general

[keen beacon]

Hi

[honest verge]

Gemini 3.1 today?

[analog canopy]

hi

[pale sonnet]

opus literally makes me feel like a genius

i just made a video that got 4k views in an hour with opus

[tender patio]

what kinda video

[north obsidian]

Can anyone generate video? In arena

I can't generate and I lose my uses

[brittle tapir]

nah they just released deepthink upgrade

it is crazy tho

[warm holly]

Wow, really? Can u share what u used

[echo aurora]

Video Arena looks to be working

[tender patio]

Is it better than 4.6 and does it forget stuff fast or no

[sly raven]

curious is it true now arena has a lot faster leaderboard update schedule? in the past often take weeks for next update

[tender patio]

Because each time i use gemini it will forget stuff super fast

[warm holly]

Just got a notification glm 5 out for pro usees

[proud bobcat]

Do you guys find it odd that there’s no benchmarks yet for Minimax m2.5

I haven’t seen ANYTHING

[stray aspen]

It sucks anyways

[proud bobcat]

You say everything sucks dude

[north obsidian]

No I can't yet it's always something went wrong 😭

[proud bobcat]

I need a reliable source

[stray aspen]

Claude opus 46 is great tho

[proud bobcat]

GPT 5.3 shadow dropped???

[left lodge]

But only for Ultra $200 plan

[proud bobcat]

Or is this older

[proud bobcat]

I

I know that

[north obsidian]

And I tried closing and cleaning cache and creating new chat but nothing

[proud bobcat]

Minimax is more of an agentic model

It’s not gonna be completely SOTA

[left lodge]

It is showing 5.2

[proud bobcat]

Ah

Apologies

[left lodge]

?

[robust sluice]

I didnt see any Gemini in battle today

[left lodge]

Hm

[proud bobcat]

The quality is worse than a CRT screen

😭

[left lodge]

Yeah lol

[robust sluice]

its more likes Flux arena

[left lodge]

Lemme send a better image

[proud bobcat]

Needs more info probably

Battle prioritizes lesser used models

[left lodge]

A lil better

[proud bobcat]

Oh wait did deepthink release?

????

New benchmarks

I remember a few months back there was a different benchmark score

[left lodge]

?

[proud bobcat]

Deepthink was a thing a few months back

[left lodge]

https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-deep-think/

[proud bobcat]

These are refreshed benchmarks

[proud bobcat]

Yeah so they did update it

Alrighty then

Still I gotta say not too impressive

Probably thinks for ages

[left lodge]

Its only for ultra user for some time I think

[late wolf]

what happens with image generation?

[proud bobcat]

Nodnod

[robust sluice]

I use battle and I always got Flux, this happen everyday but atleast I will see Gemini 2-3 times a day in battle right?

but today I didnt see any of it 2.5 or 3 nothing

[left lodge]

Yeah they released this like 10min ago

[robust sluice]

only Flux and reve

[proud bobcat]

Testing less used models

[worthy garnet]

What should we do

[proud bobcat]

Use side by side

Easy

[robust sluice]

I know

[left lodge]

I think there should be public list of which models are currently running in battle mode including codenames

[robust sluice]

I like to see droprate, funny to get only Flux there

just having fun

[left lodge]

@echo aurora Blog on this when?

[worthy garnet]

Now what should I do

[left lodge]

For what?

[worthy garnet]

I cannot understand what to do

I cannot Janet any video

[robust sluice]

its random error right ?

[left lodge]

See #announcements

[worthy garnet]

I saw

But what is the solution

[left lodge]

The Video Arena bot has been removed from the Arena Discord server.

Video Arena will still be available through the site and is unaffected by this change. This shift allows us to focus efforts into improving Video Arena with features and capabilities that aren't possible through a Discord bot.

[left lodge]

yo it’s mad easy 💀

1. go to arena.ai and LOG IN (has to be logged in)
2. switch to Video Arena (hit the lil film-reel icon / pick “Video”)
3. type your prompt (or upload 1 image if you wanna do image-to-video)
4. hit Enter → it cooks
5. wait (can take like ~10 mins), then you’ll get 2 vids back
6. watch both, vote which one’s better, and download if you want

btw vids are usually like 5–8 sec and you only get ~3 gens per 24h so don’t waste prompts lol

I had to generate the responses for these people 😭

[worthy garnet]

Bro talk kindly

[burnt sinew]

Is m2.5 any good

[left lodge]

Pineapple answer in #ask-here bro

[left lodge]

I am doing my lil test

Pineapple writing a lot ✍️

[echo aurora]

We've thought about this idea before. However, at this time we don't have plans to make a publicly available list of new/existing codenamed models. For non-codenamed models using the drop downs in Direct/Side by Side will give you that info. The hesitation to do this with codenamed models is because we wouldn't want to create incentive for users to spam Battles just to get access to a new model they're looking to try out.

Pineapple answer in ⁠ask-here bro
This message I think deserves more eyeballs, so I'm fine answering this one here.

[burnt sinew]

Pretty sure people will spam to try to find out the codenamed models no?

[left lodge]

Atleast list public models? Wait are all models in direct available for battle? I get it

I meant I asked a lot questions in #ask-here

[pseudo hemlock]

Hi

[lofty frigate]

Bro these violations im getting are ridiculous🤦‍♂️

[north obsidian]

Me too 😔

[echo aurora]

I'll get to them when I can

[echo aurora]

It'd likely increase more if they are shown.

[north obsidian]

I just wanted to make animes fights

[left lodge]

Hmm there has to be a solution

[frail stirrup]

https://x.com/Norwakar/status/2021985634530078809

Diwakar Ray Yadav (@Norwakar)

🚨MiniMax just dropped M2.5 and it's INSANE:
- 80.2% SWE-Bench (beats Claude Opus 4.6)
- 37% faster than M2.1
- $1/hr at 100 tokens/sec
- 10-20x cheaper than Opus/GPT-5/Gemini
- Stock pumped 11% instantly
Open source. The AI price war just escalated.

X · Feb 12, 2026, 4:31 PM

https://happycapy.ai/signup?invite_ref=xAhddntM

[lofty frigate]

Im just doing a harmful prompt and its getting violated, like i dont get it

[left lodge]

But they will only see the model codename after vote after which they can't use it? So surely they just send tests for models performance?

[lofty frigate]

Ridiculous

[north obsidian]

Yup

But the problem is that AI models terms

[lofty frigate]

Nah on other sites the prompt works

[lofty frigate]

On the same model

[north obsidian]

Ow

[wicked talon]

Which ai is better, Grok or Gemini (grok 2.0 Vs Gemini 3 pro preview)

[left lodge]

Where you getting grok 2 from??
Grok 4.1 is the latest

[north obsidian]

What intent?

[wicked talon]

Exactly

🙂

[burnt sinew]

I mean there is/was a bot that did exactly this as a purpose

[left lodge]

https://tenor.com/luAIsaGowfz.gif

[lofty frigate]

2 people comforting each other

[left lodge]

Both are bad use something else.
Grok can be useful for writing human like stuff but that's it , its mainly nice for chatting

[wicked talon]

Nah Gemini on top for general usage

I use Kimi for coding

[burnt sinew]

Gemini 3 pro preview

[wicked talon]

And basically everything else besides questions

[wicked talon]

W

[north obsidian]

I use Mistral Goat

[surreal zephyr]

LOL

minimax making opus a joke is crazy

[burnt sinew]

No way

[surreal zephyr]

and its 10x cheaper

[swift ermine]

Lol

[tacit temple]

I thought it was unlimited chat with AI

[swift ermine]

Glm 5 and minimax m2.5 is fire ngl

[golden ocean]

Large Language Model

[stray aspen]

glm 5 has a terrible vision though

its literally blind

[surreal zephyr]

even worse than opus/gpt/gemini?

[stray aspen]

wdym gemini

[echo aurora]

Sure but this doesn't mean that we're going to agree with it and create the same.

[stray aspen]

gemini has a great vision

[swift ermine]

? Huh glm 4.7's vision is perfect when i tested

[surreal zephyr]

it cant see gradients

gpt has better but still flawed

[remote vapor]

I don't care about geminiiiiii or glm or - or minimax-

I want qwen!

[stray aspen]

lmao

[rocky fractal]

@echo aurora i think arena needs some sort of way to evaluate agentic coding capabilities, its very clear that glm is a way better agent than gem 3 pro but the code arena itself doesn't really reflect that

[stray aspen]

great minimax 2.5 just used all the tokens for reasoning

why do chinese models overthink so much

[left lodge]

This is the result of what people choose

[whole dew]

Hi

[left lodge]

They can choose wrong

[surreal zephyr]

no its sponsored lmao

[left lodge]

Most people come here just to chat

[surreal zephyr]

theres no way opus gets even close to 5.3 codex

[left lodge]

They don't utilize the models full capabilities

[surreal zephyr]

and minimax is better than opus even

[left lodge]

That's why gpt 4o model was so high for so long

It was a great model but it was specifically great at conversation in its time which cause it's boom

[swift ermine]

Also cluade 4.6 thinking is same

[proud bobcat]

Minimax 2.5 new coding sota???

[stray aspen]

no

its claudius

jk its 5.3

[echo aurora]

evaluate agentic coding capabilities
This is something we're very interested in. The ideas are percolating

[surreal zephyr]

literally opus

[rocky fractal]

yes im not arguing that

[wicked sage]

grand left mauto 🎉 im so funny guys

[surreal zephyr]

5.3 thinks the least yet still the smartest lmao

[proud bobcat]

Right

[surreal zephyr]

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

[stray aspen]

5.3 is the greatest model

[left lodge]

100+pts below Claude-Opus-4.6 in agentic webdev tasks.
What? what is this based on?

[wicked sage]

anyways hi guys

[proud bobcat]

Claude, generate a script to tie up this man’s balls

[wicked sage]

gork: OK

[left lodge]

Don't say arena

[north obsidian]

Greatest diff from world

[rocky fractal]

looking forward to it! thanks

[surreal zephyr]

insane

[stray aspen]

glm-5 is fourth greatest model right now according to artificial analysis

[rocky fractal]

2nd place in agentic too

[echo aurora]

Sorry should have clarified, I meant more-so adding different ways for these kinds of evals. All I'm saying is agentic coding is something we're very interested in.

[north obsidian]

I thought they were clouds

[surreal zephyr]

waves

[pale obsidian]

what the hell man china is not slowing down https://fxtwitter.com/i/status/2021974501660274924

Ant Ling (@AntLingAGI)

🚀 Unveiling Ring-1T-2.5 The first hybrid linear-architecture 1T thinking model.
︀︀-Efficient: Hybrid linear breakthrough (10x lower memory)
︀︀-Gold Tier: IMO25 (35/42) & CMO25 (105/126)
︀︀-Agentic: Natively with Claude Code & OpenClaw
︀︀-Open SOTA: IMOAnswerBench，GAIA2-search & more!

**💬 14 🔁 40 ❤️ 264 👁️ 10.5K **

FxTwitter · Feb 12, 2026, 3:47 PM

[left lodge]

Yeah no this is so opposite to all other evaluations on glm5

For agentic tasks

[surreal zephyr]

glm>>>>>opus4.6

opus 4.6 is gpt 4o tier

[rocky fractal]

i very much trust the code arena ratings, i know how it works and how it differs from agentic coding, that's why i originally said that this doesn't reflect the agentic capabilities of the model

Like using glm in Kilo Code, its a miles better model than Gemini 3 Pro, where it gets stuck in deadloops and such

[left lodge]

Man gemini is not the model to compare cause of its instability consistency and reliability issues
Environments change the models performance, that is true

[rocky fractal]

take this as a soft reafirmation that its a great agent

[left lodge]

Gemini have its own benefits but it's not for coding or any tasks requiring very high precision and instruction following

[rocky fractal]

ay its the model's job to know how to deal in agentic envirionments, scaffolding and tools

[surreal zephyr]

ya ya

[rocky fractal]

id say gem 3 its still my fav model for front end designs

[surreal zephyr]

gem 3 is like opus 4.6 mini

[left lodge]

Try kimi k2.5 for frontend

[rocky fractal]

also great

[left lodge]

Thinking and non-thinking both

[surreal zephyr]

both 10000 times better than opus 4.6

[rocky fractal]

mini?

[surreal zephyr]

yeah gemini is like temu opus

just as dumb, and just as overthinking

but hallucinates even more

[rocky fractal]

about that too

its very solid

[left lodge]

Its literally the best at not hallucinating

[surreal zephyr]

60% is NOT sollid bro

[left lodge]

34%

[surreal zephyr]

thats glm

[north obsidian]

No way just 34

[surreal zephyr]

not opus

[left lodge]

Crazy I was shocked when I saw that

[surreal zephyr]

wait for 5.3 codex then xD

sub 20 🤔

nah but below 40% for sure

[rocky fractal]

5.2 is at 71% i don't see it doing better than glm

[north obsidian]

Yes adms

[echo aurora]

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

[left lodge]

:p

[rapid merlin]

https://giphy.com/gifs/news-demdebate-democratic-debate-f9eYHQ8RZ4zfc4unXx

[surreal zephyr]

5.2 is way worse

[echo aurora]

If we could avoid using the title text that'd be appreciated. That should be turned off so I'm not even sure how you're doing that

[rocky fractal]

https://x.com/OfficialLoganK/status/2021997983697711524?s=20

Logan Kilpatrick (@OfficialLoganK)

@rugubuilds @GoogleDeepMind we are very much so hill climbing on agentic coding, lots of progress and lots of work still to come : )

X · Feb 12, 2026, 5:21 PM

the salvation for gemini

[surreal zephyr]

4.6 🤣

[left lodge]

That's just heading

[cloud zinc]

3.1 is real guys there is a typo wtf

https://storage.googleapis.com/deepmind-media/gemini/gemini_3_deep_think_model_evaluation.pdf

[echo aurora]

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

[cloud zinc]

go here

[echo aurora]

Yeah that's what I mean by title text

[rapid merlin]

why is he thinking about what to say 😭

[rocky fractal]

thats just 3 dt

[north obsidian]

Hey man yesterday u sent that seedance ai video did u pay? Or free?

[left lodge]

I don't think we can disable that.

[surreal zephyr]

lol this is such bs, the "pro prev" is the current compressed version

[rapid merlin]

Me on a maths test

[north obsidian]

@echo aurora

[cloud zinc]

no look the page title

[rocky fractal]

OOOOOOOOOOOOOOOO i agree they should've named the new deepthink 3.1 tho, its confusing

[mortal vale]

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

[surreal zephyr]

yea

[surreal zephyr]

then why do #video-arena-1 channels exist and pollute the channel list?

bot

quota waster bot lol

[north obsidian]

Yes

[left lodge]

I think a timed mute would be appropriate

[surreal zephyr]

can we ban the quota waster bots fr

[frosty lava]

Guys, i actually did a test with 4-6 thinking with the same task but two different prompt, and it literally gave me a whole different result its really impressive

[surreal zephyr]

it got banned, and thats deserved. its a spambot that wastes quota/ddos

[left lodge]

What you tested for?

[mortal coyote]

bru

[surreal zephyr]

its overthinking so its rng

[left lodge]

Ocean by minimax 2.5
https://019c52e4-d3c5-7bc7-9264-bd93a98d1950.arena.site/

[frosty lava]

the task was : Create a futuristic ancient temple floating above volcanic lava, glowing runes, dramatic lighting, detailed high-quality-best 3D render

But in one of the two prompt i tried using rules, persona, and anti lazyness things, and honestly it really worked, cause the result with the prompt with rules is much better

[plucky sparrow]

wow, gemini-3-pro is outright dangerous for coding

I mean I guess they all are, but gemini-3-pro is really bad 😄

[north obsidian]

Send this here

[glad citrus]

Hi

[plucky sparrow]

it deleted a bunch of my data and didn't tell me on a rather simple migration task

[frosty lava]

Alright wait

[north obsidian]

K

[surreal zephyr]

as bad as opus

[glad citrus]

He

[red cliff]

Generete a video?

[plucky sparrow]

opus is better. at least 4.6 thinking

[surreal zephyr]

4.6 is way worse than 4.5 even

biggest disappointing model out of all in 2026

[frosty lava]

i went from this with the first one to this

when i say with the first one i mean with the prompt without persona, rules and everything

see the difference ?

[surreal zephyr]

give full prompt

[north obsidian]

Did u have any traumatic generation from it?

[frosty lava]

alright

[plucky sparrow]

what language are you using it for?

[frosty lava]

bruh my bad

i send you in dm

[surreal zephyr]

xd

[plucky sparrow]

what did it end up answering you? 😄

[frosty lava]

i sent you the prompt i used to get the good result

[surreal zephyr]

it MADE UP SHA KEY

[plucky sparrow]

so it decided not to be honest? 😄

[surreal zephyr]

thats like LEAKING WHOLE DATABASE

[surreal zephyr]

it decided to use base64, but SUGGESTED sha as alternative

[left lodge]

Omg 😭

[echo aurora]

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

[surreal zephyr]

(2nd looks insane on hdr but camera angle sucks)

[rocky fractal]

https://fxtwitter.com/MiniMax_AI/status/2021990549130281383?s=20

MiniMax (official) (@MiniMax_AI)

ohhh yesss👀

Quoting Okara (@askOkara)
︀
"Sir, MiniMax just dropped MiniMax M2.5 and it's on par with Opus 4.6 while being 20x cheaper"

**💬 6 🔁 12 ❤️ 268 👁️ 6.3K **

FxTwitter · Feb 12, 2026, 4:51 PM

[surreal zephyr]

"on par" ❌
way better ✅

[frosty lava]

im working on improving the prompt without causing hallucinations or regression in capabilities of the ai but its literally so hard

[rocky fractal]

for 20x less i can't ask for anything else

[surreal zephyr]

no the minimax is better than opus

[vagrant temple]

Has anyone tried Seedance 2.0?

[rocky fractal]

neck a neck

[surreal zephyr]

nah

[north obsidian]

Me

[surreal zephyr]

20x faster and 20x less tokens and less halucinations

[vagrant temple]

its just a benchmark bro...

[surreal zephyr]

opus sucks

[left lodge]

Someone things if a model can't do SHA256 so it's useless
So here you go

[rocky fractal]

sources?

[surreal zephyr]

5.1

[vagrant temple]

Maybe, but Minimax isn’t better

[hushed gyro]

Is m2.5 good

[opaque narwhal]

Does anyone know why we don't have quant versions of models on lm arena ?

Maybe that would help to know relative comparisons

[surreal zephyr]

ba4788b226aa8dc2e6dc74248bb9f618cfa8c959e0c26c147be48f6839a0b088

is correct

[left lodge]

I am showing the model names and prompt ,some dude just sends the cropped text and glazes gpt 😭

[surreal zephyr]

its bit better

[rocky fractal]

what are the sources of that tho

[surreal zephyr]

source try it

[rocky fractal]

of this

[left lodge]

Source: just true me

And it's literally a single point difference 😭

[frosty lava]

i tried both but sorry opus is better

[vagrant temple]

Wasn’t Minimax originally just a TTS?

[north obsidian]

Always this man

[surreal zephyr]

math vs hallucinations

[north obsidian]

Seems they work for it

[surreal zephyr]

@rocky fractal

[left lodge]

Wut this man 👨

[north obsidian]

Glazes gpt

[left lodge]

Me ??

Lol

[north obsidian]

No

[left lodge]

You messed up what you wanted to say xd

I get u

[rocky fractal]

hi

[north obsidian]

Yeah a bit lol

[north obsidian]

K

[rocky fractal]

oh damn

[rocky fractal]

try opus thinking at least

[vagrant temple]

What’s your experience using AI for R0bl0x development?

[frosty lava]

the problem is that ai are lazy, and always seek for the simplest way to do things even if that regress the quality of the work, that's why a good prompt focusing on anti lazyness can improve by alot the quality

[left lodge]

Minimax m2.5 💀
https://019c52f6-c050-7949-b9ac-06a2710f7506.arena.site/

[frosty lava]

wow on how many prompt ? and what's the prompt ?

[left lodge]

Crazy

[gloomy solstice]

hello

[rocky fractal]

4 minutes, still thinking

[vagrant temple]

hahah 4 isnt much bro

[rocky fractal]

so whos gonna check whats right?

[vagrant temple]

this looks bad

[left lodge]

I am sure it can do 1 prompt
I did use 2 cause in the first I asked for lightweight

[frosty lava]

okay thank you

[rocky fractal]

all wrong @surreal zephyr

[left lodge]

You are not seeing what I am seeing

This is bad??

[north obsidian]

Yeah lol

[left lodge]

Hello?

[frosty lava]

no its not bad its really good looking honestly

[north obsidian]

But the sun is in terraplanist mode

[frosty lava]

can you give me the first prompt you used for that ? i want to try something

[left lodge]

Uh yeah

[frosty lava]

thanks

[left lodge]

Mix these

Remove lightweight

[frosty lava]

okay thank you

[left lodge]

:>

[grand cliff]

Where's the Archives?

[mortal vale]

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

[spare path]

hello guys

[frosty lava]

i sent you a dm with my try can you give me your thought on it please also minimax m2.5

[burnt sinew]

Well obviously you wouldnt create a bot for that... I mean that the lack of a code named model list incentivises people to find them all

[proud bobcat]

whats next

[primal escarp]

Hi

[echo aurora]

It's in Search, more info on how to delete a chat session will help:https://help.arena.ai/articles/9130232616-how-to-delete-your-chat-sessions-and-data-from-lmarena

[proud bobcat]

gpt 5.3 mega ultra codex spark xhigh with sprinkles on top

[left lodge]

Another one prompt example from minimax m2.5
https://019c5315-37f0-74a4-8625-7bebec535025.arena.site/

[proud bobcat]

oh thats good

thats really good

[left lodge]

It runs so smooth too

[proud bobcat]

new SOTA?

[left lodge]

Lol maybe

[echo aurora]

I think it could cut both ways tbh. The current thinking though is if we shared a list of codenamed models it'd lead to more people spamming Battle to get the codenamed model.

[burnt sinew]

Or make codenamed models available in direct haha

[frosty lava]

I think that opus 4-6 thinking need a higher time limit for the thinking, it can think for alot of time streak but due to the limit on arena it just say something went wrong

[left lodge]

I think you guys should experiment,
Thats theoretical, maybe scenario is different practically cause there are not that much people using arena which have free time and brain power to do that

[burnt sinew]

Why was it decided to make deleting complicated btw? Why not just 2 buttons archive and delete

Is minimax 2.5 a thinking model btw

[humble summit]

Which AI is best for Coding in Lua

[frosty lava]

i hope the next best open source model will beat one of the big ai (gemini, gpt or claude), it would be fun

when i say beat one of them i mean the best one that they have released and not just a previous version

[left lodge]

Glm did in some domain and kimi k2.5 did in others

[frosty lava]

yes but that's not enough to actually say its really better you know

[left lodge]

And minimax will do in remaining lol

[frosty lava]

but i really hope this will happen

[left lodge]

Yeah but those are open labs too
They are already doing a lot , I happy with what they are shipping

Trying to make a model good at everything causes the model to be bad at everything

Specified models wins as much I know

Specific models also need less cost and compute

[crude lynx]

Ok

[jovial minnow]

how I can use Ai here?

[sleek phoenix]

you go to https://arena.ai and use it there

[mortal vale]

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

[hushed gyro]

@echo aurora Its clear that these errors aren't rate limits. I have just created a new fresh account and errors appear already.

[crystal mica]

could've you share source code and prompt?

[frosty lava]

its rate limit cause they also limit the time the ai can be doing a task it's not the limit of the account but just the limit of the thinking time

you can see it by pressing f12 and searching a bit

its literally limited

[hushed gyro]

I mean it's NB PRO...

[frosty lava]

what ?

[hushed gyro]

The errors are from Nano Banana Pro, so how is it related to thinking time

[ionic summit]

hi i have a project in arena and the ai never stop thinking what should i do

[frosty lava]

i don't know for your problem exactly, but i was talking about claude opus 4-6 thinking, he is literally thinking for alot of time before actually creating or modifying file and it almost always exceed the limit

[left lodge]

@frosty lava He is asking for the code and prompt

:>

[frosty lava]

oh yes of course

[echo aurora]

Yeah that's correct, the Something went wrong error can happen for various reasons.

[loud crag]

What level of the safety tolerance from 1-6 is the Nano Banana from LMArena using?

[left lodge]

9

[crystal mica]

infinity

i get errors literally in almost every prompt

[golden ocean]

real

[polar brook]

seedance 2.0 has officially dropped

https://console.byteplus.com/ark/region:ark+ap-southeast-1/experience/vision?modelId=seedance-2-0-260128&tab=GenVideo

[polar brook]

you can get access here

[cloud zinc]

dont work?

[north obsidian]

I made this on chatcut

It's a seedance 2 too

[cloud zinc]

why frames so bad

fps feels low

[north obsidian]

But in download it turned this

Cuz it just works on desktop

And I sent to my cell by Whatsapp 😞 so bad

[inner relic]

How we can use seedance 2?

It's no longer available for me on byte

[celest orchid]

💀

[north obsidian]

Chatcut on desktop and join in the discord for codes

U can generate 3 videos

Free

[rocky fractal]

just wait for minimax 2.5 that got 80 on swe

[north obsidian]

Check it

[celest orchid]

all open models not able to coding

[inner relic]

Yeah i will just wait

[celest orchid]

i dont want believe that open model

[ornate jay]

is that paid ai?

[north obsidian]

Nice prompt

[celest orchid]

no, its free

almost

[ornate jay]

tried making a chicken shawarma ad altough I always be getting the ai slop lol in it

that's Gemini Veo3 btw

[celest orchid]

lol

[left lodge]

I was correct, gpt-5.3-codex-spark is out lmao

[ornate jay]

have you tried making an ad in ai?

[celest orchid]

yes

in sora

its fun

[ornate jay]

isn't that too short?

like 5 seconds?

[celest orchid]

nah

[ornate jay]

oh you have paid

[proud bobcat]

such a stupid name

buddy

video arena is online

no more in discord

[north obsidian]

This was insane now

[north obsidian]

Now it comes fine

[undone saffron]

That's something easy to make
Looking [this video](#announcements message), opus 4.6 it's much better than that

[undone saffron]

https://www.youtube.com/shorts/4y7M-46QDe8

[urban mason]

hello

[compact jay]

Hi @echo aurora , sorry to disturb you! I just noticed that Gemini 3 Deep Think is out. I checked the presentation on the official Google DeepMind website and I was really surprised that Google is giving API access to such a powerful model. I’m really excited about it because it looks like a great model, possibly one of the strongest available right now. I was wondering if you could ask the devs to add it to LM Arena. I’d be really happy to try a model this powerful. Thanks!

[celest orchid]

#1372230675914031105

[proud bobcat]

yeah uhhhh minima m2.5 isnt sota

benchmarks are definitely fudged

[echo aurora]

Thank you for sharing! Would ask to make the request in #1372229840131985540 as this is the best place for that. Thanks in advance!

add it to LM Arena
Also note we're just Arena now

[echo aurora]

#1372229840131985540 is the channel for these kinds of requests btw.

[celest orchid]

oh sorry

[echo aurora]

No problem at all

[compact jay]

it's just because I used to call it lmarena before change lol.. bit thanks anyway lol 😃

[celest orchid]

gemini face reveal

[compact jay]

btw no need to do it

someone got the same idea

https://discord.com/channels/1340554757349179412/1471547369698099303

[polar brook]

Seedance 2.0 even makes sprite animations

[thorn vessel]

So you can't generate videos here anymore?

[echo aurora]

It's taking me awhile to get use to myself tbh. I'm stilling typing in l into my browser instead of a

[echo aurora]

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

[thin current]

@echo aurora Sorry for the tag, but why is the generation with Gemini having problems? It keeps freezing continuously

[echo aurora]

Check out the #1343291835845578853 channel. More specifically #1417174113092374689

[thin current]

Thank you

[nocturne stratus]

"A passenger airplane soaring high in the clear blue sky, sunlight reflecting on its wings, with fluffy clouds below, cinematic lighting, ultra-realistic details"

[still musk]

I WANT SEEDREAM 5.0

HELLOOO

[karmic vortex]

Hello brother

[still musk]

What's your favorite scary movie???

[steep glen]

Hey idk if this is a common bug, but whenever i try to have it generate a game. A shows me a “the connection was reset” thing on my labtop. But when i go on mh phone. It works fine.

[shrewd citrus]

the funny thing is that it’s easier to type l into the searchbar rather than a

[compact jay]

eheh yeah same here, lol. by the way great work made by the devs, the UI is much prettier now in my opinion

[compact jay]

is it really ai generated?

[shrewd citrus]

https://www.youtube.com/watch?v=NptAC_6J-ho

[shrewd citrus]

that whole song is ai generated

plus all the visuals

[cloud zinc]

yes

[proud bobcat]

@surreal zephyr gpt is lame at coding

[celest orchid]

gemini is naruto fan

[echo aurora]

Thank you for sharing! Thrilled to hear you're a fan of the change.

[light sleet]

If only Opus 4.6 in code arena was working 😔

[proud bobcat]

working forme

[north obsidian]

Where?

[light sleet]

it says something went wrong for me

All I want is image to code fr

[north obsidian]

Jimeng? But how did u sign in

[celest orchid]

what is ai

[shrewd citrus]

probably seedream

[unique fable]

I stop creating videos with AI for a month and now they closed the bot it can't be

[north obsidian]

Use in Arena

[echo aurora]

It's still available on the site, no changes there

[toxic verge]

Damn, if you think about it CapCut is in a really good position

[echo aurora]

It's working! The error rate is a bit higher than usual from all the demand, but would recommend these steps if you're getting an error.

[north obsidian]

Capcut?

[toxic verge]

Hopefully they don’t get hit with a cease and desist

ByteDance

[north obsidian]

Hmm

[toxic verge]

TikTok is a huge platform. CapCut is a huge video editing service

[north obsidian]

Yes I get it

[toxic verge]

They have tons of video data

Maybe even more than YouTube

[undone saffron]

I don't think so

[proud bobcat]

Yeah no I gotta say

Minimax M2.5 is just not that good

[shrewd citrus]

wait why tho

[keen beacon]

Hey guys, is there any estimate for when Gemini 3 Pro will go back to normal? After a while in the chat, it keeps generating replies.

Or have you already returned?

[karmic vortex]

Could you send me the promo code for making these kinds of videos?

[echo aurora]

You'll want to then follow the steps in If the problem continues in the same article.

[toxic verge]

https://youtu.be/cjIzGMN7Bdg?si=3YN6tdAp6kIxUCla

I love mini max❤️

[proud bobcat]

As much as it’s a good model in general

It’s just not that good for coding

[toxic verge]

I got accepted in there cpp

[proud bobcat]

I think the best workflow for vibe coding right now is a combo of Claude Opus 4.6 and Kimi K2.5

[toxic verge]

No doubt, but we know how fast everything moves those spots flexible

[surreal zephyr]

codex 5.3 made bootable os with ui in 1 prompt (C + assembly)

[shrewd citrus]

i wonder if codex could make exploits

like just for games

[surreal zephyr]

guess why it requires ID verification

(yes it can it made rdr2 mod menu)

[shrewd citrus]

cause it would probably refuse actual “real world” exploits

[surreal zephyr]

"proof of concept"

🥀

[inland barn]

Ok

Done

[gritty spindle]

36

[half mist]

What video generator is this?

[cloud zinc]

seedance 2.0

[north obsidian]

But where did u use itM

[cloud zinc]

jinmeng

[rugged abyss]

Gemini 3 Deep Think is kinda insane

[north obsidian]

How did u log in?

That Chinese tik tok

[cloud zinc]

by downloading it

[north obsidian]

Ok thnx

In cellphone?

[cloud zinc]

yes

[north obsidian]

Ok

I couldn't 😞 I can't create a China tik tok account

[ornate jay]

ain't no way this is a cap haha

[ornate jay]

and has limits too now

[ornate jay]

codex is not free though right?

[surreal zephyr]

ofc its not

[ornate jay]

yeah can't imagine

so far no bugs with your programm?

[fiery gull]

Deepseek v4 better that glm 5?

[echo aurora]

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

[sand matrix]

I can generate videos. What's wrong

[echo aurora]

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

[sand matrix]

So I don't i create a video

[echo aurora]

Correct, it's still available on the site.

[sand matrix]

How do I generate videos

[echo aurora]

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

[echo aurora]

This help center article should help

[slate nimbus]

hi frens

[echo aurora]

howdy

[rose sky]

Which track you guys like most? All of them I created myself

[mortal vale]

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

[livid jacinth]

Why pls

[stoic zodiac]

hello!

[frosty lava]

someone have a fix for "something went wrong with this response" i already tried to delete cookies, refresh, log out, change browser, and it doesn't fix anything, this happen with claude opus 4-6 thinking almost everytime.

and it does not happen with other models

[echo aurora]

The announcement explains why.

[echo aurora]

I like oblivion the most tbh

[loud verge]

Guys. Have y'all heard of liberty ai?

[echo aurora]

Also, TIL you can play multiple audio tracks in discord at the same time.

[echo aurora]

Can't say I have, nope.

[loud verge]

https://lbrty.ai/

I just discovered this on reels.

[junior spoke]

What's happening with nano banana pro

[echo aurora]

There are going to be some higher than usual error rates with high demand models. Would recommend this section of the help center article if it continues.

[loud verge]

Grok 4 search never works.

Even on incognito. Can't be just me.

[junior spoke]

Nano banana pro is giving inconsistent results

[echo aurora]

We've been hearing some reports of this happening, if you're able to share examples in #1470529575367868498 that'd be helpful.

[junior spoke]

Ok

[burnt sinew]

@echo aurora i was wondering, how does battle mode choose what models to pick? does it take into account how good the models are? like will better models appear more?

[loud verge]

Matchmaking

Or is it random?

[burnt sinew]

i feel that it is pretty useless to compare 2 bad models

its why i am asking

what i care about and want to vote for are for which model is the best

[loud verge]

It feels random.

[burnt sinew]

not which is least bad

[loud verge]

Cuz no way amazon nova is anywhere near gpt 5.2

[burnt sinew]

yea its trash

[loud verge]

It's fine in the sense it complies without safety bs.

[fathom kayak]

Any ai website can make images to videos

[junior spoke]

Ye battle should be 2 of the best models already not random models

[loud verge]

You can do it yourself by selecting versus mode.

[proud bobcat]

The point of battle is to give smaller models a chance to be scored

[junior spoke]

Oh fr

[proud bobcat]

Just use side by side

[loud verge]

Battle is to give new competitors a chance to be in spotlight.

[gleaming roost]

[junior spoke]

Just checked i cant

[loud verge]

Are you sure?

[junior spoke]

Ye it only works with images not videos

[loud verge]

Oh. That.

[junior spoke]

Ye

[frosty lava]

thank you, but since its a model that really take a very long time to think before creating and modifying file i think it's due to that, and it may reach a potential hardcoded limit and that can cause the error.

when given a medium or hard difficulty task it will think too long

and may exceed a limit then cause the error

[echo aurora]

I think we have an interesting blog post answering this, let me see if I can find it (or if I'm imagining this)

[echo aurora]

Yeah this is true. Would note that just today we put out a patch where before models would error out if the response/thinking went past 6 minutes. The patch today raised this to 10 minutes. Unfortunately, for now this the highest we can raise it without a large overhaul.

[frosty lava]

Oh so you increased it already ? thank you i didn't know that, its very probably due to the limit, that'll explain why it happen more on the thinking version.

Thinking a longer time will probably be on every SOTA model that's coming, so this error will probably happen with the next one too

[echo aurora]

Yeah fix went out a couple of hours ago.

[placid flame]

@heavy latch Note that Video Arena has been removed from the server. More information can be found in this announcement. Also, this is an English only server.

[bronze olive]

hello

[proud bobcat]

Anyone know what Aurora Alpha is on openrouter

It’s REALLY FAST but it’s knowledge cutoff is June 2023

[plucky whale]

im still waiting for the deepseek r2

[undone saffron]

The term “rate limit” is being confused with problems such as this

It's not necessarily a rate limit, it could be other errors, such as these

[fossil heart]

#general

Hai

[young blade]

Yes

[toxic verge]

Which one do you guys prefer?

[north obsidian]

The last in right

[toxic verge]

Bottom or top?

Im experimenting with color pallets and grading

[stray aspen]

seedance 2,0 is so good

[wicked talon]

Bruh

Kimi 2.5 Vs Gemini 3 pro

[tawdry coyote]

what is bro yapping about

[undone saffron]

https://klipy.com/gifs/elonmusk-R4M

[toxic verge]

I really dislike anthropic I never use Claude ever

I think they’re super fake the whole team at anthropic

[wicked talon]

Claude worst AI oat

Defo

[toxic verge]

I don’t think so. I think that it’s a very capable model a lot of people like it and a lot of developers, especially so I’m not dogging on the actual model.

[wicked talon]

Nah Gemini better

[toxic verge]

Gemini low-key is has something wrong with it

[toxic verge]

With memory or whatever the hell there’s something not right about that model

[wicked talon]

Memories fine in my opinion

[obtuse smelt]

yeah gemini is better

[toxic verge]

Hmm

Maybe it’s me maybe there’s something wrong with the app

It blanks out and just forgets context completely randomly

[wicked talon]

Not for me tbf

[toxic verge]

Wtf look at this lol

[obtuse smelt]

is like super sentai lol

[wicked talon]

Lmao

[obtuse smelt]

https://tenor.com/view/go-onger-go-onger-super-sentai-engine-sentai-henshin-gif-6658737206021766588

is same colour right ?

[wicked talon]

Lol

Battlefild

[toxic verge]

This is Gemini pro but I guess you can’t really do it because this might be challenging for other models to do also so

It was close

[frigid carbon]

Donald trump dancing Lebanese dabke on a wwe ring

[wicked talon]

You'd be lucky

America would sue Google

[toxic verge]

😂

You know that there is it’s called an alliance

[obtuse smelt]

what!? mark meet AI lol

[toxic verge]

I like Japanese approach to AI

https://youtu.be/Z_ZXSUxQF-Y?si=sZSIFg6r2VPFoG0E

[obtuse smelt]

lol

[tough fox]

Reh

[toxic verge]

https://fpf.org/blog/understanding-japans-ai-promotion-act-an-innovation-first-blueprint-for-ai-regulation/

https://thediplomat.com/2025/12/japans-long-return-to-artificial-intelligence/

[obtuse smelt]

japan followed AI too lol

[toxic verge]

I just learned about this yesterday

Very cool very interesting

Completely different approach than what we’re doing here in the west

[obtuse smelt]

beefore they hate wtih AI but they helped AI in japan

[toxic verge]

https://www.shinkalabs.com/blog/230718-japan-ai-history

[obtuse smelt]

Japan is good country and very advanced

[toxic verge]

It’s a very strange country

In many areas, it is advanced in many other areas it’s not which is so strange

https://youtu.be/Q3pjnp2ojig?si=bSUWU6wgQU5wgGay

[obtuse smelt]

right is like year 2050

[toxic verge]

No, I’m just too lazy to write out what I’m trying to say, but these videos will do a better job than I ever could

Like I said, I just recently learned about this

[obtuse smelt]

lol, yeah is very smart

[toxic verge]

Crazy some countries have no AI no tech sectors no nothing

And another country develop these like it’s nothing

[obtuse smelt]

yeah right, is japan time is 12:56

in japan Even small children ask to earn money to work

[toxic verge]

Yeah, they’re a really hard-working. People actually watched another documentary just last night.

https://youtu.be/pQvbqzp8Km0?si=4f8KcMLYTk7er2hN

This was an eye-opener for me

It’s so different from how business is done in America

[mild knot]

Hey guys

[undone saffron]

#general message

[bleak lake]

Who's aurora alpha in open router?

Pony alpha was glm 5 btw

[obtuse smelt]

i use pony too before gemini released

[bleak lake]

Aurora alpha seems like a deepseek model

[obtuse smelt]

yep

[left lodge]

It doesn't seem much good. I think it's a light model

[bleak lake]

Definitely

It even said sorry I can't tell that , when I asked are you any good in creative writing

then it answered when I started a new session

[bleak lake]

I’m an OpenAI‑based language model (the GPT‑4 architecture).

It said this😭

no wonder ts didn't seem much good

[left lodge]

Yeah its just hallucination it from its training data

[bleak lake]

Real

its not something good either way

Not total trash as well just that

[gloomy onyx]

not sketchy at all

[cobalt laurel]

Bcbv

Jchff

JH fdddg

Hfhg

Jddg

Open access

[burnt sinew]

any update on this?

[forest helm]

#video-arena-2

#1397655624103493813

#1397655624103493813

[sand furnace]

Wear man city cloth and keep dancing

[toxic verge]

Pretty good

[restive plaza]

Hey someone help anyone here from usa

[burnt sinew]

@echo aurora you own [e-mail]??

[undone saffron]

https://arena.ai/?chat-modality=video

https://arena.ai/?chat-modality=video

[burnt sinew]

yo try the others ones too 😄

bottom right

[burnt sinew]

how are you so good at it lol

well

[undone saffron]

Only move the mouse?

[burnt sinew]

18 minutes

maybe not

[burnt sinew]

yeah but its a lot of ball pushing

you like it though?

[undone saffron]

You need to add more
Because this is the limit

[burnt sinew]

oh

btw

if you want you can click space it turns to sandbox

[undone saffron]

Top ten: Best offline games I've played

[burnt sinew]

im cooked i have an error i cant even fix with opus 4.6 thinking

[strange sluice]

great idea

[burnt sinew]

😲

🤯

[burnt sinew]

log out

delete cookies cache

[undone saffron]

That's not the problem
Btw, that's my video

[burnt sinew]

OH

you impersonated pinepapple? lol

[undone saffron]

Something like that

[undone saffron]

And that email is generic

[burnt sinew]

yea no but you put way too much effort into that 😭

[undone saffron]

Do you think?

[burnt sinew]

crazy

[undone saffron]

Ye

[undone saffron]

Well, good game
10/10

[burnt sinew]

ive never seen anyone do that to it

[undone saffron]

#ai-memes message

[cunning snow]

But how you made

[burnt sinew]

yeah like what

[undone saffron]

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

[reef moat]

/image to vedio

[floral urchin]

is something up with nano banana? 👀 results look of a very different (& much lower) quality than normally

[astral helm]

is NB pro 1k gonna be back anytime soon ?

[undone saffron]

https://arena.ai/?chat-modality=video

[grim scarab]

i still get the "Something went wrong with this response, please try again." error even when clearing cookies

Only with some models

[undone saffron]

You did [this](#general message)?

[grim scarab]

Jup, just did it

i could try it again perhaps

[undone saffron]

Can you check the console in Network tab?

[grim scarab]

Ofcourse!

[grim scarab]

It says status code okay for everything, everything is greeeen

[undone saffron]

What says the response?

[grim scarab]

Its all just status ok

[undone saffron]

So there's no response?
Like this?

[grim scarab]

OOHH im so so so sorry yeah thats correct

[undone saffron]

In that case, clearing your cache and cookies will not work
Report all of that in #1417174113092374689

[undone saffron]

Try another browser
If the problem persists, it is not the browser but the website

[grim scarab]

Alright ill give that a try

a3:"prompt is too long: 216099 tokens > 200000 maximum"

My prompt is only like 6 words lol

@undone saffron

[undone saffron]

Let me guess
GPT?

[grim scarab]

claude-opus-4-6

almost ANY Model does this

Gemini-3-pro doesnt

grok also does it

[undone saffron]

If you've been talking to the AI for a while, that message is normal
#general message

[undone saffron]

I know, because Gemini has a lower limit on tokens

[grim scarab]

aaah

dang

so starting a new chat should work?

im guessing there is no way to continue in this one huh?

[undone saffron]

You have two options

1. Stay in the same chat so that the AI you use has all the context
   Use AIs that have fewer token limitations (GPT 5.2 or Gemini 3 pro)
2. Start a new chat

[golden ocean]

https://media.discordapp.net/attachments/1006341705924087919/1200758925205708870/caption-2-2.gif

[grim scarab]

just started a new one 🙂 thanks for the help dude!! really appreciate it :)!

[boreal roost]

#video-arena-2

[undone saffron]

https://arena.ai/?chat-modality=video

https://arena.ai/?chat-modality=video

[golden ocean]

https://cdn.discordapp.com/attachments/1276918863975682079/1299435078153273435/attachment.gif

[left lodge]

🤣

<@&1349916362595635286>
Can you delete the message please

[marble otter]

i think claude opus 4.6 is down

i refreshed even changed chat but it just something went wrong with this response or just endlessly generating.

[left lodge]

Sir video generation is now only available on arena website at https://arena.ai/?chat-modality=video

I suggest to not share prompt and photos of your loved ones here.

If you want to just share prompts to help other people, you can do that here [#share-prompts]

[left lodge]

Both thinking and non thinking are working for me

[frail oriole]

why i cannot generate video ???

[left lodge]

@frail oriole

[tawdry robin]

can the motivational got monitize

[fossil heart]

26

[craggy owl]

hi

[tulip kindle]

Everyone know how fix captcha after every massage?

[gray oyster]

I stopped getting them

[sturdy mica]

new minimax model seems promising!

[gray oyster]

VERY promising 😢

[tulip kindle]

It's came again

[sturdy mica]

it hasn't stopped and won't stop

i can't stop it

its been doing this for 5 minutes

poor thing

[fleet steeple]

#video-arena-1

[gray oyster]

Hmm, also there's a weird issue where Gemini models feasibly stop understanding out to use the tools. It just spits out its thinking then deploys the project

Or gets confused because it thinks it edited the file then it checks it but then it spits out thinking again of changing it then it loops

[sturdy mica]

almost there bro

[full heron]

Cannot generate video

[sturdy mica]

so many people in this server have no clue whats going on

[sturdy mica]

uhuh wait 24 hours

[full heron]

Not allowed bro

I don't know why

Not about wait 24 hour

Check out

[sturdy mica]

no

[full heron]

you don have permision ???

[surreal zephyr]

they disabled it

[sturdy mica]

scroll down

[surreal zephyr]

because there are spambots that are made to waste quota

[sturdy mica]

and read

[sturdy mica]

really?

do you have a screenshot of that

[surreal zephyr]

yeah like each hour one gets banned

[sturdy mica]

i want to see

[full heron]

why the other can do ?

[sturdy mica]

show me

[surreal zephyr]

like this

[sturdy mica]

dude you need to go to the website

people here need to learn how to read

[surreal zephyr]

are you the spambots owner?

🤔

[sturdy mica]

oh i didnt think they were bots

i thought they were random braindead people

only joining to make videos

some random 50 year old

who needs an ad for his jewlery buisness

[surreal zephyr]

yeah random people create discord accounts then join this server under 10 seconds and paste ai generated prompt in wrong channel 😭

[gray oyster]

Ikr

[sturdy mica]

yeah it gets pretty old

[full heron]

that same thing have not done , never done only like that!! i a hour need to wait it but not done any more

[sturdy mica]

what kind of ancient language is this

[full heron]

sorrry that mean i need a hour to wait this and not yet ready until now

wkwkwwkwwkwkwk

[toxic verge]

https://youtu.be/lNNH-Ox_r04?si=roZKz90ELAGsMczj

[surreal zephyr]

claude literally sucks

lol

as bad as gemini 💔

[sturdy mica]

what are the best models

codex 5.3 ofc

and then uhhh

idk

[toxic verge]

I’m glad you can relate and understand

[sturdy mica]

@surreal zephyr

[toxic verge]

Best model in what category, coding?

Or just a general purpose?

[sturdy mica]

coding yes

and general purpose

two lineups

[surreal zephyr]

5.3c then minimax 2.5 or 5.2h unsure

[toxic verge]

Do we judge the metrics on popularity?

[surreal zephyr]

opus 4.6 is creative but braindead

[surreal zephyr]

by how much it hallucinates and how good it finds bugs without creating new ones

[toxic verge]

Well, then, if it’s not popularity, how the hell we know what the best model is?

Let me go look at the leaderboard real quick

[surreal zephyr]

lb is sponsored, and even then, people vote only by how the thing looks

[frosty lava]

its in battle mode when you don't know name model so you judge by result only

[surreal zephyr]

someone here made a discord copy with opus, and after 10 mins, someone hacked it with codex 🤣

[toxic verge]

Damn Claude is killing it

[surreal zephyr]

claude is like apple

overrated

also 5.3 is not api unlocked so its not on lb

[toxic verge]

Well, that’s what I’m saying if it’s not popularity then what the hell and how are we gonna get the best model?

We don’t have the metrics here to evaluate something like that

[surreal zephyr]

because if you let opus fix bug, it will delete your whole database

and codex will fix it

thats the difference

[toxic verge]

We need a new in the middle evaluation

Something unconventional

[surreal zephyr]

waterbench

[frosty lava]

what do you mean the evaluation is based on the user feedback just like i said some second ago

[toxic verge]

There needs to be like a stupidity benchmark

[surreal zephyr]

opus could never

[frosty lava]

and we don't know the model name

[surreal zephyr]

halluciantion vs understanding bench:

[toxic verge]

I’m not sure what I mean, but I’ve been thinking about it for a while

There has to be a way for people to be able to judge the model

[surreal zephyr]

opus 4.6

waterbench:

[toxic verge]

On its weaknesses

Common sense stuff lol

[surreal zephyr]

yeah its weakness - halluciantions and dementia

[toxic verge]

That’s what I mean needs to be away for people to accurately judge that too

[sturdy mica]

minimax 2.5 ?????

[surreal zephyr]

and costing 30x the price per token as minimax 2.5

[sturdy mica]

really???

minimax sucks

[toxic verge]

I like mini max

[frosty lava]

if the model hallucinate i mean you can see it, when it work and it'll have consequences on the final result so there's no need to add another test

[toxic verge]

Straightforward

[surreal zephyr]

exacly - hallucinations make models unusable

[light sleet]

Will Gemini 3 deep think be added in arena

[toxic verge]

I don’t wanna sound like a broken record, so I’m just gonna let it go

[sturdy mica]

WHAT

[toxic verge]

People only want to see and hear good stuff about ai

[surreal zephyr]

5.2h damn that lava is crazy

https://arena.ai/pl/c/019c5302-89e5-7be0-ba3a-ca2ee1fca204

[frosty lava]

the waterbenchmark is really wild but i don't know if its accurate did you used the same prompt everywhere ?

[surreal zephyr]

the prompt was very vauge so the models could choose

opus and codex decided to write custom shaders

but opus failed

[light sleet]

Will Gemini 3 deep think be added in arena

[frosty lava]

yes you have to "force" it to choose the best option

[surreal zephyr]

(the opus water looks 2d but its worse on performance as the codex water)

[paper island]

is anyone else getting the 'something went wrong' error randomly during generation on claude 4.5 opus?

[surreal zephyr]

or maybe be like codex and have experience to CHOOSE THE BETTER ONE?

like thats what the bench is about

[frosty lava]

we can do a better comparison when both of them will be on the arena

[surreal zephyr]

they wont be btw

codex requires ID verification due to hacking concerns

[frosty lava]

they will at some point release the 5.3 with an api

[toxic verge]

What’s funny lmarena got a bunch of money from a company called light speed venture