general

[wicked sage]

thanks trey

[stray aspen]

it sucks

[echo aurora]

Time moved forward

[stray aspen]

ill just stick to claudius

[wicked sage]

claudius

[modest prism]

Is Gemini 3.1 good

[wicked sage]

best name oat

[lost patrol]

i bet 3.5 earliest on google I/O. So in on May 19th

[thorny schooner]

Going to be real though i keep forgetting how much direct chat get get so disrupted by random competitive mode it's crazy

[sick mantle]

Pineapple WHEN IS SEEDANCE 2.0 COMEING

[stone tundra]

why si 3.1

so low

[modest prism]

It is AGI

[inner relic]

huh

bro

[echo aurora]

Sorry going to have to give the boring answer -> I won't be able to share details about what new models or features are upcoming until we're ready to share more. Would recommend to keep an eye on our announcement channel.

[cloud zinc]

80.6 on swe

[loud verge]

Thanks socky

[sick mantle]

Pineapple last queston can u check my dms

[stray aspen]

is the gemini on lm arena high thinking

[errant zodiac]

Why do I not have permissions to use the video arena channels? For some reason, Arena won't let me use those channels.

[cunning birch]

Gemini pro is just so dumb... More stupid than GLM

[sick mantle]

Its deleted.

[rigid holly]

I dont mean to be a nag, but its been like 2 days and yet the thinking model for sonnet 4.6 still did not come out while usually the previous ones came out withing the hour

Is it like a whole intentional thing or what @echo aurora

[echo sinew]

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

[stray aspen]

nah

[inner relic]

what are ya yapping about, this thing just released

yes

[stray aspen]

claude is better

[inner relic]

it created better

building

[cunning birch]

At all !!!

[modest prism]

Because sonnet 4.6 thinks for so long. The API cost would be too high

[echo aurora]

Yeah I will when I get to it.

[errant zodiac]

What does that mean? The channel is deleted? Or my permission is deleted? And if my permission is deleted - why?

[celest orchid]

and cheaper)

[golden ocean]

existensial crisis

[sick mantle]

The bot its gone

[echo aurora]

The bot was removed from the server.

[rigid holly]

Ah. Ok. Atleast i got a answer.

[echo aurora]

Moving Video Arena off of Discord and just onto the site.

[inner relic]

Okay it's your opinion

[errant zodiac]

Oh. Does Arena plan to add it back?

Ah. I understand.

[echo aurora]

The request for thinking versions has been raised. There are various reasons why we will or won't have specific models added.

[stray aspen]

bro this thing has early deepseek reasoning times

takes forever

[echo aurora]

No, not at this time. We're too limited with the bot. Unable to add new functionality to Video Arena so it's best we have it just on the site so we can focus efforts.

[stray aspen]

why did bro send his pfp

[thorny schooner]

You did see my PIN right in feedback ( just want to make sure)

[remote vapor]

why send fake chicken.,.,

[errant zodiac]

Thanks for replying and letting me know.

[rugged abyss]

Really interest in coding performance in real tasks

[sick mantle]

ai chicken

[stray aspen]

that chiken aint real

[arctic basin]

with video removal, new users intro process to this discord is perma locked on "Generate a video"

[remote vapor]

yes exactly

[sick mantle]

We got ai chicken before gta 6

[remote vapor]

yea fair-

[stray aspen]

<@&1349916362595635286>

[remote vapor]

oh come on...

[sick mantle]

Fake

[remote vapor]

just like chicken?

[stray aspen]

is anyone even still falling for that scam

[celest orchid]

Thank You

[sick mantle]

Ofc he is

[proud bobcat]

Oh shi

Gemini 3.1 pro

[echo aurora]

Tried to combine both our pfp

[proud bobcat]

It’s not supposed to be groundbreaking

[sick mantle]

YOOOO

[stray aspen]

this coding is terrible

[sick mantle]

SICK

[proud bobcat]

It’s a refresh

[stray aspen]

theres no order

gemini 3.1 sucks

[proud bobcat]

Hold on let me check this out

[sick mantle]

I love it!

[proud bobcat]

Let me give it a pretty simple task

[remote vapor]

why peeps posting slop?

[proud bobcat]

I think people forget video arena moved

[cunning birch]

Same problem 😂 😂

[proud bobcat]

Yeah Gemini code was never good

Me and my friend tested it out and it always was just messy

Did it work? Yeah

But it’s still ahh

[proud bobcat]

HELL NAH

HELLLLLLL NAHHH

[sick mantle]

But pineapple remove limts

[proud bobcat]

I DONT BELIEVE THAT AT ALL

[sonic swallow]

/vídeo

[fickle venture]

LMAO GEMINI 3.1 IS GOOD

[stray aspen]

seems like it bunched up a lot of code in a single line to reduce the total lines by 700

[cunning birch]

I thought Google will destroy all her competitors..
Now the competitors are safe 😆 😆 😆

[fickle venture]

It will get nerfed tomorrow

[inner relic]

there's a thinking level

medium

I think yall using high thinking

[cunning birch]

It is terrible today what about tomorrow

[fickle venture]

Probably more terrible

[proud bobcat]

Video arena is moved to the website

[lost patrol]

source
https://x.com/ArtificialAnlys/status/2024518545510662602

Artificial Analysis (@ArtificialAnlys)

Google is once again the leader in AI: Gemini 3.1 Pro Preview leads the Artificial Analysis Intelligence Index, 4 points ahead of Claude Opus 4.6 while costing less than half as much to run

@GoogleDeepMind gave us pre-release access to Gemini 3.1 Pro Preview. It leads 6 of the

X · Feb 19, 2026, 4:16 PM

[proud bobcat]

Yeah I know it’s from artificial analysis

Yeah no this is ridiculous

It’s a literal refresh model

INSANELY slow too

[shrewd citrus]

yeah lol like it should’ve achieved 1st place on the arena leaderboard

[proud bobcat]

Depends

[shrewd citrus]

with those “insane” benchmark results

[thorny schooner]

Well from what I'm hearing I guess it was a good thing I did not use that most recent AI the granted it probably wouldn't had help right now with the compares and stuff so still funny to hear some of the complaining ( cuz well I mean I was So nice to see I'm not the only one)

[proud bobcat]

GPT 5.2 isn’t a refresh

[lost patrol]

hallucination gone back it seams

[proud bobcat]

That’s a whole ass new architecture

“it actually is” 5.1 was a refresh

5.2 was completely diff

Yeah

[proud bobcat]

Hold let me check this out

How is Gemini 3.1 pro number 4 in speed this is slow as hell

Number one in coding?

[lost patrol]

at least at AA 😉

[proud bobcat]

Nahhh

[harsh flume]

we at a point of dinimishing returns. More compute, better inference, better artificial data, but still same paradigms

There wont be another GPT3 moment in a long time

[low patrol]

W GEMINI in the chat

spam w gemini

[celest orchid]

nice

[low patrol]

use in aistudio

[royal sail]

hopefully this shows in usage

[lost patrol]

yeah

[royal sail]

older models hallucinated all the damn time

[proud bobcat]

Wow this is slow dude

Holy hell

[harsh flume]

but thats nitpicking, in general compute is increasing at a linear pace

Opus 4.6 wasn't a gamechanger. This is what I mean, diminishing returns

[plucky sparrow]

is this good? 🤔 doesn't seem that impressive to me:

[sick mantle]

Yea

[lost patrol]

lets see, how it's on simple bench 😄

[stray aspen]

average

[echo aurora]

Since it's going to be a new/high demand models there errors may happen, team is monitoring this

[stray aspen]

they migrated from the video arena channels

[echo aurora]

I think it's people trying to use Video Arena

[sick mantle]

@echo aurora I wanna use sora on arena but it chooses for me random model please fix

[royal sail]

I think we've been at a point of diminishing returns for a bit, but improvements are definitely still meaningful

[celest orchid]

WOW

[echo aurora]

Isn't broken, that's intended

[low patrol]

safety censor has reduced

in gemini website and aistudio

[stray aspen]

is this lmarena or ai studio

[proud bobcat]

3.1 PRO PASSED THE CAR WASH TEST

YESSSSS

[stray aspen]

lmarena or studio

[proud bobcat]

Arena

[low patrol]

why are you asking? is studio worse?

[stray aspen]

yes

it gave me a non working script

lmarena gave me one that works

[harsh flume]

There's a chance compute is too high, I predict there'll be a inflexion point of algorithmic paradigm shift way more radical than what deepseek did and we get to a way higher joules efficiency

[sick mantle]

I know i just want side by side video

[quartz light]

my screenshot btw

:(

[harsh flume]

I don't mean compute is high in an absolute sense, rather energy output towards compute is too high

[surreal zephyr]

ok so gemini 3.1 pro is comparable to 4.6 sonnet nonthinking

[harsh flume]

The human brain is enough evidence of that

[surreal zephyr]

but much worse than sonnet high or opus high

[low patrol]

i thought it is better than them

in leaderboard artifical analysis

[proud bobcat]

Artificial analysis is benchmarked

[low patrol]

what does that mean

[patent bane]

thought all the models get it?

[surreal zephyr]

not even close

[frosty lava]

how does 3.1 not really better than 3 ?

[proud bobcat]

I

Sorry

[surreal zephyr]

gpt models fail, google and claude models pass

[proud bobcat]

Benchmaxxed

[proud bobcat]

Not at all

[frosty lava]

when we got opus 4-6 it was much better than 4-5 and gpt 5.3 codex is much better than gpt 5.2 then why not for gemini 3.1

[low patrol]

so it is worse than gemini 3?

or better

[proud bobcat]

Uhhh

It’s like

[stray aspen]

great i just lost my massive manually made prompt to a somethnig went wrong error

[proud bobcat]

Marginal improvement

[low patrol]

🥺

[quartz light]

wrote a 9k char prompt with no new lines today without noticing

[half mist]

I would say Gemini 3.1 Pro is very good besides coding

[frosty lava]

they released a new model knowing it was just a small improvement ?

[proud bobcat]

Benchmarks wise it’s defo a lot better

Actual use?

Eh.

Defo a LOT more token efficient

[stray aspen]

benchmarks say theres a massive leap

[harsh flume]

Thats why LMArena is the goat

[sick mantle]

True

[frosty lava]

if that was really the case then we should see it

[half mist]

That’s why it’s Gemini 3.1 Pro instead of Gemini 3.5 Pro

[simple kayak]

claude is probably still better lol

[proud bobcat]

To be fair it is a refresh

[frosty lava]

like i said for example opus 4-6 is much better than 4-5

same for gpt

[harsh flume]

in b4 Gemini 3.105 Pro

[proud bobcat]

Gemini 3.1.2 pro xhigh ultra spark codex

[half mist]

Except for use on Code Arena which is very glitchy like every other new model added to Code Arena

[proud bobcat]

Gets 0.2% increase in swebench

[frosty lava]

but that's arena problem

not the model

[proud bobcat]

Yeah this is a joke Gemini 3.1 pro refuses to code

[half mist]

I still think Claude is better at coding, and I use it for my personal projects

[proud bobcat]

OH

WAIT

ITS DOINF SOMETHING

[frosty lava]

i would liked to see as much improvement from gemini 3 to 3.1 than what we saw for opus and gpt codex

[half mist]

I tried Gemini 3.1 Pro, and it couldn’t even get the start button right 💀

[proud bobcat]

I’m asking it to make a simple 3D fps shooter and it’s installed 6 packages

I’m sorry?????

[surreal zephyr]

same, it sucsk

[hoary elbow]

Is it still optimized or is it like when 3.0 first came out?

Cause after it came out, it got Nerfed

[proud bobcat]

What’s optimized

OH

[half mist]

Just use Claude, or even Gemini 3 Flash (which is somehow better at coding than Gemini 3.1 Pro)

[proud bobcat]

Quantization

[hoary elbow]

Oh yeah, is it still quantized?

[proud bobcat]

Uhhh no looks like it’s working at full precision

[hoary elbow]

Or is it like Gemini 3.0 prime

[proud bobcat]

As a tradeoff

Abysmally slow

[surreal zephyr]

quite sad imo

[plucky sparrow]

it's quantized

[proud bobcat]

No it’s not

It’s slow as hell dude

[surreal zephyr]

not even close

[proud bobcat]

That’s a full precision model

[plucky sparrow]

probably just cause everyone is using it. it doesn't seem much better

[proud bobcat]

Not how compute works

Google doesn’t do batch inference

[celest orchid]

its fun

[surreal zephyr]

bad

[pulsar crystal]

how do you know?

[proud bobcat]

Looks fine

[proud bobcat]

He doesn’t

Lmao

[surreal zephyr]

honestly

it probably is quantized

[plucky sparrow]

he's right, I don't, I'm just basing it on the fact there was a model 1-2 weeks ago that was miles ahead of the current release

[surreal zephyr]

doesnt mean we are ever getting the real model lol

[proud bobcat]

If it was quantized it would be fast

[stray aspen]

i hate this so much

[hoary elbow]

I’m getting mixed reactions. Can I make a poll here? I can’t find rules anymore and the last time I read, it was a long time ago

[surreal zephyr]

its quantized and slow

[frosty lava]

cause that was more like a new thinking mode no ?

[surreal zephyr]

it was very big before release

[proud bobcat]

https://tenor.com/view/patrick-spongebob-spongebob-meme-patrick-meme-dumb-patrick-gif-9974665538168463324

Two conflicting points

Models are quantized to be faster

[surreal zephyr]

opus 4.5 is quantized too

[proud bobcat]

How do you think opus is so fast

Yeah exactly

[surreal zephyr]

or cheaper

[surreal zephyr]

its not fast 💀

codex spark is fast

opus isnt

[proud bobcat]

Yeah it is?

[harsh flume]

For Claude users, what is the best EV in getting usage? Their plan or just go to Openrouter API?

[proud bobcat]

For a behemoth of a model?

Absolutely fast

[surreal zephyr]

LOL

[proud bobcat]

LMAO

[stray aspen]

this one

[proud bobcat]

Dude even Qwen crushes Gemini

This is bad

[surreal zephyr]

SIDEF VIEW

LOL

[harsh flume]

i'm too dumb to get the joke

[cyan harbor]

why the "remove" option is deleted?

[surreal zephyr]

2.5 flash does better than this bro

[harsh flume]

ill wait for the subtitled version

[frosty lava]

i think its lazy as hell

[proud bobcat]

Reminds me of that one panzer

[frosty lava]

you have to prompt it well

[proud bobcat]

🥀

I KNOW WHY IT SUCKS

[surreal zephyr]

?

[proud bobcat]

WAIT

YEAH

ITS BEING TOKEN EFFICIENT

[proud bobcat]

It’s being aggressively token efficient

The files it’s making don’t even go above 100 lines of code

It’s strangling itself

[surreal zephyr]

the tank has 300

[proud bobcat]

Not enough at all

See what I mean?

It’s been trained to use the least amount of tokens as possible to copy Claude

[surreal zephyr]

the noita prompt had like 3 times repeated to never take any shortcuts and such

900 LOC

not remotely close to what opus did

[royal sail]

A lot of people have been noticing this with 3.1 and I have too

[surreal zephyr]

LOL

[royal sail]

The model literally feels lazy

it also makes tons of silly syntax errors

[surreal zephyr]

gemini worse than gpt 5.2 low?

[stray aspen]

this is grok levels of disappointment

[royal sail]

I mean

The model isn't terrible by any means

It just doesn't feel great with some coding tasks

[surreal zephyr]

syntax errors

lol

[hoary elbow]

[lost patrol]

i think google is positioning it more for common user and maybe into the scienece direction.
Not for coding

[royal sail]

i'll admit that's pretty bad for an LLM to have lol

But majority of the time, it generates pretty reasonable code

[stray aspen]

i think claude could do this better

[royal sail]

high 🥶

[stray aspen]

considering it made me this

[sick mantle]

<@&1349916362595635286>

[mortal vale]

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

[stray aspen]

this is way better

[surreal zephyr]

welp

[stray aspen]

holy cook

[surreal zephyr]

the models not even close

[stray aspen]

lmao

[analog steeple]

god forbid third world countries from touching latest tech 🙏

[whole swallow]

WHEN DID 3.1 PRO COME OUT

[proud bobcat]

Try qwen 3.5 lmao

[whole swallow]

Is opus 4.6 that good?

[echo aurora]

TODAY

[proud bobcat]

Yes

[shell pewter]

anyone want to share how does gemini 3.1 vs opus 4.6 feel?

[bleak lake]

Benchmarks or tests?

[stray aspen]

@echo auroracan max route you to gemini 3.1

[proud bobcat]

Gemini 3.1 is great in everything except coding

Claude Opus 4.6 just gives you the same performance

Oh and Gemini is slow ahh hell

[shell pewter]

as in did any user test it extensively to have some own opinion, cus the benchmark from google said it really good but the rank on arena is a mixed bag innit

[shell pewter]

thx man its always you haha 😂 🙏 👍

[proud bobcat]

Got no babes…

Only ai…

[shell pewter]

same lmaooo 💀

[proud bobcat]

I love how my breakup was the reason I became invested into llms

I literally had no other hobby

[shell pewter]

really gemini 3.0 was the fastest among gpt and opus for me
it gotten slower?

[proud bobcat]

Gemini 3.0 still peak

[shell pewter]

but for grounding i like gpt 5.2 more than gemini 3 search tho

[stray aspen]

grok 4.2 is the best research model

[proud bobcat]

GPT 5.2 search is great if it would USE IT MORE OFTEN

[shell pewter]

is it i haven't use it actually

[proud bobcat]

Agreed

Grok makes the best research models

[limber panther]

penguin riding a motorbike

[shell pewter]

alright guess i gotta check it out then thx bud

[stray aspen]

yeah sometimes it goes through 250+ sources

[odd geyser]

Why is Claude 4.6 better than Gemini 3.0 pro in the text?

[proud bobcat]

It’s just peak

[crystal mica]

i will honestly say. gemini 3.1 pro is great in terms of solving math and etc. problems, but i hate it speech type and position

[odd geyser]

Oh, good then. I just didn't understand why it was better if it looked like it was going to be worse.

[proud bobcat]

Haven’t tried it too much yet

But it really goes crazy in coding

Not in a good way.

It installs way too many packages

Is slow

[sonic swallow]

gerar Prompt de vídeo profissional de uma pessoa que faz flexão rápido

[proud bobcat]

Lane

[whole swallow]

Overwhelming right?

[rigid copper]

hi guys

[proud bobcat]

It took it a solid 3 seconds to make one line of code

[odd geyser]

Idk. I like the style of gemini models.

[odd geyser]

Hi.

[rigid copper]

i'm trying to deal with the error in my arena.ai side lol

[proud bobcat]

Also Gemini 3.1 is heavily token efficient but not in a good way

It has too many shortcuts when coding

Claude is thorough

Every other model is thorough

[rigid copper]

umm @echo aurora i would like to get some help with this

[stray aspen]

it puts like 30 lines of code in a single one just to reduce total line count

🥀

[proud bobcat]

Real

[odd geyser]

Hmm. Idk. Maybe the limit, but I'm not sure.

[rigid copper]

the error was already like 3-4 hours and the conditions still the same

[surreal zephyr]

maybe garbage system prompt?

[proud bobcat]

THE CODE FOR THE MAP IT MADE IS BARELY 64 LINES LONG

WHATTT

[rigid copper]

i don't think it's rate limit

[royal sail]

I feel like it's actually not bad speed wise considering it's so good in regards of verbosity

Definitely faster than Opus 4.6

[proud bobcat]

Opus 4.6 was faster for me

[echo aurora]

It's hard for me to say for sure with the information I have, but based off this image I wouldn't be surprised if this was rate limit. Regardless, you'll want to follow the steps outlined in this message: #1417174113092374689 message

[rigid copper]

if it's a rate limit, it shouldn't affect other model like claude since i often use gemini 3 pro

[echo aurora]

Note that there is a rate limit for specific model use, but also a rate limit for overall modality use.

[proud bobcat]

One thing I do like is that Gemini 3.1 bug checks after it makes files

That’s good

[echo aurora]

Without being asked to?

[proud bobcat]

Mhmm

Yeah real

Oh

It finished coding

Let’s see

Yeah it’s garbage

It’s complete garbage

Oh my god

Hold on let me send the link

[soft matrix]

Opus 4.6 is good with some languages💗

[proud bobcat]

https://019c76f6-61ee-793f-938e-620cf50eca9d.arena.site/

This is god awful

[shell pewter]

kekw

[proud bobcat]

Lmarena isn’t nerfed

It’s the same model from Google api

[shell pewter]

so code is still opus (how good is sonnet btw?)
search is grok?

[proud bobcat]

Studio sucks right now

Arena has seemingly been better

Lmao let me get Claude sonnet to do the same thing

[shell pewter]

its really hard to just use logic without any search, or is it just me?

i agree actually

[proud bobcat]

Gemini losing the plot

[rigid copper]

it's normal for specific model to hit rate limit, but at my side, all model are unusable due to error as well (it happen to all model and even the model that i have not being interacting before)

[proud bobcat]

Claude is cooking

[shell pewter]

hey brada would you say
code is still opus and research is grok?
what else? how good is sonnet btw?

yes 100% this

[proud bobcat]

Sonnet is fire

I think a mix of opus and kimi right now is goated

Maybe GLM 5?

Research is always grok

Grok crushes research

Oh my god sonnet is COOKING

[echo aurora]

Yeah this sounds like it could be something else. I assume this is happening with different modalities as well?

Note I would like to move this convo to #1417174113092374689 so we can keep #general clear for chatting.

[plucky sparrow]

no opus deep think is best at research

but it's also super expensive

[shell pewter]

how much is sonnet performance of opus?

[shell pewter]

yup this

[inner relic]

Any leaks about deepseek v4?

The whale is really quiet

[proud bobcat]

I’d say 90%

[proud bobcat]

Hopefully today?

[wind hinge]

/image

[rigid copper]

possibility

[shell pewter]

thats really good then, consider that 4.6 sonnet is already better than 4.5 opus right?

[rigid copper]

tried all model and it won't work :/

[proud bobcat]

A bit yes

I’d say sonnet is 5-10% better than 4.5 opus

[stone tundra]

why is 3.1

so low

😢

[proud bobcat]

Gemini sucks…

It’s joever

[shell pewter]

I have been using AI for research and code, so for now i guess i only need to use grok and opus/sonnet?

[stone tundra]

text leaderboard

[proud bobcat]

LMAO I IUST REALIZED GEMINI 3.1 PRO SKIPPED MAKING THE ACTUAL HUD AND GAMEPLAY

BRUHHH

Ass model

[shell pewter]

i guess we need to give it a little bit of time for people to vote more?🤔

thx mate 🙏

[stone tundra]

lol

[astral vortex]

me too

[stone tundra]

yeah

idk how it fialed os badly

[lofty quartz]

मनुष्य किसी भी दुःख को, सहन कर सकता हैं, लेकिन गृह क्लेश उसकी, आत्मा को निचोड़ देती हैं..!

[inner relic]

I have to agree, Gemini 3.1 isnt that good

[proud bobcat]

It’s such a downgrade

Remember how crazy 3.0 was at coding to the point we were blown away

[shell pewter]

the benchmark numbers from google looks good idk 🤣

[crystal mica]

i like style of all gemini models , except gemini 3.1 pro

[astral vortex]

isnt is nano banana pro

[astral vortex]

why

[crystal mica]

it refuses to understand my meta-thought

[inner relic]

Alright, Let me guess, I think deepseek v4 releases tommorow

[shell pewter]

nah i think it wasn't that good still, the point for me was opus 4.5, that was actually usable good

[astral vortex]

I make use of json prompts to enhance and make it understand

[proud bobcat]

DUDE

[crystal mica]

can you tell me any?

[proud bobcat]

SONNET COOKED GEMINI BY A MILE

https://019c76ff-e7e8-7a61-aa48-00b1950d77cb.arena.site/

LOOK AT THIS

[rigid copper]

not sure is it i'm the only one who get this problem in arena.ai or anyone else got the same

[inner relic]

Gemini 3.1 is cooked

[balmy mist]

https://x.com/OfficialLoganK/status/2024516665740833156

Logan Kilpatrick (@OfficialLoganK)

Introducing Gemini 3.1 Pro, our new SOTA model across most reasoning, coding, and stem use cases!

X · Feb 19, 2026, 4:09 PM

[balmy mist]

anyone tried g3.1?

[astral vortex]

no it is general

[shell pewter]

yeah thats pretty good eh

[proud bobcat]

It’s ass

Like

I’m not even hating

It’s just ass

[balmy mist]

how??

like worse than g3?

[proud bobcat]

Terrible coding ability

Slow as hell

[astral vortex]

have u seen it

[crystal mica]

yes, thanks

[proud bobcat]

Yep

[balmy mist]

wtf

no way

[cunning birch]

Yeah

[balmy mist]

smh

[proud bobcat]

Sonnet cooked it

[cunning birch]

Google always make bad models on february

Gemini2 and 3.1

[royal sail]

This is just slamming the model for nothing lol

[proud bobcat]

If this was Gemini 3.1 flash it would’ve been understandable

[royal sail]

The model is not that bad

[proud bobcat]

It is

It IS

Sonnet wiped the floor with it in less time and 10 times higher quality code

[royal sail]

The only problem it has is the verbosity and uncommon syntax errors

[royal sail]

10 times higher quality based off what?

[surreal zephyr]

WILD

[proud bobcat]

For one the game actually

Functions
Has working dynamic lighting
High quality code
Multiple files
Robust systems

Gemini 3.1 made a mess

https://019c76ff-e7e8-7a61-aa48-00b1950d77cb.arena.site/

Here’s sonnet

[royal sail]

It's a bit unfair to base an entire model's performance off 1 prompt no?

[surreal zephyr]

it has awful system prompt

gemini sys prompt SUCKS

[proud bobcat]

Tried it twice

Failed both times

[royal sail]

this i can agree with lol

[royal sail]

same prompt?

[proud bobcat]

Mhm

[royal sail]

That's the problem

[proud bobcat]

The third time the game functioned

[royal sail]

You're basing your entire opinion of a model on a single prompt

[proud bobcat]

On the same prompt

[proud bobcat]

Not only that but community impressions here

Like ampro’s tank

[surreal zephyr]

look at this

[proud bobcat]

For a SOTA model that looks awful

[surreal zephyr]

gemini is BETTER if you not use system prompt

its WAY worse with system prompt

[proud bobcat]

Wow it really is

Oh my god

[meager tinsel]

I do not think 2 hours of testing is nearly enough time to form a general finalized opinion on a large multi-modal model like Gemini. For me it's generated some pretty nice stuff but in completely different medium than what everyone else is posting.

[proud bobcat]

Safety fine tuning fluff

[fading atlas]

OK

[surreal zephyr]

heres with system prompt

[proud bobcat]

Wow you aren’t even lying

God damn

[balmy mist]

did gemini deep think get updated as well?

[surreal zephyr]

gemini is actually better if you tell it to work hard and ignore sys prompt

the sys prompt DEMANDS it to be cost efficient and token efficient

[gleaming heath]

Fotor: Easy-to-use AI imaging editor at your fingertips. https://mobile.fotor.com/

[simple perch]

I was wondering why my photos don't look good in terms of quality anymore. I thought my app was bugged because it wasn't saving in 2K. Lol.

[rigid copper]

@echo aurora starting a new chat does the fix, but that means i can't use the old chat anymore because my project (usually from code arena) still going on.

anyway thanks for the tip

[astral vortex]

Kk

[rigid copper]

try to start the new chat in arena.ai, but that means you will have to start it over.

[mortal vale]

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

[echo aurora]

We’re exploring how occasional Battles in Direct chat might work. Our mission is to measure and advance the frontier of AI for real-world use, and integrating Battles into Direct is a meaningful step in that direction. The help center's article about the experiment can be found here.

[proud bobcat]

Gemini subreddit glazing the hell out of 3.1 pro

Woah it can make an svg…

[coral axle]

Hey everyone. Trying to use opus4-6 here in the arena, but it's clearly not the right model. It's running some version that doesn't even know its own existence lol. Any routing or deployment issues going on?

[shadow prairie]

Hi everyone, what can these lags be related to?

[fringe carbon]

ugh

anyone surprised 3.1 didn't top

[surreal zephyr]

look

[shadow prairie]

I have stupid lags.

as shown in the screenshot

[rotund seal]

no, it did?

[fringe carbon]

look at what?

[echo aurora]

You'll want to follow the steps in this article - #1417174113092374689 message

[half mist]

This is single‑handedly one of the least helpful features I’ve come across.

[surreal zephyr]

its worse than gpt 4o with default prompt

[fringe carbon]

you didn't label the pics

[echo aurora]

This is a typical hallucination. Essentially the model isn't able to perceive it's own existence when it's being trained, as it doesn't yet exist.

[mild dagger]

why i could not make video?

[half mist]

Battles should only appear if the user is in Battle mode

[echo aurora]

The Video Arena bot was removed from the server. Video Arena is only on the site now.

[shadow prairie]

I'll try to follow the guides, but it's unlikely to work out.

for Russia ❤️

[mild dagger]

Link?

[robust sonnet]

Yo guys getting like mad errors

[surreal zephyr]

I did but ur in dark mode

[echo aurora]

If it doesn't work be sure to follow Step #3. This is the step that helps us by getting us more information. Think of the Something went wrong error like a car's check engine light. A mechanic wouldn't be able to tell you what's wrong with just with the information that the check engine light is on.

[echo aurora]

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

[shadow prairie]

It's just that I have a huge project there, and if I create a chat, I can throw him a repository that he would copy?

[stray aspen]

thats insane

[echo aurora]

I'm sorry to hear you're not a fan. It's worth noting that this is still in the experiment phase, so there could be changes before being fully rolled out.

[surreal zephyr]

yup

[coral axle]

I get the pre-training argument, but in production, a properly fine-tuned model should have its identity explicitly defined in the System Prompt. If it's guessing whether it's Sonnet or Opus, it usually means the system prompt injection is either missing or misconfigured on your end, not just a random pre-training hallucination. Can you check the prompt wrapper?

[surreal zephyr]

i agree with him - the whole point of direct is no battles

[half mist]

At least make it so you can turn the feature off somewhere (if it does get fully rolled out)

[drowsy mural]

https://discord.com/channels/1340554757349179412/1474108826390106124

I don't even want to say anything—this feature is utterly ridiculous.

[coral axle]

@echo aurora ?

[fringe carbon]

when you vote in side by side does that count?

or does it need to be in battle

i thought only battle counted

but it lets you vote side by side which is weird

[echo aurora]

I can flag this, but this tends to be pretty common. I've seen this happen with a lot of different models.

[half mist]

No, no it does not. It’s a new Experiment that puts Battles in Direct Chat that is pointless in my opinion

[echo aurora]

Only votes in Battle are what power the leaderboards.

[fringe carbon]

what is the vote for in side by side then?

[sleek crow]

bruh

[coral axle]

@echo aurora Hey team, following up on the model identity issue with some visual proof.

Image 1: The official UI. The model knows exactly what it is (3.1 Pro) because the system prompt is properly configured.
Image 2: Your gemini-3.1-pro-preview endpoint. It's defaulting to a generic, outdated response (claiming it might be 1.5) because it clearly lacks a proper system prompt wrapper.

By exposing these raw endpoints without injecting the correct identity context, you're essentially lobotomizing state-of-the-art models. Users come here to benchmark, and when a model acts lost, they blame the AI companies, not the platform's infrastructure. It's tarnishing the models' reputations. You really need to update your system prompts for these new endpoints.

[stray aspen]

where can i utilize gemini 3.1 no system prompt

[stray aspen]

holy yap

[proud bobcat]

Standard hallucination

Models don’t know what they are

[mortal vale]

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

[stray aspen]

yeah what is tguy yapping aobut

[whole swallow]

Buddy that's not how it works

[proud bobcat]

Probably jus doesn’t know about llms

[whole swallow]

If I tell gemini that it's gemini 4.5 it doesn't become magically more powerful

[coral axle]

@proud bobcat Models don't know what they are unless you pass a proper system prompt at inference time. The official UI injects it. Your API wrapper clearly doesn't. Calling a missing system prompt a 'standard hallucination' is wild for a testing platform. But anyway, good luck with the benchmarks.

@whole swallow I never said a system prompt magically boosts capabilities. That's a strawman. I said it grounds the model's identity. If you don't inject the correct context (like the official UI does), the model falls back to older SFT data and hallucinates an outdated version. It's a basic context injection issue, not a capability debate. But I'll leave it at that.

[lost basalt]

This is the worst possible thing arena.ai could do with there loyal members, lost everything.

[stray aspen]

@surreal zephyr where can i use the best version of gemini 31

[surreal zephyr]

api

and tell it to put max effort ect

[stray aspen]

lm arena?

[surreal zephyr]

no shortcuts

yea

[stray aspen]

ok

[surreal zephyr]

its trained to be lazy for money

[hollow imp]

Vertex

[surreal zephyr]

you have to tell it to put more effort

[hollow imp]

How did you lose everything

[coral axle]

Look, I was just trying to give you a heads-up so you could actually update your system prompts. I have zero interest in using this outdated, misconfigured crap anyway. I'll just take this to social media to expose how you're doing false advertising and dragging these AI companies' names through the mud by serving crippled models. Have fun with your broken benchmarks. I'm out.

[stray aspen]

dude quit yapping

do wahtever you want

[hollow imp]

@surreal zephyr exhaustively enjoy 3.1 while you can, it's gonna be nerfed to trash after 2 days

[stray aspen]

ive never seen someone so upset because a model doesnt tell you its name

[surreal zephyr]

yup

[lost basalt]

I had a month old chat history which i build with time, now i can't build it again. It took alot of time. i tried to create a master prompt for a new chat but it doesn't work the same.

[hollow imp]

What problem you're facing mr white444

[lost basalt]

It randomly starts a battle and now claude 4.6 also has a rate limit which asks to wait for 21 minutes.

[hollow imp]

Why can't you wait

Do you have to catch a train?

[lost basalt]

& Something went wrong with this response, please try again.

[stray aspen]

its what it is buddy

[sick mantle]

Happends with Video too

[stray aspen]

if you want more pay aa claude subcription

[coral axle]

kkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkk

[hollow imp]

If the something went wrong with this response error is not fixing after repeated generations switch the model and continue the chat

You get enough rate limits that you can copy paste every single message from the error chat instead of scrolling instagram and wasting time

[coral axle]

just to say it's one thing when it clearly isn't

[lost basalt]

you're right but you can't make these changes suddenly.

[hollow imp]

1 blocked message

[stray aspen]

wdym its always been there

[pale sonnet]

3.1 pro is soo good

[lost basalt]

Already have done it but it starts battle mode during chat chats which is more annoying

[sick mantle]

'A blocked message from taco bell.'

[stray aspen]

where are you using it

[hollow imp]

Vertex bro

Vertex ai

[stray aspen]

vertex aint free

[hollow imp]

It is

[sick mantle]

Yea yea we know

[hollow imp]

200$ google cloud free credits

[pale sonnet]

9/10 ragebait

[hollow imp]

Chicken go lay eggs

[sick mantle]

Already did, i have 1000s of eggs

[proud bobcat]

Dude

[lost basalt]

Can only wait, maybe they will fix it. we can only criticize

[proud bobcat]

Even with a system prompt

It still doesn’t know

Gemini has told me it was 1.5 pro on the official app

[shadow prairie]

Oh gods, nothing helped how to live?

[stray aspen]

lmarena's gemini 3.1 is way better than the one on gemini app and ai studio

[pale sonnet]

skill issue?

[proud bobcat]

Yeah real

[sick mantle]

Bro, it does not only happen with claude but with video too.

[shadow prairie]

the problem with the error

[lost basalt]

Nah, everyone is struggling here

[quartz light]

claude models feel like that too but its probably the system prompt idk

[hollow imp]

@stray aspen vertex ai bro

https://docs.cloud.google.com/free/docs/free-cloud-features

[proud bobcat]

Claude processing 5 billion safety parameters before Something went wrong with this response, please try again.

[shadow prairie]

He just can't give me the code! I've been doing this for an hour now.

[proud bobcat]

Reset browser cookies

[shadow prairie]

I've tried

[sick mantle]

Gemini gives a free trial on gemini business for 30 days

[hollow imp]

No this is different

300$ api credit

90 days

[gleaming roost]

😊

[stray aspen]

gemini business is overloaded

it sucks

[pale sonnet]

https://tenor.com/view/patrick-spongebob-spongebob-meme-patrick-meme-dumb-patrick-gif-9974665538168463324

[shadow prairie]

Now endless coding

[echo aurora]

The rate limit just needs to be waited out and should start working again.

[midnight marlin]

Are the battles still in direct mode? 💩

[stray aspen]

yes

[lost basalt]

what about the Something went wrong with this response, please try again. ERROR?

[shadow prairie]

IT'S THE SAME FOR ME

[sick mantle]

I think pineapple has to delete the limt

[stray aspen]

no

these models aint cheap

[proud bobcat]

Oh it’s SLOW

trust me

[sick mantle]

Yea its the same for me

[shadow prairie]

bruuuhhh

Xs what should I do, I need to download the repository

[echo aurora]

You'll want to follow these steps: #1417174113092374689 message Also note that the rate limit can result in the Something went wrong error.

[echo aurora]

Yeah this is likely rate limit, spamming the chat when you're getting this error probably isn't helping.

[hollow imp]

@surreal zephyr @stray aspen

https://gemini.google.com/share/5689d70d3c7c

[stuck orchid]

Is Gemini 3.1 better than CLaude Opus 4.6 according to your tests?

[stray aspen]

insane

whats the prompt

[quartz light]

really efficient too

[surreal zephyr]

with proper prompt?

its not even close

bro

its literally AGI

[quartz light]

wow, it works on firefox so this could be used as liquid glass

[surreal zephyr]

https://019c7741-bf0d-7a29-b05d-7f7e191208a3.arena.site/ see this water

[surreal zephyr]

gemini

[quartz light]

bad test

[stuck orchid]

Claude Opus 4.6 > gemini 3.1 pro?

Okay

[hollow imp]

I have gemini deepthink which is based on 3.1 and it is NOT AGI

[olive mesa]

3.1 pro was agi

before it got nerfed

[stray aspen]

so thats deepthink

?

[hollow imp]

No

Normal

[stray aspen]

slide your ptompt

prompt

bro what hte hell is trinity large

[hollow imp]

If you want to test Deepthink send me your prompt

[woven peak]

what is trinity larg

large

is it good

[brittle tiger]

https://x.com/voxelbench/status/2024558946023854437

Voxelbench (@voxelbench)

We have a new leader.
Gemini 3.1 Pro!

X · Feb 19, 2026, 6:57 PM

[quartz light]

WTF ITS ONLY 350 LINES LONG

[woven peak]

exactly never even heard of it before

[quartz light]

THE CODE

[hushed gyro]

guys what the hell is trinity

[quartz light]

us based ai company

bad models

why?

oh they just released it on arena?

loooooooo

so late

[hollow imp]

Opus 4.6 non thinking 😭

[fickle venture]

Tf is trinity

[quartz light]

us based ai company

[echo aurora]

Try it out and let us know

[quartz light]

trinity large released a while ago

its not very good

but

[fickle venture]

USA?

[quartz light]

decent for first release

[quartz light]

yea?

why

[fickle venture]

United States of America?

The heck that's a spying model

[turbid timber]

is kimi better than chatgpt

[woven peak]

depends on the task and the chatgpt model you are talking about

[fickle venture]

If your a free user yes
If you are paid user no

[woven peak]

I think coding yes that's at least from what I've heard

[brittle tiger]

i've been using 4.6 opus in antigravity for days and 3.1 pro is fixing problems opus wouldn't

[celest orchid]

Who is trinity large

[woven peak]

USA company

[olive mesa]

even the quantized version of 3.1 is better than 4.6

[quartz light]

yea idk why people are happy that a us based company released an ai model

[celest orchid]

It's good ai?

[olive mesa]

which is what we have rn

[stray aspen]

no

just another open source model

[woven peak]

thinking or not?

(opus)

[olive mesa]

yes

at least from my and other people's tests

[fickle venture]

Cuz now they aren't gonna see a random Chinese worker watching them and selling their data 😭

[olive mesa]

the unquantized version which was only around for a short time at launch was also noticeably better than the 3.1 we currently have access to

[bleak lake]

Trinity-Large-Preview is a frontier-scale open-weight language model from Arcee, built as a 400B-parameter sparse Mixture-of-Experts with 13B active parameters per token using 4-of-256 expert routing.

It excels in creative writing, storytelling, role-play, chat scenarios, and real-time voice assistance, better than your average reasoning model usually can. But we’re also introducing some of our newer agentic performance. It was trained to navigate well in agent harnesses like OpenCode, Cline, and Kilo Code, and to handle complex toolchains and long, constraint-filled prompts.

[woven peak]

idk if this is crazy to say but I'd rather let china collect my data not USA cuz at least it's not going to israel or something

[bleak lake]

512k tokens btw

[surreal zephyr]

try api

[woven peak]

are u thinking about Deepthink or base model? cuz how could they quantize the day it releases, if u are thinking about base model then I'm assuming the short time is before it got added to Google ai studio?

[olive mesa]

i'm talking about 3.1 pro w/ thinking

[turbid timber]

what i really like about kimi is the agent mode

[fickle venture]

Correct

[turbid timber]

but i think ur right

[zealous sparrow]

https://019c774c-8fa7-7a3f-bc95-f01faad869df.arena.site/
someone fight me

[turbid timber]

chatgpt is definitely more worth paying for generally

[fickle venture]

Btw are you using CLI agents cuz I got something interesting for you

[zealous sparrow]

who beat me

[fickle venture]

You can use GPT 5.3 codex xhigh for free (idk about limit)

If you got macos install codex
If you got windows / linux install codex CLI
or opencode and link your account there

[shadow prairie]

And what should I do in this case? Wait?

[rocky mauve]

Game code

[sterile sun]

@echo aurora Why do you come with so much armour?

[zealous sparrow]

i left you can make your lobby and share codes

[stuck orchid]

479f2ec5-b5f9-47fd-9a38-2495390686f3

[echo aurora]

You should follow the steps in this message: #1417174113092374689 message

[echo aurora]

[sterile sun]

Pineapple’s armour

[rocky mauve]

Ggs u just smoked me

[sterile sun]

I bet if a pineapple would have fell on netwon, we all would be floating today

[leaden ivy]

Hey

[echo aurora]

Gravity defeated

[hushed gyro]

oh no.... i thought this is a dream... BATTLE MODE IN DIRECT???!!!

[proud bobcat]

its fine

[normal abyss]

mhm yeah, sorry bro

[surreal zephyr]

Its really good

[proud bobcat]

hold on let me try studio

[surreal zephyr]

Pineapple is heavier so gravity would be stronger

[surreal zephyr]

Studio is bugged

Llmarena is best

[hushed gyro]

dw side by side will work for me '

[surreal zephyr]

Make sure to force it to think much

[proud bobcat]

without sys prompt

[surreal zephyr]

Lma has no sys either

[normal abyss]

if you specificly want a clean chat with no interuptions, google AI studio is now the current best option. (if you like gemini)

[surreal zephyr]

But repeat many times to think excessively

Its trained to be lazy

[hushed gyro]

has anyone noticed that the quality of NB Pro has degraded after the errors?

[hushed gyro]

does it have NB Pro for free?

[normal abyss]

no i dont think so, higgsfield does though, but its pretty limited

[hushed gyro]

omg why does it have errors again STOP THE TORTURE...!!!!

[proud bobcat]

oh

damn

[drowsy mural]

Welcome, fresh victim. You're just in time for the worst update ever. Highlights include: destroying your conversation flow.

[robust sluice]

NB still errors but come with low resolution when it works

[surreal zephyr]

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

[hushed gyro]

for infinite generation glitch on arena: no login, generate 2 prompts, save pictures, clear cookies, repeat

[hushed gyro]

the quality has degraded since the errors

[proud bobcat]

"create a detailed 3d simulation of a tanks suspension with a map, driving controls, accurate physics, and polish. do not cheap out on code, think to the maximum and install as many packages as needed."

trying this out

lets race it against sonnet lmao

[normal abyss]

whats your opinion on this?

[robust sluice]

did they buff on Flux or something

[hushed gyro]

yes

based, W

[robust sluice]

never seen them work so fast

[echo aurora]

If you have experiences a lowering of quality we are looking for examples in #1470529575367868498, the top of the thread has instructions on what we're looking for.

[robust sluice]

so its a bug ?

[hushed gyro]

i feel like after nano banana pro released, all the other models suddenly become dogwater compared to NB

[proud bobcat]

the hell

lmarena is down

[echo aurora]

I'm not sure, we're trying to gather more examples

[hushed gyro]

this is dogwater lmao

[sick mantle]

Yea

[hushed gyro]

and it's NB Pro 🤣

[normal abyss]

is battle in direct still classed as an experiment or a main element of the site now?

[echo aurora]

Testing out Code Arena now, can't say I'm seeing the same (stll mid generation), but doesn't appear to be having issues.

[whole sundial]

I should bump this again as it is still heavily discussed here, please remove the battles in direct

[proud bobcat]

odd.

maybe i overly specified to install packages?

[echo aurora]

It's currently considered and experiment.

[normal abyss]

kk thanks

[echo aurora]

Oh wait now I'm seeing it. Reporting to the team. Thanks

[proud bobcat]

no problem

[echo aurora]

What model did this happen for you btw? I'm using battle and it appears is only happening with one of the options.

[proud bobcat]

sonnet and 3.1 pro

more specifically sonnet 4.6

[hushed gyro]

WHAT?!

Today vs... 2 weeks ago

[fickle venture]

@echo aurora edited: nevermind the scam got deleted

[split kayak]

Gemini3.1

[fickle venture]

Oh sorry it got deleted

[robust sluice]

yes sadly I rarely see them in battle

[echo aurora]

All good. Btw don't hesitate to use the @ Moderator ping for mod purposes.

[fickle venture]

I thought that doesn't work?

[echo aurora]

If you can share in #1470529575367868498 and include the Eval IDs for both sessions that'd be helpful.

[echo aurora]

No it does

[fickle venture]

Alr I'll try it when I see a scam again

[hushed gyro]

the prompts for the second one... were deleted

[echo aurora]

Okay good to know, if you notice this again don't hesitate to use that thread and flag the images/Eval IDs.

[golden ocean]

DOG WATER

[hushed gyro]

better or no?

[surreal zephyr]

https://019c775e-143e-7242-92b4-0d936da78f8c.arena.site/ the gravitational lensing is insane

[hollow imp]

Create a 3D simulation of a blackhole including accurate light bending, you should be able to pan the camera with the mouse, the blackhole should have a accurate accretion disk, web based

@quartz light

[stray aspen]

what model

[surreal zephyr]

3.1 pro

opus didnt get even close

[simple kernel]

Hey I've been waiting two weeks I don't see this feature in claude ai if it's going to be added soon?

[hushed gyro]

again... @surreal zephyr

[hybrid osprey]

Is nano banana being nerfed it’s generating 600kb files instead of the usual 6mb and failing to generate much more frequently

[meager tinsel]

They are good for different tasks.

[crystal mica]

guys, someone know how to export lmarena chat .txt?

[hushed gyro]

bro nano banana pro across all platforms are dogwater lmaoooo

[rapid otter]

.

[half mist]

They are not equal

[rapid otter]

i can send u code for this

add to friends

[rapid otter]

(i just needed it for myself)

[crystal mica]

sure

[surreal zephyr]

via api? gemini is infinitely better.
normally? opus

[quaint trail]

it just never works on arena

[pale sonnet]

rip🥹🕊️

[random violet]

Remove battle mode in direct chat... ☠️🤌

[gloomy onyx]

I doubt there is a public model which is able to overperform Opus-4.6-thinking in pure coding tasks.

Maybe Gemini 3 Deep Think Could actually be better, but only by reading its description card provided by Google, it is more of a research tool than a model made only for coding.

[scarlet spire]

GPT-5.3-Codex was pretty competitive in a fair few marks, wasn't it?

Given it itself anecdotally way outperforms GPT-5.2 and at least matches GPT-5.2-Highest in abilities

[pale sonnet]

Γεια σου!

[quaint trail]

everytime i try to use nano banana pro i get this error lol

[surreal zephyr]

https://019c779f-32f1-7eab-83f0-e1cfe2cf7544.arena.site/ custom wormhole shader

[gloomy onyx]

I guess so, but as Anthropic stated in its benchmark outputs, Opus 4.6 manages to be extremely good in agentic coding environments, extremely long context coherence and tool interaction / call

Those are parameters of evaluation which in my opinion are worth more than a single benchmark when we’re talking about pure vibecoding in real scenarios

[scarlet spire]

as Anthropic stated
^
first party rankings should always be externally verified for cherry-picking or unrealistic comparison factors

[gloomy onyx]

Yeah, true, I still remember Google faking their benchmarks for Gemini 3 pro. I just hope Anthropic isn’t that naïve 💀

It’s like steroids for AI at this point

[scarlet spire]

I'm not saying faking but I'm moreso saying if you use a metric that's a subset of something else that can only ever be measured as a whole, what does that tell us? nothing much.

[scarlet spire]

Because the context of the metric matters. If you remove its context, it can look phenomenal. But if it's doing a phenomenal job at X while being subpar at X, Z, U, V and W, it's mostly pointless.

[scarlet spire]

Congrats! You've asked the model for their identity

[gloomy onyx]

Classic hallucination case

[scarlet spire]

It's not a hallucination if you mean "GPT-4 class"

[quaint trail]

what does it mean by gpt-4 class

[scarlet spire]

It's aaaah. An unintended lens into how OpenAI got to it. It just means that GPT-5 is, as we know, a further-trained version of GPT-4.

[gloomy onyx]

I guess so, the model does not actually know the specific version of itself

[scarlet spire]

It doesn't

It never "knows". It isn't told.

[stray aspen]

ai studio gemini 3.1 sucks

[quaint trail]

id assume that

i have no idea how you even begin to train AI models, i can understand how they work

but training makes no sense

also coding in arena has recently just never worked for me, one time today but every time i get this error, and this error is common when using nano banana pro. not sure why

nano banana pro works fine when i use the actual gemini

[scarlet spire]

Simple concept of a Transformer:

1. You give it random noise created from adding noise to a real piece of data
2. You punish it until it guesses what the original was correctly every time.
3. You do this concurrently with a f"ckton of other data! :)

[scarlet spire]

Have you tried clicking "Re-generate"?

[quaint trail]

after multiple tries before i have, but i only tried on nano banana pro. ill see if coding with opus will work

[surreal zephyr]

it happens often

[muted bolt]

I dont like the idea that when we generate we can choice to have 1 generated at a time.. by the time I generate the 2nd image its saying im up to my limit which is wrong bc I normally get 5 chances

[surreal zephyr]

<@&1349916362595635286>

[vocal axle]

i dont speak english

[mortal vale]

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

[scarlet spire]

Yes it might. That means you'll have to try clicking re-generate "often" as well! :)

[quaint trail]

when they punish it, do they cut the power

[scarlet spire]

lol
no, they don't. It's a high level overview.

[vocal axle]

i am nexs user

[quaint trail]

https://tenor.com/view/silly-funny-too-silly-beating-beaten-gif-1299076303322216660

[surreal zephyr]

quota waste

[gloomy onyx]

There should be an identity dataset where they train the self perception of the model about its metadata, but I guess that to preserve the efficiency of the hardware allocated to the training session, it just reinforces the company name and maybe the model series

[scarlet spire]

Then just sulk, be my guest. Let me know how you get on and continue. :')

[light sleet]

When gpt 5.3 Codex😭

[scarlet spire]

This has nothing to do with it haha.

They can tell the model.
They usually do in the system prompt.
The API doesn't inject one.
Arena doesn't add it either. because that would be antithetical to the point.

[gloomy onyx]

But across all the training data (which, in case of ChatGPT we’re talking about terabytes of stuff), its impact is very limited

[scarlet spire]

Ask GPT-5.3 on the ChatGPT.com website and you'll see that it has no problem telling you.

[gloomy onyx]

I doubt it’s just prompt engineering

[quaint trail]

it worked on claude, but gemini 3.1 pro keeps failing to create anything

[scarlet spire]

That's not really how the training data works. Seeing the training data as "this is two kB and this is ten TB" is trying to count apples by moving oranges on your counting screen.

[quaint trail]

https://019c77a5-8f3a-7afc-a599-d7115b44f4ba.arena.site/

[gloomy onyx]

Otherwise it would be easier for the model to “remember” that stuff

[scarlet spire]

The system prompt is not included on Arena, as I said

That's why it doesn't truly know to tell you its identity. It's intentionally not part of the post-training.

[gloomy onyx]

It matters to the point where the model learns to have a better inference performance if the training dataset is generic enough (not implying that a lot of data = more generic, but it was an assumption that I made for how it’s trained a fundational model like ChatGPT)

[scarlet spire]

Good show! Yes, the ChatGPT website (first image) will indeed be able to tell you its identity just fine. It's told about what its identity is and what the date today is as well if I'm not mistaken.

[scarlet spire]

I'm saying that bytes aren't a thing in the neural net

[quaint trail]

i remember in like 2023-2024 using chatgpt when the website was green themed, and i kept asking it the year for its training data

i was pissed when it said october 2022 or 2023

[coral axle]

too

[quaint trail]

also it didnt even have web search back then

[coral axle]

fk lie

[scarlet spire]

Byte count doesn't scale from binary storage count to nonbinary neural net gradients. Your comparison of "compared to many terabytes" is one that doesn't hold up because the data when trained, is not transformed into bytes of any quantity.

[quartz light]

on opus its cuz of timeout

[quaint trail]

whats timeout

[lilac nest]

As a free user I've been using Gemini Flash 3 as my preferred everyday model. I use the limited free Pro for more complex tasks. Interestingly, Sonnet 4.6 from Claude was made the default for even Claude Pro users, so I thought I'd give it a try.

Initially, my thoughts are that it seems pretty good, but can't tell if it's better than Gemini Flash. I was wondering when it would be added to the leaderboard?

[scarlet spire]

That was likely correct for the dataset though. A year of delay between dataset termination and post-training completion is not unreasonable

[quartz light]

ai providers have different timeouts and basically if the response is taking longer than the timeout plan theyre paying for then it just stops

[coral axle]

I was very upset, swearing I was using opus 4-6 when in fact it's a clear derivation of 3-5. Shame!

[scarlet spire]

Once it accumulates enough votes and ELO confidence

[quaint trail]

that was when chatgpt started getting viral and nobody knew how LLM's worked

[scarlet spire]

You'll understand why this is the case when you conceptualise what the models are based on. They don't start from scratch every time.

[quaint trail]

so we just assumed chatgpt was like google if it could talk back

well, it basically is that nowadays

since it can use the web

[scarlet spire]

No, nowadays it's a teenager that talks back (and brattily so).

[lilac nest]

Ah so it's available to be voted on currently then. I searched the change log for it but I guess that must not include when things are added for initial voting

[stray aspen]

is gemini 3.1 nerfed already?

[coral axle]

@scarlet spire Dude, I'm developing my own model in Brazilian Portuguese from scratch. I've already developed my rag Online using some free models, and you can clearly see the difference when you use the latest templates directly from their respective platforms to the Arena platform.

[gloomy onyx]

Performance drift time record?

[scarlet spire]

Okay! That's cool. How are you finding the experience?

[coral axle]

@scarlet spire Look, I'm using Antigravity with Opus 4-5 and I was sometimes using the arena because I don't have the paid Opus plan, and I noticed that some days it always sends the same mess instead of fixing what it should do to continue training... then I just asked it to find out the version and realized I was using something that clearly wasn't it.

[scarlet spire]

The scoreboard displays a vote count, but the actual ranking isn't based on raw votes. The system that underlies this ranking is based on a longer-term heuristic-type averaging, the "elo" system.

[scarlet spire]

Look... at? Look at what exactly?

You're sending me random pings with absolutely no reference to what you are commenting on. You sound to me like you're trying to reply to something but are effectively failing to use the reply feature? Use the reply feature.

[stray aspen]

what is this ai conversation

[coral axle]

@scarlet spire Just configure the correct API to use the correct version and stop running those crappy system prompts from older versions.

[honest verge]

BRO I CAN'T TAKE GEMINI 3 FLASH ANYMORE

It forgets everything

After 3 prompts

I need 3.1