data

And tons of computer systems + concurrency + graphics + DSA textbooks use C

With thinking on both, yes

It solves race conditions better

Not necessarily

C is the one used in a lot of textbooks

Ideally it should be good in Rust

For bigger projects, C++ has classes and namespaces, which is handy

Which allows easier application of OOP/software design principles

You can connect C to Python too, for expensive operations

Well you can embed assembly in C

Probably, but I've only seen it in Linux code, which uses C

Not too sure, but C++ is younger

Not sure about the history, but I think systems ppl prefer to manage a lot of things themselves too

C++ is mainly for bigger projects that benefit from applying OOP techniques ig

C# is fine too if on Windows

It depends what you're doing ig

Larger apps often combine languages

If you're on Android for example, it'll be better just using Kotlin than JNI since the OS will take care of a lot of things

Java Native Interface

You can use it to write C/Rust/C++ code for Android

It's probably not very noticeable, there are entire games written in C#

C# is quite common in game development

Godot, Unity, MonoGame, etc

Probably less, but probably still good

Only issue is that it might be hard for agents to setup the IDE, but there might be tools for that

Well, low languages like C/C++ have a whole set of issues to consider that higher languages don't

Like garbage collection, pointers, memory leaks, etc.

Even basic stuff like reversing a string

Not familiar with Delphi, but it probably would do fine in Rust

AI companies would probably prioritize it since it's quickly rising in popularity

Also it's very easy to use via the CLI

Go is more for things like multiprocessing I think

No, only C and JS so far

Claude seems fine at Rust

Not familiar with Delphi

Node is pretty powerful

Well they solve different tasks

For speed you can also use WASM if in browser

Not sure, Claude seems a bit poorer on pure JS