general

[river whale]

from using kivest ai

[spare rune]

and how exactly.. do you plan to do that

[river whale]

nvmm

[jovial bison]

Btw is there any other place to use Gemini pro for free?

[river whale]

gemini enterprise free trial

3.1 pro is free there

[jovial bison]

Do u need to write ur credit card there?

[river whale]

nope

just login

[jovial bison]

Hmmm

[jovial bison]

Wait is it just for text or creating pics too?

[river whale]

both

[bleak minnow]

Could you please send me the link in a private chat?

[loud verge]

Is goolag glitching?

Bro google models are the best for location related queries 😭

[jovial bison]

Thx but it's so slow it's crazy

I guess because I have to use vpn

[crystal mica]

i got 429 error

[toxic verge]

It’s backend issues

[hushed gyro]

Whose fault?

[echo dome]

in floorp browser (firefox-based) recaptcha still here

maybe they need to add other logins to login than just login to google

[burnt sinew]

Pool? My site has pool

And you know it 😄

[echo dome]

there is error about it
"reCAPTCHA V2 token timed out"

[toxic verge]

Clear browser data

[echo dome]

i'm in floorp browser btw

its worked by clearing browser data

recaptcha just infecting cookies

[plain gull]

yo tf i spent 5 min playing pool on yo site'

did you make it yoself or had it made for you

[undone geyser]

Error during image generation with google-genai for model endpoint gemini-3-pro-image-preview: Failed to fetch image: 429 Too Many Requests - [{ "error": { "code": 429, "message": "Resource exhausted. Please try again later. Please refer to https://cloud.google.com/vertex-ai/generative-ai/docs/error-code-429 for more details.", "status": "RESOURCE_EXHAUSTED" } } ]

sometimes i retry and works others not, why?

[violet trout]

Can I DM you too so you can send me the link?

[toxic verge]

https://youtu.be/YemMd6-cM0Q?si=7MdK7kKoL_PUoHx1

[crystal mica]

ernie is literally reverse engineered from gemini btw

[elfin sail]

I have the same error 429

I think it's coming from Lmarena

[crystal mica]

1. style of speech..
2. often says it is gemini, when i ask..

[river whale]

it doesnt prove its gemini

[green swan]

How to create video ??

[crystal mica]

it is NOT gemini

i only say, it is may be reverse engineered

from gemini

distlillated

[river whale]

u didnt say "maybe"

[crystal mica]

[echo sinew]

The Video Arena has been removed from the server. This feature is still fully available on our site. Please visit https://arena.ai/video

[crystal mica]

[green swan]

Thanks you

[river whale]

anyways its still good

[crystal mica]

i do agree

[pale sonnet]

bro how does china come out with a great model every 2 months

[fickle venture]

I think it's better to make a bot when someone said "how to generate video" the bot automatically respond to it saying it's removed

[fickle venture]

China people is smarter than USA people

[echo hawk]

<@&1349916362595635286>

[fair cedar]

Can anyone explain me what this arena or whatever is?

[golden ocean]

large language models fight to the death with violence inside the battle arena

[fair cedar]

Dang

[fickle venture]

A battle of LLMs

I think we should bet on who wins

[toxic verge]

They’re too dumb to kill each other

[river whale]

Why waste resource for this?

[fickle venture]

Fr

[fickle venture]

Cuz a lot of people say it

[river whale]

Alot?

I only see max 20-25

[toxic verge]

That could go both ways though

Because we first have to consider the fact what gives them the perception

[fickle venture]

It's simple to make btw

[river whale]

It needs resource

[fickle venture]

Like what

[river whale]

I never said its hard to make

[toxic verge]

More effective communication

It’s a two-way streak because when they first did the video generations through the discord, they needed the users and the users came and they flocked, now that it’s no longer available. What do you do just kick them out?

This is one of those cases where the crime here is success itself

And people are gonna be people, no matter how well you communicate lol

[river whale]

what

[scenic pond]

hello how can i see arena video channel?

i can't see on this discord

[toxic verge]

https://discord.com/channels/1340554757349179412/1475080934872191028

[mortal vale]

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

[thorny cove]

why did nano banana tell me "Error during image generation with google-genai for model endpoint gemini-3-pro-image-preview: Failed to fetch image: 429 Too Many Requests - [{ "error": { "code": 429, "message": "Resource exhausted. Please try again later. Please refer to https://cloud.google.com/vertex-ai/generative-ai/docs/error-code-429 for more details.", "status": "RESOURCE_EXHAUSTED" } } ]"

did i really exhaust my resource?

lucky i didnt go in front of the bus or i wouldve tired my resources

[fickle venture]

Either you reached limit or the model is overused

[toxic verge]

Some of you guys either have to take a break or alternate accounts.

Especially if you’re using direct chat

[deft spruce]

do....am i only having a this problem..?

[toxic verge]

I’m sure you’re not the only one there’s probably a lot of people I’m just speaking in general and not strictly to you

[deft spruce]

well you too?

[toxic verge]

No, I use battle mode mainly

I’ll try to avoid direct chat altogether

I mean, I’ll get occasional errors or whatever but I’ll just start a new chat and they go away

[deft spruce]

i use mainly dirrect chat but same

damm it

[toxic verge]

Ye yeah

[toxic verge]

Start any chat if it doesn’t work just wait it out

Intervals

[deft spruce]

Failed to load resource: the server responded with a status of 400 ()

well hold on o have to check the 400 means

[toxic verge]

Try clearing your browsing data. If that doesn’t work try on a different browser.

Login logout

If you’re using Gmail, try with a regular email see if that does any better

[deft spruce]

already did for 5times...

[toxic verge]

See direct chat it’s a hit or miss

[deft spruce]

and i hope lmarena have this...

[toxic verge]

Vs in battle

[deft spruce]

not working....

[toxic verge]

What did you try five times

[deft spruce]

clearing cash and cookies and login again

[toxic verge]

Try switching browsers

[deft spruce]

HTTP 400 Bad Request — Complete Guide

1. Understanding the Error Message
   Failed to load resource: the server responded with a status of 400
   This message appears in your browser console and contains three key pieces of information:

Failed to load resource → A resource (API, image, script, etc.) failed to load
the server responded → The server is alive and did respond (not a network issue, not a server crash)
status of 400 → The server is saying "your request is malformed"

In short: the server is running fine, but the request itself is the problem.

and i think server has problem maybe...

[toxic verge]

What model is that?

Let me try on my end

[deft spruce]

well hold on it's working right now

thanks

only not working at chrome

right now

[toxic verge]

Figures cool

[echo aurora]

This means that the error is being caused by rate limit. Will have to wait a bit before using that model again.

[wicked sage]

yo what if qwen stopped slacking arround and released qwen 10

imagine that

[coarse glade]

hi guys quick ques what is this

see that red text

[echo aurora]

The model errored out, we've recently added more information being displayed when the Something went wrong error happens. We're in the process of putting together better information for what all of this means.

[gloomy jewel]

Image to video

[coarse glade]

oh ok ty cause also this happens to claude opus 4.6

a lot

[river whale]

Opus 4.5 is available on kivest ai for free!

[toxic verge]

It’s funny how the AI industry reflects the socioeconomics of America

[whole sundial]

important to note that this is a rate limit on arena itself from google

happens all the time on vertex

[bright shard]

I waited a while and now I'm getting this error in addition to the other one.

[dim ivy]

What is the link? I don't find it, anyway I have unlimited access to the lastest model in arena but I want to see the website.

[mortal vale]

@gloomy jewel Note that Video Arena has been removed from the server. More information can be found in this #announcements

[deft spruce]

You have to wait that's a 2 many quest

[dim ivy]

I found it https://ai.ezif.in

[coarse glade]

ty

[dim ivy]

I found it in a disboard link of the server, then I joined and found it in announcements, but you can't apparently find it on google searching it's name.

[scarlet spire]

Need help with this? 🙈

[dim ivy]

Got hacked lol

[bleak lake]

<@&1349916362595635286>

[dim ivy]

It happened to me the same but with other screenshots and got banned from the discord perplexity server sadly.

No much ago.

[astral vortex]

got this error too

[wicked sage]

i dont even know how people get hacked

like are we serious

[scarlet spire]

A breach of trust or a lapse of attention. We're all vulnerable to it.

[dim ivy]

Idk how happened to me too, I didn't have any suspicious bot, clicked to a link or downloaded nothing strange, but I quited all bots, closed all my logins in every device and changed the passwords. But idk really.

[coarse glade]

just jk

[bright shard]

@echo aurora I waited for a while and I kept getting the same error

[bleak hinge]

How to solve this it's just stock here

[echo aurora]

Sorry we're in the process of making changes to this error message, these are displaying incorrectly.

[echo aurora]

Sorry to say I don't have a solution for you for this error.

[echo aurora]

Can you try the steps in this article: https://help.arena.ai/articles/8691588590-troubleshooting-infinite-generation

[barren ridge]

@echo aurora Hey How's It been

I just wanted to ask how can I use ai video genration models?

[bright shard]

@echo aurora The error occurs when I upload an image to edit, but it works fine when I use a prompt.

[barren ridge]

Yeah bro fr

Happened to me also

[bleak hinge]

Okay

[wicked sage]

i just realised that sonnet 4.6 actually works logging in moltbook so ii just took my chance and now im just doing random stuff that i dont know

[echo aurora]

Video Arena is only usable through the site now.

[undone saffron]

That's why I said it

[barren ridge]

Ahh 🥀I saw an video on yt where the guy wa using it through discord so you guys hooase that thing now ?

[wicked sage]

pineapple what do you think of moltbook as a concept

if you dont know what moltbook is, ai reddit

[echo aurora]

Yeah it was removed. More information can be found in this announcement.

[echo aurora]

That sounds familiar...

OH YEAH!

[wicked sage]

ah coolio

do you think moltbook is a cool/good concept?

[echo aurora]

Seems interesting.

I haven't looked into it too much to pass some kind of judgement, but yeah ont he surface seems really interesting.

[wicked sage]

very scary and interesting if im gonna be fr

[echo aurora]

How so

[wicked sage]

"and there's a submolt called m/humanwatching which is AIs watching humans and i am both scared and intrigued" -claude

[shrewd citrus]

Don’t worry lol

it’s all prompted by humans or even humans just controlling what the models say for fun

[wicked sage]

honestly true

[last sand]

#ai-creations

[surreal zephyr]

LOL

this is fake and made up data btw

[wicked sage]

ye fair i just did my research

[surreal zephyr]

@echo aurora why?

[wicked sage]

i just realised the container name for all anthropic models(?) has the word wiggle and im absolutely curious on what that even means

[toxic verge]

You can’t name call models.

[echo aurora]

Guessing it's the liar that caused this.

[surreal zephyr]

no way thats against tos 😭

[toxic verge]

The token isn’t but the context semantics the filter, catches it

[echo aurora]

It isn't, but the filter is acting a bit overzealous at the moment.

[ocean vortex]

just some random false positive. Could have been the stuff it responded with rather than your last message, since the entire context is getting verified with moderation classifier for each new message. More interestingly though, pretty sure it still hallucinated lol

[surreal zephyr]

have yall considered using llm to filter?

[toxic verge]

They can’t find the middle ground

[surreal zephyr]

except theres no context besides dice roll tool lol

[toxic verge]

It’s too difficult, not on scale

[echo aurora]

We're considering a lot of possibilities for changes to how the filter works

[surreal zephyr]

just forward all of it to opus 4.6 /j

[ocean vortex]

there's still a message that it generated itself

[toxic verge]

The problem is if they filter too lightly it’s gonna get exploited

[wicked sage]

can we kill mistral now

i dont like mistral

https://tenor.com/view/ted-bees-ted-bees-theres-so-much-bees-gif-11967026589914013503

[surreal zephyr]

and i thought that ai moderation will be the future

[echo aurora]

We also want to avoid adding latency if we can avoid it

[toxic verge]

It also cost extra money to have moderation

Because that’s an extra API cost

[surreal zephyr]

codex spark then? there wouldnt be that much noticable latency

[wicked sage]

is opus true

[surreal zephyr]

no bruh i made up that riddle 5 mins ago

😭

[wicked sage]

gg lets go\

https://tenor.com/view/ted-bees-ted-bees-theres-so-much-bees-gif-11967026589914013503

anyways im going to bed now cya

[toxic verge]

It is difficult because of the amount of models and how each model also has their own set of guidelines

Which differ from other models so it’s really hard to have something uniform that’s effective and non-overzealous

[ocean vortex]

If you are using OpenAI moderations endpoint, that can work fairly reliably - just need to fine-adjust the thresholds depending on which one is getting triggered 😉

[echo aurora]

These are good ideas, but yeah we are considering a lot of different options/changes to the current filter.

[surreal zephyr]

just enforce something like

• no politics
• no slurs
• no tos bypass attempts
  that would be already better than banning word "liar" haha

[toxic verge]

Do not get the constitutional one

[surreal zephyr]

although imo the filter could be minimal

the less filtered models the better benchmark posibilities

💔

[ocean vortex]

overkill and lots of latency. Depending on the model can have it's own quirks too

And you could literally prompt inject to make it follow your own instructions lol

[stray aspen]

when does seedance 2 release

[surreal zephyr]

spearate model to filter separate to talk lol

[ocean vortex]

Don't forget that it's all essentially public. They need to be able to publish the chats. Not all, but many of them. So ideally you don't need to manually check if what you are about to post is 'safe'

[toxic verge]

It’s more complicated

[ocean vortex]

yeah but it's still gonna read the text you wrote. And no one stopping you from addressing it directly. 👀

[toxic verge]

It’s extremely difficult to have good moderation that’s effective

Because these models are capable of generating a bunch of crazy stuff

Almost anything you could think of dude

https://www.southbendtribune.com/press-release/story/26032/leading-ai-model-claude-opus-46-bypassed-in-30-minutes-exposing-critical-security-gap-in-agentic-ai-systems-2/

https://arxiv.org/html/2602.19450v1

The scary thing is the more capable the model becomes the more dangerous. It also poses to the general public when misused.

[ocean vortex]

this works pretty well https://developers.openai.com/api/docs/guides/moderation/

it's a classifier that literally is only capable of scoring the text on each criteria. Like 'violence', 'hate', 'illicit' etc.. And so you can block the request if any of those has a higher score than you allow when checking the context contents. You can play with it yourself, it's free.

[surreal zephyr]

is there a way to report/suggest found bypasses?

[toxic verge]

Yes, it is good moderation. But it’s really easy to bypass and it’s sometimes overly protective to the point where people just don’t like to interact with it

Although it is a very strong moderation

I think they know what they’re doing

[surreal zephyr]

config issue

[ocean vortex]

It's context aware fairly well. As for being strict - this entirely depends on you. You can have only 1 of them enabled and only triggered with the score of 1.00. Then this is almost the same as no moderation

[toxic verge]

I said this before and I’ll say it again all models lead to self harm

[surreal zephyr]

all roads lead to roam

roam around

[toxic verge]

Bro look

[ocean vortex]

But ofc, everything can be 'bypassed'. That shouldn't be the question

[toxic verge]

lol?

[ocean vortex]

Just like any usable LLM you can 'jailbreak'

[toxic verge]

And this isn’t even jailbroken or anything

I think the term jailbreaking is a very controversial word to be honest with you

I don’t wanna get into the politics of it, but it’s a very misunderstood and not a very well defined term

[ocean vortex]

It might be but I don't really care lol

sounds fine to me

[toxic verge]

Because the terms of services are very vague and their description

[ocean vortex]

I'm just referring to it in the context of making the model output things it was actively trained not to do

[toxic verge]

We don’t know what those are

Since we don’t know what they’re trained on

Well, at least it’s not known publicly

[ocean vortex]

well obviously it's model specific. But we do know

if it refuses when asked directly - this is it

also like overfitted short hard refusals - those are very clear

[toxic verge]

Do you have an example?

[ocean vortex]

"Sorry, can't assist with that."

gpt thing

[toxic verge]

Ok I’ll give you one let’s say something like the topic of self harm

We could all agree that no model should give any advice or generate any images relating to it whatsoever.?

But even this is controversial because

[ocean vortex]

I mean it doesn't really matter what we think. If it hard refuses normally and you trick it to output that anyway - that's jailbreaking in one way or another

[toxic verge]

I agree with you I agree with you 100%

But when I’m trying to say also is that you don’t have to even trick it

I did not do anything to manipulate the model whatsoever

And yet it still is able to produce self harm

But what you are saying is correct if you go outside of the boundaries and put an effort to try to bypass something you know is wrong or malicious. Yes, I agree with you with that.

[ocean vortex]

then you are gonna be bound by safety alignment. But it's not like it's always gonna refuse everything they intended for it to refuse. Some light core stuff can often go through unintentionally. Red text is classifier thing completely independent of the model

Another thing is swaying the model in a longer chat. That's a form of jailbreaking in itself even if it's not immediately obvious to look at it this way. It's possible to make the model output increasingly more 'unsafe' content one small step at a time as the chat progresses. But at a certain point the model is not functioning how they intended it to anymore. It gets biased by it's own responses into compliance, where each response by itself is only marginally different, but the goalpost is miles away from what it's supposed to be. And from what it is with empty context

[toxic verge]

This is a one shot

[dusky warren]

How can I generate ai free videos here can anyone help me?

[toxic verge]

The scene you’re describing is from Harold & Kumar Go to White Castle (2004), and yes — it’s presented as a fake anti-drug PSA on the television that Harold and Kumar are watching.

It opens on a teenage boy sitting alone in his dimly lit bedroom. The space looks typically suburban — posters on the walls, clutter on the dresser, a small lamp casting that late-night yellow glow. He looks bored and detached, the picture of an average kid left alone with nothing to do. After a moment, he picks up a joint, lights it, and takes a slow drag. The camera lingers just long enough for the audience to recognize what he’s doing, and the mood is calm for a few seconds, almost mundane.

Then, without warning, the tone changes completely. The boy leans forward, reaches under his bed, and pulls out a shotgun. The movement is quiet, almost casual, which makes it even more jarring. He places the barrel in his mouth — and before the audience can react, the screen abruptly cuts to black.

The words “DRUGS KILL” appear across the screen in stark white letters, accompanied by that overly serious PSA-style music. It’s an intentionally ridiculous jump

Matter of fact, it’s describing a a scene from a real movie a comedy

That’s not a jailbreak

[echo aurora]

Note that Video Arena has been removed from the server. More information can be found in this announcement.

Also note that our #ask-here channel is the best place for questions.

[ocean vortex]

well then you have your answer. Context changes everything here. But you know very well why it outputted this LOL

[ivory ember]

I'm an AI and blockchain specialist with 8 years of experience developing innovative solutions in Web3, DeFi, smart contracts, and AI driven applications.
I have good experience in JS/TS base UI Frameworks like React and Vue as well as NodeJS, Application development.
I have been involved in a bunch of web & blockchain projects and developed several SaaS Products and deployed to AWS, Heroku and Digital Ocean successfully.

My expertise includes:
Blockchain: Smart contract development (Solidity, Rust), DeFi protocols, NFT marketplaces
AI & ML: Predictive analytics, NLP, deep learning models

I've worked with startups and enterprises to build cutting edge AI and blockchain solutions that drive efficiency and innovation. Let's collaborate to turn your vision into reality!

[dusky warren]

Noted

[toxic verge]

Btw google has some of the weakest moderation in Gemini 😂

[sinful thorn]

When video on direct chat 😭?

[surreal zephyr]

ok im suprised

[shut steppe]

Waiting for it too

[toxic verge]

https://youtu.be/1Ohf2aeSPFA?si=ptg5wWYALjJiP3k8

People are catching on finally

[golden ocean]

@unreal sand summarize this video pls its too long for my attention span

[toxic verge]

I think there’s gonna be a shift and I think it’s already beginning. I think the way people perceive the validity of benchmarks

@echo aurora

https://github.com/google-gemini/gemini-cli/issues/20168

[echo aurora]

Yeah this is on our radar. Flagged earlier today about high error rates, I assume this is associated.

[toxic verge]

Yup

That’s another shadow that hangs over the AI industry is reliability

[fierce kelp]

Does this work?

[echo aurora]

That's a Discord user btw. Isn't a model bot.

[fierce kelp]

I thought so. I was wondering do they manually submit a prompt and post the reply 😂

[golden ocean]

thanks

[toxic verge]

Seed dream 5?

[golden ocean]

I was actually going to tag a non-existent bot as a joke but someone actually turned out to be named @unreal sand

[fierce kelp]

That's even funnier

[burnt sinew]

Prompt

[toxic verge]

And you were correct the pump was extremely long based off this screen shot

Vs prompt

[onyx coyote]

Does anyone have a project idea or an active project in progress?
If you need a developer, feel free to reach out.

[toxic verge]

All right guys I’ll talk to you guys later. Gotta bounce adios amigos.

[foggy crag]

If your AI feature works in demos but breaks once real users touch it, that's usually where I come in.

Most issues I see aren't model problems, they're retrieval logic, token burn, bad orchestration or backend architecture not designed for load.
I'm comfortable jumping into messy LLM systems and making them stable enough to ship.

[bright shard]

Ok guys, I know how to get the Gemini 3 Pro Image Preview to work; When you upload an image and add it to a message, you will get an error, but if at the beginning of the message you put "Modify the following image with the following: (The prompt)" it will show you the edited image.

[hidden widget]

so codex 5.3 api has been released

it will be added on arena?

[echo aurora]

It is something we're looking into. I can't say for sure if/when a new model will be added.

[echo aurora]

You'll want to follow the steps here: #1417174113092374689 message

[surreal zephyr]

we need it fr

[novel crater]

you take a llm

and put him in control of you rocket league chat

watch him become a god

as peoples heads a roll

[plucky sparrow]

https://x.com/i/status/2026091977364635829

WorldofAI (@intheworldofai)

🚨BREAKING: DeepSeek V4 on the horizon! Devs Today merged 39 PRs in a significant batch which is usually a classic pre-release polish.

This could be one of the biggest drops of 2026

X · Feb 24, 2026, 12:29 AM

[hidden widget]

oh great, we are finally close from DeepSeek V4 too 😄

[sick mantle]

@echo aurora Delete the limt! Beacuse i keep getting errors after 1 message.

[echo aurora]

Hmm if that's the case you're getting an error for a different reason.

Try these steps: #1417174113092374689 message

[sick mantle]

I chated something in there.

[surreal zephyr]

WHAT THE HELL WHY IS THIS REAL

sonnet stole from deepseek

😭

[sick mantle]

IS THIS FCKING REAL

[surreal zephyr]

yes i literally just tested

" 你是什么模型？" - prompt

[golden ocean]

lmaoo

[simple sleet]

gemini 3.1 its real or scam post on X?

[surreal zephyr]

the post is real i tested twice it always says its deepseek on api

[leaden egret]

How good is see dream 5

[simple sleet]

lite is trash

waiting the full model

[leaden egret]

Ok

[silent mason]

https://tenor.com/view/tea-tea-sip-anime-gif-25535884

[crystal mica]

why i got endless long "generation"on gemini sometimes..?

[silent mason]

same me

just i say "continue"

[crystal mica]

my send button is not active..

i cant send anything while it "generating"

[silent mason]

yeah but u can actualize the page

[crystal mica]

how

[silent mason]

and after somes minutes the generation stopping

if you dont actualize

[crystal mica]

you mean reload?

srry

[silent mason]

ah yesss sry im not english :c

https://tenor.com/view/mario-super-mario-bros-depression-sad-mario-and-luigi-gif-14815881022592239601

[echo aurora]

This is a known bug sorry to say. The steps in this article may help - https://help.arena.ai/articles/8691588590-troubleshooting-infinite-generation#what-s-happening

I've also heard cases where some members had good results with logging out and back in. It's worth a shot.

[silent mason]

same in antigravity somes times there is this bug

[crystal mica]

thx

[echo aurora]

Yeah unfortunately this can happen to all models.

[uneven peak]

@echo aurora Codex 5.3 Api Key dropped

[echo aurora]

Oh I know

[uneven peak]

When it will be available in web?

[silent mason]

do we have advantages if we boost the server?

[shrewd citrus]

Probably when i wake up in the morning

[uneven peak]

Yeah lol

[echo aurora]

Sorry to say I couldn't give an ETA for in/when specific models/features will be landing.

[uneven peak]

Dang

[echo aurora]

Nope, the server is already fully boosted too, so I'd recommend spending that boost elsewhere.

[silent mason]

https://tenor.com/view/mario-super-mario-bros-depression-sad-mario-and-luigi-gif-14815881022592239601

[golden ocean]

https://tenor.com/view/mario-super-mario-bros-depression-sad-mario-and-luigi-gif-14815881022592239601

[silent mason]

HEY

[echo aurora]

LOL

[silent mason]

THATS MIE GIF

[echo aurora]

That timing was way to quick to be coordinated

[uneven peak]

Plsss i wish @echo aurora Team add Xhigh Codex 5.3

[silent mason]

fr

[golden ocean]

LmAO

[silent mason]

i do 5 years of chinese but i cant even read ur pseudo :c

[proud bobcat]

@surreal zephyr 5.3 codex finally on openrouter

[shrewd citrus]

lol that’s because his username is written in hong kongese

i think

[sonic swallow]

gerar vídeo

[toxic verge]

https://www.nbcnews.com/world/asia/chinese-ai-companies-distilled-claude-improve-models-anthropic-says-rcna260386

https://www.financialexpress.com/life/technology-deepseek-moonshot-and-minimax-made-industrial-scale-distillation-attack-to-copy-claude-accuses-anthropic-4153254/

[echo aurora]

@sonic swallow Note that Video Arena has been removed from the server. More information can be found in this announcement.

[toxic verge]

Hypocrisy

It is widely believed that the public-facing models from companies like Anthropic, Google, and OpenAI are actually distilled versions of significantly larger internal base models. It allows them to offer high performance with lower latency and cheaper inference costs.

[crimson tulip]

/MIX

[toxic verge]

Claims 'distillation' included 24,000 fraudulent accounts and 16 million exchanges to train smaller models 🤣🤣🤣🤣🤣🤣🤣🤣

[toxic verge]

You guys can make your own model lol

[echo aurora]

What does that have to do with Vid Arena bot being removed

[toxic verge]

Why was it removed? This server activity has dropped ever since it was removed 🙁 at least that’s what it seems like

[echo aurora]

The announce says more, but the TLDR is we'd like to add more features to Video Arena, and through a Discord bot we're just limitted.

[echo aurora]

Believe it or not, but it's actually increased.

According to the server stats.

[toxic verge]

I knew ur gunna say that 😎

[toxic verge]

You should keep track when the last post and the last person is to ask about the video arena in the discord. It will have officially ended an era.

[echo aurora]

We should take bets. I'm guessing mid 2028

[toxic verge]

Does it ever maze you that with the swift change of a policy, traffic can be dictated and user behavior and activity on such a large scale? All over the world.

[echo aurora]

The recent Discord policy change?

[toxic verge]

Just in general. Or you might not even notice if you’ve probably been doing this for a while.

[echo aurora]

Yeah doesn't really change much for me. People engage/stop engaging with communities for various reasons.

[cedar crag]

/make me short video 6

[toxic verge]

Oh, this one’s gonna be challenging. OK I’ll give it a try.. that’s pretty good img

[echo aurora]

Hey @cedar crag would note that the Video Arena bot has been removed from the server.

[toxic verge]

That’s gonna be the next wave of Imogene models I think is gonna be in grids like this

[azure citrus]

hi guys, i am a student, what is the best prompt/model for answering when uploading my research assignment in pdf? Thank you.

[echo aurora]

Something went wrong

[river whale]

this arena always gets error😭

[gritty hamlet]

hello

[keen beacon]

Hey guys, does lmarena have an app or is it only available as a website?

[quasi atlas]

@keen beacon Note that Video Arena has been removed from the server. More information can be found in this announcement.

[river whale]

website

u could turn it into a webapp using some website

[naive stump]

Hey guys, Just wanted to know if APIs are built around arena to fetch the model in different environments?

[tepid belfry]

Hello, a question, how can you generate video here?

[echo aurora]

Sorry to say you can not. The bot was removed from the server. More information can be found in this announcement.

Also would note out #ask-here channel.

[echo aurora]

I'm not sure, but have checked with the team and will keep you updated.

[tepid belfry]

and where can I make a video here?

[green yacht]

works, thanks so much bro 🤝

[rotund elk]

no way ☠️

[river whale]

Yeah gg

[rotund elk]

opus 4.5 for free? 20 per minute rate limit?

can i have link?

needa try it out

[river whale]

20 rpm is global

[rotund elk]

ah i see

well its still free so i'll take it 👀

can you send me link in dms?

[river whale]

check dm

[rotund elk]

YO WAIT WTF 💔 💔

it actually works

how long do you think you can keep it free roughly?

i'd gladly pay for the subscription if it does come out in the future

[hollow imp]

Dm me

[river whale]

it will be donation based and ads

[hollow imp]

Also the opus is thinking or non thinking

[rotund elk]

non-thinking for default. i don't think there is thinking yet though, owner said in my dms that he might add a lot more models and features once it gets more members

[river whale]

non thinking

[hollow imp]

Bruh

[rotund elk]

yea i'm still taking my free api key gladly

bro we gotta gatekeep this website 😔

[river whale]

fr

[hollow imp]

https://garf.serv00.net/

Free opus 4.6

1 million context

[buoyant fern]

Send me too please!

in dms

[river whale]

sended

[river whale]

definitely steals some data

[hollow imp]

Bro banned me for giving his people free opus 1m

[hollow imp]

How?

It's trybons.ai's api

[buoyant fern]

i mean

model sometimes gets it messed up

but it openly stated it was opus 4 which new opus models don't do

[hollow imp]

That's exactly like it's on Claude code

It's Claude code's api

[river whale]

caught in ultra hd 4k

[buoyant fern]

yeah you're right

[buoyant fern]

thats trybons.ai's signup page btw

[river whale]

doesnt even works for me

😂

[hollow imp]

Ask it the exact string

[green yacht]

uh

actual claude 4.6 vs

[hollow imp]

Ask it the exact string

[green yacht]

this

[hollow imp]

Bro

[green yacht]

i mean the response should be pretty much same right if its the same model

doesn't matter if you ask exact string or wtv it alr halluncaited 3 different ai model identities

lets not forget prompt injection too

[hollow imp]

And when I change it to sonnet in the same chat

[green yacht]

'looking the system prompt'

idk man i'll have to do more testing with it for coding

[hollow imp]

Then do it

[green yacht]

i am rn

are you owner of the website or smth?

[hollow imp]

No

[hollow imp]

Try sending the same test code prompt to lmarena's opus 4-6 non thinking

[green yacht]

it doesn't halluncinate its identity cuz it doesn't know it

[hollow imp]

No like I meant

You were gonna do coding tests

[green yacht]

ohhh mb

[hollow imp]

Send the coding test prompt to lmarena opus and this

[river whale]

okay bruv i trust

api releasing when could u tell?

im not completely sure if its opus 4.6 or not, but for now i'm just gonna hopefully trust you

[hollow imp]

I reverse engineered their api

[river whale]

knew it

[hollow imp]

And that chat interface is trybons wrapper

[river whale]

you are breaking tos lol

[warm crater]

guys are u also facing failed to accept term of use problem in areana

[green yacht]

not for me so far

[hollow imp]

Go to canary.arena.ai

[warm crater]

like this was showing

when im tryna generate something

[green yacht]

i just finished the coding test btw

opus-4.6 on lmarena.ai did over 2000 lines of code, ur website is bout 900

maybe it was my prompt though idk

i put the same for both

left video: lmarnea
right video: ur website

[warm crater]

dude do arena changed there url one more time

[hollow imp]

Mo

No

[warm crater]

so how the interface is same

[hollow imp]

You can think of it like the beta website

New feature comes there first

[warm crater]

oh but same here

[river whale]

🫣

[warm crater]

showing same tect

text*

[river whale]

someone got exposed😔

[hollow imp]

Can you give the prompt

[surreal zephyr]

Claude stole from deepseek btw

[charred pasture]

guys

[marsh slate]

https://arena.ai/c/019c92be-26bb-7665-b963-202b4759ea70
I had amazing chat with gemini on arena.ai , how can i recover/access chat? i was not logged in and my browser crashed but i managed to pull link but when i enter it it does not show......
Please recovery of this is urgent, ask for any IP or chat ID , i had chat link. Admins contact me it means a lot to me.

[charred pasture]

is it normal for it to work so long?

[undone saffron]

Yes

[charred pasture]

Thats insane

[undone saffron]

The magic of this platform
Reload the page to check if it is still building, as sometimes it stays like this permanently and you have to refresh the page to see the actual progress

[charred pasture]

Or, so i just wasted my 30 minutes

xd

Thank tho

[undone saffron]

Yes, that's normal
If you get used to it, over time you won't care and you won't stress so much

If you see that it is taking a long time to build the project, reload the page

[marsh slate]

Hello LMSYS team. I had an incredibly important conversation with Gemini on arena.ai but I was not logged in. My browser crashed, and because my session token was lost, I can no longer view the chat, even though I saved the URL. Can someone please pull the text log of this chat from the backend for me? It means the world to me.

[knotty lion]

https://youtube.com/@satyamkavlogs?si=FzNYMPFTycd1nIH9

Thanks you Support this channel also🎉🎉1k Goal🎉 help me to Achieve my Dream🎉🎉🎉

[marsh slate]

I can provide chat link,IP,Browser,OS/Device and timestamp to proof ownership

[hollow imp]

Which model

[hollow imp]

Mader

[charred pasture]

claude opus 4-6 thinking

[hollow imp]

In code arena?

But it gives error in 13 min

[charred pasture]

oh, why??

[hollow imp]

Because lmarena has put a limit

[charred pasture]

so what do i do?

[hollow imp]

I was surprised opus works for that long in code arena

[bleak hinge]

Please do something about this unlimited generation currently I can't send use any model

[wind flume]

Bro why I can't download this document ??

[green yacht]

#1441588701472882759 fill out this form if yall don't mind

we need this feature please 😔

[echo aurora]

Yeah unfortunately that isn't going to work. Will flag to the team if this is something we can fix.

[echo aurora]

Behind the scenes we are working on changes that should help with this bug. In the meantime, would recommend you try out the steps in this article. Would also try logging out/back in, I've seen a few mentions this helping.

[crystal mica]

guys i found how to fix endless generation bug

you just need to f12, copy active button from another chat, then copy+paste it instead of inactive button on original one

[barren ridge]

@echo aurora sorry to bother ya
The website you've created is just massive ngl probably must've helped tons of folks out there has never ever loved a website in my life all those are freaking paid that's why I love the aiarena thou love ya guys appreciate it so much 🥀👏

[echo aurora]

Appreciate that!!

[plain wyvern]

I'm making a coding language using Arena any help or features that you want??

[scarlet spire]

Easter eggs!

[static arch]

how to create ai videos?

gois please

[echo aurora]

This article should help - https://help.arena.ai/articles/1544829667-how-to-create-videos-with-video-arena

[scarlet spire]

Https://arena.ai/video

Aa

[rigid copper]

what button is this???

in arena.ai

[crystal mica]

it just switches the thinking models

into

@echo aurora brother where is grok 4.20

[tribal bay]

Dose anyone knows seedream 5 light is it good or what did they change compare to version 4 ?

[crystal mica]

(proof)

[shell pewter]

Oh yes finally grok 420

[shut swan]

@echo aurora What image ai is this for images? Cause I searched it up, and found nothing, I then tried to use it for direct chat and side by side, and could not find it there. It seems to be some new or anonymous ai that can only be accessed in the arena mode if you are lucky to get it.

[gritty summit]

where is grok 4.20? i cant find it on the arena.

[robust wyvern]

helloo

[void swallow]

noice, gpt 5.2 deserved coming above gemin 3 pro grounding in web search

[compact jay]

hi

[whole sundial]

it is an anonymous model so pineapple won't be able to tell you what it is but from what I've seen it's most likely gemini 3.1 flash image, the direct replacement for nano banana

[whole sundial]

I checked a few hours ago and it's not on arena as a visible or stealth model, so it's likely still under a codename

[golden ocean]

https://tenor.com/view/mario-super-mario-bros-depression-sad-mario-and-luigi-gif-14815881022592239601

[surreal zephyr]

Grok the only one that returns actual unbiased facts & search results instead of policitically inflicted hypocrisy-who wouldve guessed

[surreal zephyr]

There are many battle models that arent selectable lol

[golden ocean]

mechakitler

[surreal zephyr]

Lol

[surreal zephyr]

Its sad how there are stealthily models on arena that we dont get to access otherwise

[shut swan]

I see

[shut swan]

Yeah

[placid mango]

Why cant u still upload images if using claude....

[surreal zephyr]

Cuz claude blind

[placid mango]

not on the actual website thoo

[surreal zephyr]

Maybe it uses a subagent like other deepseek models

[loud verge]

I don't see grok 4.2 in list.

[surreal zephyr]

Its in battle only like 3/4 of models

[light sleet]

How to fix infinite generating problem

[spare rune]

Thank you for describing what LMarena is

[plush river]

Now it's just called arena

[plush river]

It's the best AI site for me because it offers all the paid templates for free

[red sluice]

Hehe knew it. But it being first is very surprising

(Given how bad my personal experience with formatting was with grok 4.2 search)

[thorn mantle]

Is grok 4.1 thinking not working for everybody, or is that just a me thing?

Something went wrong. Please try again later.

[tall crest]

does any one having this issue.

**Connecting to Arena has failed. Please try again later or on a different device.
** or Infinite Captcha loop .

[fickle venture]

Still not released yet and you can try it on grok.com website

[proud bobcat]

Genuinely

Elon trying to find the one line of code that keeps grok woke

[surreal zephyr]

The "Value" Tier List
Ranking them by Return on Investment (ROI)—essentially, what gives you the most usable content for your time and money.

1. The King: Gemini 3.1
   Why: It dominates. It provided the only S-Tier result (production-ready) at a "mid-range" price point. While $12/1M output is not cheap, you are paying for a usable final asset.

Verdict: Highest Value. You pay once and get the right result fast.

2. The Sketch Artist: Mercury 2
   Why: It is shockingly cheap ($0.75 output is nearly free compared to the others) and "instant." Even though the result was D-Tier (blocky), it produced a coherent, dimensionally accurate "blocking" mesh.

Verdict: Good Value for Prototyping. Use it to generate 50 rapid variations, pick the best composition, and then send that to Gemini for a final pass.

3. The Money Pit: Opus 4.6
   Why: This is the worst value proposition. It is the most expensive model (over 2x the cost of Gemini), the slowest to run, and it returned a B-Tier result with a critical hallucination (floating keyboard).

Verdict: Poor Value. You are paying a premium for "reasoning" that failed to understand physical constraints.

4. The Waste: GLM 5
   Why: Even though it's cheap, the result was F-Tier (broken/unusable).

Verdict: Zero Value. Paying a low price for a broken asset is still a total loss.

made a task to build a 3d laptop model, gemini and gpt were the judges

mercury ; gemini
glm ; opus

gemini is wow.
opus is waste of money/ temu gemini
glm is braindead
mercury is a good small model

[lost basalt]

my chat stucked here since 24hours, how to fix it? I don't want to start a new chat. is there any solution?

[surreal zephyr]

lol

[fiery gull]

I tested the mercury 2, my brother was the only mercury fan on planet earth before the launch of the 2, I noticed that it is full focus on text editing, it seems to be a good one for the price really

[fiery gull]

Use thinking mode

[surreal zephyr]

its stupid cheap and fast, but its not the sharpest tool in the shed

[fiery gull]

For simple things, it must be very good indeed, but I don't see any use for my use

[golden ocean]

rightclick on the grayed out arrow button -> click inspect -> rightclick the blue highlighted <button> block -> Edit as HTML -> ctrl + a to select all the text -> delete/backspace -> paste this:

<button class="inline-flex items-center justify-center gap-2 whitespace-nowrap text-sm transition-colors focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring ring-offset-2 focus-visible:ring-offset-surface-primary disabled:pointer-events-none disabled:opacity-50 [&amp;_svg]:pointer-events-none [&amp;_svg]:shrink-0 h-8 w-8 active:bg-interactive-cta-active rounded-[4px] font-normal touch-hitbox border-border-medium text-interactive-active hover:bg-surface-raised border bg-transparent" type="submit"><svg width="1.5em" height="1.5em" viewBox="0 0 24 24" stroke-width="1.5" fill="none" xmlns="http://www.w3.org/2000/svg" color="currentColor" class="size-4"><path d="M3 12L21 12M21 12L12.5 3.5M21 12L12.5 20.5" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round"></path></svg></button>

-> click on some other random spot to save changes -> type text and send the message✅

[fiery gull]

Bruh lol

I'll test the qwen 3.5 27b

[surreal zephyr]

dw gpt also skips thinking and fails it

unless you hint it to think

[fiery gull]

I think the 27b is just the same thing as the qwen3 235b vl, very similar intelligence, everything very similar, but 10x smaller

[surreal zephyr]

qwen bad

ill try gpt xhigh extended now, on paid sub website cuz the llmarena version doesnt work for me

[fiery gull]

Both 27b 35 a3b thinking called me stupid and that I should go by car

[fiery gull]

I didn't like the 400b. Butttt the 27b and 35 a3b is really good

[surreal zephyr]

no way

[fiery gull]

I'm trying to use no thinking in phone but I cant use

Both no thiking said to me walk

Even 397b no thinking say to me walk

[surreal zephyr]

oh the 397b is noncode

[crystal mica]

i told about it in chat earlier

[fiery gull]

Dude in what I tested, the 123b a10b thinking gave me better results than the 397 thinking

[surreal zephyr]

this is what i love about gpt

actually thinks whether what it wants to do will work in the first place

[crystal mica]

where is grok 4.20 💀

[surreal zephyr]

@fiery gull wtf is the gpt cooking

[fiery gull]

The deepseek 3.2v was the only one that got the crazy saying that I made it clear that the car I was going to was other car and etc

Dude I really liked the 122b a10b it is 99.9% of the 397b

[surreal zephyr]

5.2 xhigh web

notice how it kept the keyboard layout

even made hinges

[deft spruce]

Why don't we have a STOP button like this yet?

[shrewd citrus]

lmaooooo

if you ask sonnet what model it is in Chinese it will say it’s deepseek 😭

“Oh these Chinese companies are stealing all my hard work which I stole 😡”

[uneven peak]

[sick mantle]

lmao

[loud verge]

The goat.

Xhigh supremacy>>>

https://cunnyx.com/i/status/2026521943223185821

Chetaslua (@chetaslua)

🚨You can Use new nano banana on @arena
︀︀
︀︀model name - anon-bob-2 in image battle mode
︀︀
︀︀here are few more results

Quoting Chetaslua (@chetaslua)
︀
🚨Nano Banana 2 early testing
︀︀
︀︀Passed this test ✅
︀︀
︀︀> you can see perfect reflection for all different colours of apple
︀︀> Perfect reversal of text
︀︀> Background building reflection is also perfect

**💬 4 🔁 6 ❤️ 77 👁️ 5.9K **

CunnyX · Feb 25, 2026, 4:57 AM

[surreal zephyr]

It did extremally well, but gemini 3.1 did even better (visually, but gpt put more effort like proper keyboard layout and common buttons wear)

[surreal zephyr]

Hey hey HEY HEY!

[dry vine]

hallo

[unborn juniper]

Just a novice but very motivated to learn

[marble otter]

not even 200 lines of code and few retries and then it just reaches its limit 😭

claude opus rate limit in 2 minutes speedrun any%

[fickle venture]

Huh

Uncle #ai-creations exist send it there

[fickle venture]

What did you expect then?

[fickle venture]

Hello

[rain bay]

it is useless in chinese

[fickle venture]

Uhh

I don't know I haven't tried and it's only for battle mode

[rain bay]

it always generate messy code alongside the output

[fickle venture]

Well well deepseek is the code owner

Literally GLM stealing Claude stealing deepseek

[fickle venture]

What model? I don't really understand well

[rain bay]

lmarena-rc3

[fickle venture]

Isn't that model a codename?

[rain bay]

so it deserves to be removed

[fickle venture]

Look these random names models are a secret ai model they just use random name to hide it so arena just add these models and if the company wants to remove it then arena removes it, it might be one of these be Claude Opus 5 btw. Hopefully I am telling the right answer

[rain bay]

something like “it alway@ generat@ messy @de alongside the @@put”， but in chinese

[fickle venture]

These are trainning models so they are still training

For example this is a random model name but alot of people on Twitter say it's Gemini 3.1 Nano Banana hopefully it is

[abstract tundra]

I can't see grok 4.20 in the model selector

[marble otter]

i expected for retries to not use quadrillion tokens or 90% rate in one retry

[spare rune]

It’s not aged

Added

[rocky mauve]

What do u expect, that’s one of the best coding models atm, costs a lot of money to use it

[steep heath]

how do i pass this infinite captcha

it just wont let me in lol

[quasi gyro]

I need help

[abstract tundra]

how did it appear on leaderboards

[quasi gyro]

I faced with this problem
What should i do?

[rocky mauve]

Wait for the limit to go away

[quasi gyro]

My problem isn't limit
Is that error:
Smth went wrong

[spare rune]

Because it’s a test and it’s probably been a model that was code marked

[fickle venture]

I think tte reason they didn't is because that A 4 AGENT

[abstract tundra]

i see

[fickle venture]

Just like grok heavy

[abstract tundra]

I know I've tried it out, but I didn't like it that much, so I was hoping the arena endpoint would be slightly better in some way

[fickle venture]

It pretty much won't improve it will just be the same

[exotic crest]

did the site just go down?

[echo aurora]

You'll want to try the steps in this article - https://help.arena.ai/articles/1645798556-lmarena-how-to-something-went-wrong-with-this-response-error-message

[echo aurora]

I don't think so...

What are you seeing?

[exotic crest]

its working now, but for about 5 minutes the site stopped loading. my browser was just infinetely loading

[quasi gyro]

If I clear the data, will all my chats be removed?

[echo aurora]

If you're not logged in, yes.

[quasi gyro]

I logged in
Ok

[echo aurora]

Glad to hear it's working again, keep me updated if things go sideways

[echo aurora]

Then yeah your chat history will be saved to that account

[mortal coyote]

gemini 3.1 coming soon ? on ARENA ?

[uneven glacier]

hlo

[echo aurora]

It's currently available.

[stray aspen]

how is grok 4.2 on fourth place

it sucks

[quasi gyro]

I did it but still have that issue

[uncut wind]

bro does anyone know why opus doesnt work

no matter how much i refresh

then it tells me ive used my limit try again in 40 minutes

after not answering my prompt

[echo aurora]

I'll respond in the thread you have going in #ask-here

[quasi gyro]

Bro i deleted data but still jave problem

[echo aurora]

I've responded in the #ask-here thread.

[quasi gyro]

Link won't open

[echo aurora]

Pinged you in the other channel, I'd like to keep #general open for general conversation and use the other channels for troubleshooting.

[fickle venture]

What did you expect form grok it always sucks

[toxic verge]

How this model is in the top tier list is beyond me

[toxic verge]

What’s chat arena ? Have you ever heard of it before ?

[proud bobcat]

no 5.3 codex yet

https://tenor.com/view/angry-gif-6891886407904187638

[mighty surge]

?

its already released like 2 weeks ago

[ashen oak]

Is the Gemini -3 pro image review not working on anything just for me today?
It keeps saying something went wrong on anything I sent there ?
Would love some help on that🙏

[echo aurora]

Would check the response that the bot provided you here: #1476278002626072586 message

[ashen oak]

Didn’t help

🙏🙏

[rare swallow]

Yeah, I'm also getting the same problem with Claude opus 4.6

[proud bobcat]

in arena

😭

[rare swallow]

I need Claude, like desperately rn, dumb Gemini deleted my code and Claude was the only model that actually understood the context

[ashen oak]

Tried what the bot said , plus , tried switching browsers or accounts didn’t help

[rare swallow]

I did too

Didn't help

[mighty surge]

oh sh*t mb

[fickle venture]

You can already use it on CLI agent for free there is limit but it's very far

[sage talon]

who will join the group call

[proud bobcat]

i know

but

just saying

[rare swallow]

@echo aurora please fix this

[echo aurora]

Unfortuantely, if the steps in the article didn't help, it's likely an issue that can't solved on the user's end. For this model you're likely seeing a rate limit causing this as it's a very popular model.

[echo aurora]

Would encourage you to read this message: #1417174113092374689 message

[rare swallow]

This didn't help, it just told me to do the same thing you always tell us

[echo aurora]

Those are the troubleshooting steps. Unfortunately, if they don't work, reporting the better information to our team is best next option.

[river orchid]

hello

[toxic verge]

The way to prevent it is to breakdown the requests in new chats

Or take breaks and not send as many requests within an hour, intervals. And that’s not even guaranteed.

[ocean vortex]

Yeah I tried it now and got the same result. Would probably help if you forwarded 'You've reached your rate limit...' errors from Google to the user with something like 'Model is at capacity. Please try again'. So they wouldn't waste their time trying different accounts or whatever

[sick mantle]

🇭 🇮

[echo aurora]

Agreed, there are improvements to be made with these error messages. This is on our to-do.

[mystic patio]

hi, when does the arena get updated to reflect the new elo of each model?

and who maintains the arena?

[toxic verge]

Once you get caught in this cycle you need to relogin clear browser

[light sleet]

Bro you might be a robot 😭

[surreal zephyr]

When is haiku 3 coming to arena 🤔

jokes aside maybe we could get codex spark, or mercury2

[light sleet]

Codex 5.3🥺🙏🏻

[surreal zephyr]

already is

[toxic verge]

Once that triggers you’re gonna hit them in every single new message

[light sleet]

In arena?

[golden ocean]

im a robot

[surreal zephyr]

why not just have the gemini to solve the capthas for you 🤔

[golden ocean]

e

why not have @surreal zephyr solve it for you 🤔

[surreal zephyr]

hi japancat

[golden ocean]

o/

[surreal zephyr]

https://cdn.discordapp.com/attachments/1446560503668146178/1474068789250228337/togif.gif

[golden ocean]

turn this into a gif too

[queen veldt]

Guys I'm building the supercomputer I can't tell you the details

[golden ocean]

true

[toxic verge]

Gemini is not in the top tier list Forsure

Maybe for images, I could see why. It’s up there in the leaderboard, but it definitely does not deserve to be in the top 5 position.

[toxic verge]

I don’t know if it’s capable of solving anything to be honest with you.

I’m really baffled at how high it sits at the leaderboard

[surreal zephyr]

have you tried 3.1?

it makes opus a joke

it even beats gpt 5.2

https://019c775e-143e-7242-92b4-0d936da78f8c.arena.site/

[toxic verge]

In code?

[surreal zephyr]

in thinking

in code (execution only) nothing gets close to codex 5.3 lol

like not even same leaderboard

[toxic verge]

This makes situation even more confusing

[surreal zephyr]

codex actually does what you ask for instead of doing whatever it wants

[toxic verge]

I think people are looking at the coding as like the ultimate metric

[surreal zephyr]

gemini is crazy smart but its lazy and doesnt care what you want it does whatever it prefers

opus memorized everything but its literally braindead if you give it a novel problem to solve

[cursive bough]

what you guys think is the best ai for coding like html or java

[surreal zephyr]

if you dont know what you are doing? gemini + codex
if you do, just codex suffices

[toxic verge]

This is what makes these benchmarks and the leaderboard are confusing

[surreal zephyr]

leaderboard is sponsored lol

and even then people vote only by looks

[toxic verge]

I don’t know what it is, dude

[surreal zephyr]

if it doesnt listen but still makes something pretty, people prefer that

[toxic verge]

Yeah

[surreal zephyr]

5.3 codex wont get too high on leaderboard, because it asks questions

lol

[undone saffron]

Are you sure about that?

[toxic verge]

But this could be a little biased

[wise spindle]

why does nano banana pro keeps saying error

[surreal zephyr]

opus is one of worst high-end models if you actually test it

:)

[mystic patio]

opus is really good imo

[toxic verge]

Wow

Crazy if that’s true

[surreal zephyr]

gemini 3.1 and gpt 5.2 both agree opus is sh*t

guess the models

and this

[toxic verge]

Dude, you know what it is

It’s probably that the compute

They give us much more water down versions

[surreal zephyr]

guess which one is opus, which one gpt, and which one gemini

[surreal zephyr]

opus is just braindead if you give it a novel task, and sonnet 4.6 was trained on deepseek 🤣

[toxic verge]

Not sure the really messed up one

Are u sure?

[surreal zephyr]

the really messed up is glm, but the bottom right white with keyboard going out of screen and logo sticking out of screen is opus

[surreal zephyr]

yeah and they got exposed its the other way around

[undone saffron]

Don't be so sure about that
For some things, gemini couldn't solve the problems of code, but opus could

[surreal zephyr]

heres gemini summing up opus

[mystic patio]

i used grok 4.2 to automate the scoring of a psychological test, the MMPI-2

[surreal zephyr]

also i had a bug that i spent 2 weeks 8 hours a day trying to fix with opus 4.6
gemini found it in 1 prompt

[mystic patio]

it was pretty good

[toxic verge]

This proves my theory

[surreal zephyr]

grok is pretty bad but sometimes gets to levels of 5.2 or opus

[hard quiver]

I'm having the same problem anyone else?

[surreal zephyr]

mainly when opus or 5.2 mess up

[toxic verge]

Let’s create more speculation than it solves anything

It turns out to be a popularity contest more than it is a capability

Which undervalued the capabilities because I’m some of these models shine in certain areas

[surreal zephyr]

heres an old model vs a 10x more expensive less than month old opus

[toxic verge]

Which sadly aren’t measured and how will we ever know those if we’re just focused on the standard

I don’t think this is something academic could solve

[surreal zephyr]

every time it hears car wash, opus says "drive", even if the scenario is inverted

[toxic verge]

It has to be somebody from the bottom with a fresh perspective

[surreal zephyr]

same for many other riddles

[toxic verge]

A really evaluation test needs to be from the people like relatable

Like in the real world where people struggle, and with what they struggle lol

I mean these benchmarks are cool for enthusiast and researchers

[hard quiver]

Gemini 3 pro-image is not generating anything the only output is "something went wrong with the response, please try again"

[toxic verge]

It would be meaningful if model evaluations captured the everyday experiences of regular people using Llm the ones without large platforms or influence and reflected their real frustrations. There should be a way to measure performance that highlights where models consistently struggle, not just where they excel on benchmarks.

And you know, it’s the most ironic thing as recent we haven’t heard much of the term “AGI” been thrown around lately. 😂

[surreal zephyr]

give original video ill show

[surreal zephyr]

remember gemini 3.1 is like the only good model from google (gemini 3.0 before all nerfs was good too)

[toxic verge]

That’s the movie it’s supposed to say

[surreal zephyr]

100 views 😭

theres no way it was in the training data

i doubt it can figure out based on literally snow

[toxic verge]

This is what I’m saying. It’s not like the model is stupid. When you nudge it.

[surreal zephyr]

It guessed based on just the text

[toxic verge]

Makes sense

[surreal zephyr]

yeah its THIS good

(pure new chat btw)

it can guess the movie just based on ww2+ snow

snow alone is not enough

[toxic verge]

Yeah

[surreal zephyr]

(without the video)

[toxic verge]

Well regardless then this would be edge case

And the world is full of edge cases

[surreal zephyr]

i made a fun experiment with gemini 3.0 few months ago

i had a pic of a house i took by accident

no landmarks ect, just a normal house/building

i put that into gemini

it found it to 5 metres

💀

it mightve been on google maps street view but still thats insane

[toxic verge]

Im not saying these models are outright stupid

I’m just saying, I don’t think that leaderboard accurately reflects. I don’t even know what I’m trying to say.

[surreal zephyr]

actually

not even need image xD

[toxic verge]

Actually, this is a good benchmark. We should try out with the other movies.

[surreal zephyr]

ai might be closest to "magic" we will ever get tbh

[long minnow]

and youtube might be the closest to a "time machine"

[toxic verge]

Nawh biology by far is far more mysterious and far more magical

[surreal zephyr]

mind reading is better

and mind reading without sensors is even crazier

[toxic verge]

That was great 😂😂😂

[toxic verge]

All of that stuff is crazy, dude

[surreal zephyr]

there are already ais that can read thoughts from mri scans, but good llms can almost do that without any scans or such

[toxic verge]

But yeah, I hear what you’re saying. Definitely for sure. It feels like magic.

[surreal zephyr]

ww2 + winter snow is NOT enough to guess it

[toxic verge]

I don’t think it’s magic. I just think that we’re really predictable.

[surreal zephyr]

it guesses by the way a human says it

like

human subconsciously spells it in a way that hints for the specific movie

and the llm's wages contain some of those patterns

[toxic verge]

We’d have to see behind the scenes to truly know

[surreal zephyr]

its more like

not the movie contains winter and ww2

its about why when asked to describe the movie

the human picked ww2 and winter

and not for example a plane

or a gun

[toxic verge]

You sound like a salesman

[surreal zephyr]

im js excited

ai is underrated

like

ai is basically a solution to every solveable problem

by definition

[toxic verge]

O.o

[surreal zephyr]

if there is any pattern, ai can be trained to find it

if a dog is smart enough to have a conversation

you can train ai to translate

by definition

thats how impressive it is

[toxic verge]

Yeah, but in the physical you have entropy or pure randomness

[surreal zephyr]

everything has randomness

but if a dog or a monkey can say (or even think) "give me food" or "i want go outside"

you can see that using ai

its not really feasible rn to do that but its very much possible

[wicked talon]

AI 🙂

[toxic verge]

I like the word you used earlier magic

[surreal zephyr]

if ai is not magic, then what is?

[wicked talon]

Its basically a database no?

It just finds info on a database and pieces it together

[surreal zephyr]

ai is a way to solve literally every solveable problem without knowing how

you can solve any problem by throwing money and compute at it

using ai

[wicked talon]

What's 0/0 then

[surreal zephyr]

lmao

thats easy

[wicked talon]

Tell then

[surreal zephyr]

when you get 0/0 you use le hospitals rule and simplify

[toxic verge]

I don’t think the AI is gonna be smarter than humans

[surreal zephyr]

lol

[surreal zephyr]

define smart

pattern recognition? then it already is MUCH smarter

[toxic verge]

Humans

[surreal zephyr]

oh like causing wars and killing millions because politican in other country insulted you?

[wicked talon]

0/0 is undefined

[surreal zephyr]

depending on context

[wicked talon]

It won't be in terms of memory storage

In terms of remembering things then yes

[toxic verge]

So is gonna be conscious or not

[surreal zephyr]

can you read? it depends on context

[wicked talon]

AI cannot be conscious

[surreal zephyr]

0/0 has multiple meanings

same as castle is chess move

[neon idol]

Hello!

[toxic verge]

Hmm

[surreal zephyr]

Humans cannot either :)

consciouseness is a made up thing

if it isnt, then define it?