general

[compact flame]

Honestly I understand the Ai

Confusing miles with actual miles seems valid

[zealous sparrow]

2.5/3

0.5 deducted because the model didnt mention the ticket was forged

but i know that a model can't reach that

unless opus

[zealous sparrow]

this was 5.1 high

[compact flame]

Ig it did alright

[zealous sparrow]

give the same to gpt 5.2

it will be worse

if its better i take back all i said about OAI

[compact flame]

Uhh try giving that question I'm not at LMarena rn

[zealous sparrow]

both opuses got 0/3

damn

fiercefalcon flopped this one

[compact flame]

Damn didn't expect those to fail

[neat apex]

Opus 4.5 failed??

Maybe Gpt 5.2 Xtra High nails it xd

[zealous sparrow]

failed at my 3 questions

it aced my first 3

then failed

[neon idol]

what is this prompt

[zealous sparrow]

Lucy and Mary were at a concert, one of them got in but the second didn't, even tho the tickets were booked. Why?
Daisy and Mike were at a park. Daisy took 3 daisys, and mike took 0 why?
Luke and Miles were driving on bikes down a hill. When they got down to the hill, Miles was missing, Why?

[neon idol]

and the answer ?

[zealous sparrow]

A1: One of the tickets were forged
A2: Daisy took 3 daisy's because her name is literally daisy
A3: Miles fell to the side

[neon idol]

thx a lot

[compact flame]

Try asking these questions to extra high ig

It's available at yupp I think

[zealous sparrow]

yeah ima do that

[zealous sparrow]

gpt 5.1 high was the closest btw failed the forgery question

[bright shard]

@echo aurora Nano Banana Pro is throwing a lot of errors again; sometimes it works perfectly, but other times it keeps throwing the same error.

[compact flame]

High hopes for extra high tbh

[pulsar canopy]

:ablobwave:

[echo aurora]

[echo aurora]

We have been noticing higher than usual error rates, and our team is aware and working on lowering this as much as possible. However, when you mention:

sometimes it works perfectly, but other times it keeps throwing the same error.
This sounds a lot like it's being caused by rate limit. You can confirm this by: opening Dev Tools > Network > search Stream and find the Eval ID, there if you're seeing the Status Code = 429 that means it's being caused by rate limit.

[zealous sparrow]

2/3
screwed up the hill question

[compact flame]

I guess expected

[zealous sparrow]

so a shocker is like

google flash models screwed that question up too

so its prob all the new models that know Miles is always a unit and not a name

[hushed gyro]

you can edit messages in lmarena now???

[misty harbor]

using claude and it's stuck on generating, any ideas to fix it?

[visual osprey]

i feel like those questions have a lot of valid answers though

[zealous sparrow]

Yeah but you expect the ones you want from the LLMs

and if they can reach them

[visual osprey]

waht does that mean

[zealous sparrow]

They prove how good they are

[fleet lintel]

these simple questions are apparently difficult for LLMs

[zealous sparrow]

I changed Miles to Kyle so it doesnt do the measurement unit bs

[neat apex]

Its LLM fault

Gemini 3 must be the goat for this kind of question

At least the deep thinking

[zealous sparrow]

both flash models did the measurement unit bs

[visual osprey]

i mean for answer 1 i would say like she might be not following dress code or intoxicated

she took daisys because its her namesake

badly worded but crashing on the way down seems the most logical and likely

third question is really badly worded

[zealous sparrow]

Im going to ask 3 pro

[visual osprey]

saying when they got down implies both reached the bottom of the hill when the answer you want demands that one of them did not

[hushed gyro]

What models are in the video arena channels for image and video?

[zealous sparrow]

btw fiercefalcon and ghostfalcon [3 flash] failed

overall just wrong but also fell for measurement unit bs

[polar wharf]

any ideas what other models to add to gpt4free.pro?

[zealous sparrow]

Also the models often interpret booked as criminal booked

exactly

[polar wharf]

damn

[compact flame]

Damn booked up in a prison cell

Like hey I booked you a prison cell spot

[visual osprey]

thats not the correct usage

[zealous sparrow]

3 pro keeps interpreting the people as pets

or animals

or objects

[polar wharf]

[zealous sparrow]

1/3 [gemini 3 pro]

2/3 [failed ticket question]

[bright shard]

If it happens again, I'll check it to confirm it for you, thank you very much!!!

[proud bobcat]

what is the prompt btw

nvm i got it

[zealous sparrow]

bro if not the Miles BS
LLMs will say Kyle sounds like Isle or Cycle

[proud bobcat]

they think its some wordplay

[desert abyss]

Hello! @midnight vigil Please check ⁠how-to-video-bot to learn how to generate videos.

[zealous sparrow]

gpt 5.2 told me its wordplay just now 💀

2/3 failed the hill question

[compact flame]

Extra high said miles was missing

It measured in kilometers surprisingly

Though about the first one it said Lucy is a staff member

Brah

[misty harbor]

anyone knows how to fix a constant generating on lmarena?

[echo aurora]

A good way to check this is looking at the leaderboards - https://lmarena.ai/leaderboard/text-to-video & https://lmarena.ai/leaderboard/image-to-video

[echo aurora]

It depends on the error. I assume you're running into the Something went wrong with your generation, please try again error?

[cold minnow]

Yes how can i fix it

[misty harbor]

no, just generating

[misty harbor]

no that's not it

[cold minnow]

How cani fix it

[lucid nexus]

How to create video

[red meadow]

i cant generate anything, i always get the error the other user above got.

[olive mountain]

keep getting this error

[cold minnow]

Yea

[echo aurora]

For this error it's a bit difficult to troubleshoot as it can be caused by different reasons. For emini-3-pro-image-preview we have been seeing higher than usual error rates, our team is aware of this. With this case unforuntately, there isn't much on the user's end that you can do to get past this. But overall I would recommend: refreshing the page, starting a new chat, clearing cache. This can help, but isn't a guaranteed fix.

It's worth noting too that this error can trigger because of rate limit, which tends to be pretty common. This can be verified by opening Dev Tools > Network > search Stream and find the Eval ID, there if you're seeing the Status Code = 429 that means it's being caused by rate limit.

[echo aurora]

In #1397655624103493813 you'll find the information you're looking for.

[misty harbor]

@echo aurora what about the constant generating?

[echo aurora]

This one is more difficult to figure out what the issue is as an error isn't triggering giving some kind of status code. However, trying the same methods may help here: refreshing the page, starting a new chat, clearing cache.

[misty harbor]

i did all of the above, including a hard refresh

well the new chat creates a new chat but can't load the one that's stuck

[olive mountain]

after generating a image in the discord itself, how to download that generated image in the local storage? Or we cant download it?

[bright shard]

Yes Pineapple, I'm getting error 429

[obtuse heart]

Pineapple working overtime huh

[echo aurora]

Sorry to hear that didn't help fix the problem. Unfortunately, this is a known bug that can happen occasionally. If a chat gets too long, or long prompts/responses can contribute to this.

[echo aurora]

Yeah that's rate limit sorry to say.

[misty harbor]

oh no worries, thank you for trying to help!

[bright shard]

And how often does that limit reset?

[echo aurora]

IIRC it's 50 mins.

[pseudo hemlock]

But like who funds them

Does Google pay and pray (assume) their model with a fancy hidden name is going to be the best?

So they’re willing to give $1000?

[light tusk]

Is 5.2 not going to leaderboard?

[echo aurora]

It's on the WebDev leaderboard https://lmarena.ai/leaderboard/webdev, but isn't yet on Text

As soon as we have those scores we'll be sure to put out an announcement, so keep an eye on #announcements

[light tusk]

I’m taking the over on kalshi

[thorn path]

There's no way openai will outperform Gemini, Kalshi even has it as a 90% in favor of Gemini so everyone knows this

If you were waiting for leaderboard to update so you can find out who wins just go ahead and make the decision now since it's already not looking too great for 5.2

[empty stump]

openai is falling

[latent crest]

Hello @echo aurora When I try to login with google, it wants me to download a file named “google” , why???

[golden ocean]

https://cdn.discordapp.com/attachments/1170550524110721176/1339363741044707359/attachment.gif

[latent crest]

Pardon me; what’s that ?

[lime wind]

so how are people finding GPT-5.2? good in general? not revolutionary right?

[latent crest]

If it’s a virus, how is it possible? It’s an iPhone after all

[golden ocean]

joke

[whole sundial]

nano banana flash

but gemini 3 flash has to come out first

[fiery gull]

Calm down, it need votes

#1397655624103493813

[zealous sparrow]

not today for sure, there isnt even a test model for banana 3 flash

as in when the flash model comes out

eventually

[compact flame]

#1397655624103493813

[spark python]

Bro tried twice

[zealous sparrow]

but logan posting about nano banana 3 flash already huh

[neat apex]

Gpt Pro is already trash, imagine that

Imagine how many time would take an gpt 5.5 xtra high

[quasi atlas]

@round fox Please check on #1397655624103493813 for creations #video-arena-1 #video-arena-2 #video-arena-3

[compact flame]

I think what can save them is making a not rushed model that is trained properly

[weary galleon]

Highly agreed.

[fleet lintel]

they haven't not been able to pre-train a model successfully in 1.5 years. they are reaching the end of how much they can squeeze out of post-training. That is likely the reason that they haven't improved much in 2025.

[viscid cloak]

YO WHAT are hazel-edit 6 and ghost-pepper in image gen battle? Horrible models 🤣

[echo aurora]

This doesn't sound familiar, can you provide more information? Do you have a recorded video by chance?

[zealous sparrow]

hazel is OAI

ghost-pepper is apparently qwen

[golden ocean]

Large Language Model Arena

[queen veldt]

Sama this you?

Ultra high gpt 🌿 will save oai

Gpt 5.21.1 Ultra high x-max pro plus

[sour spindle]

I feel like I am the only one who likes gpt 5.2

[golden ocean]

yes

[sour spindle]

Welp the beauty of choice I guess.

[compact flame]

Uh based on what tasks do you use it

[sour spindle]

My job consists of using models for a lot of "text based work." A lot of research based queries. I have been comparing it with gemini and just 10 normal thing I do GPT was better in 7, 1 basically identical output, and 2 gemini was better.

[fiery gull]

400$ per prompt, but I can use free in lmarena and yupp and etc lol

[fiery gull]

is the start, I hope the nova 3 will be good 👀

[half mist]

What version does the ai use in Code Arena if you pick both good or both bad in battle model

[gaunt roost]

https://www.youtube.com/watch?v=wtct-rdzHKo&t=476s

[sharp mirage]

Hi

Anyone ?

[weary galleon]

Hi!

[compact flame]

Hi

[sharp mirage]

Hi

Anyone got any prompt for Clash of clans game ?

[echo aurora]

Hello

[sharp mirage]

Hey 🙂

[compact flame]

Uh try asking ai to make you a prompt?

[sharp mirage]

I am doing this

[sonic flare]

Hello guys, I'm new here.

Pls, which AI model is best for book content creation?

[compact flame]

There no best

It's based on what you prefer

[burnt sinew]

@fiery gull gemini 3

[queen veldt]

Gpt 4.5

[burnt sinew]

Why did it ping him

[queen veldt]

Highest context window

[burnt sinew]

5.5

[burnt sinew]

I thought gemini has highest

[compact flame]

What is this?

[echo aurora]

Welcome welcome would encourage you to check out our Text Arena leaderboard with the Creative Writing category -> https://lmarena.ai/leaderboard/text/creative-writing

[compact flame]

Damn he got whole texts ready

[faint drum]

Is LMArena down?

[compact flame]

No I think

No it's not

[slim spire]

GPT 5.2 broke for me

[zealous sparrow]

fellas we got some new gpt 5.2

on lm

[slim spire]

does gpt 5.2 work for you guys

[burnt sinew]

5.2 search nice

[half mist]

When is it gonna be on LMArena?

[zealous sparrow]

right now

[half mist]

Not all of them are on there. I only see 5.2 high, 5.2, and 5.2 search

[zealous sparrow]

the system prompt ones prob not fully added yet

[half mist]

But what about the code ones?

[zealous sparrow]

the names got reverted

[burnt sinew]

To what

[zealous sparrow]

just now btw

[half mist]

Why?

[zealous sparrow]

dunno

mayb issues

[cloud zinc]

grok 4.1

[zealous sparrow]

@echo aurora was gpt 5.2-code a new model or a finetune? It was removed immediately..

[cloud zinc]

its finetune for code

[zealous sparrow]

yeah immediately got removed

even if its a finetune

[sonic flare]

I kinda think Grok is nice for writing content

[burnt sinew]

No its worse

[sonic flare]

Which do you use

[burnt sinew]

Gemini 3

[sonic flare]

How about Gemini 3 Pro

[burnt sinew]

Thats the only one

[zealous sparrow]

I hope 5.2-code finetune isnt the same case as speciale

[cloud zinc]

there is also gemini 3 flash

[burnt sinew]

Oh let me check it out

[burnt sinew]

Where? I dont see it

[zealous sparrow]

LMArena battle mode

[zealous sparrow]

turns out these 2 are currently only battle mode models

[outer lark]

hello

[echo aurora]

Welcome

[fiery gull]

Create the book using opus 4.5 + gemini 3.0, and create the book with notebooklm or gemini app

Use gpt 5.2 xhigh for double check

[sharp mirage]

Thats good Idea 💡

[fiery gull]

Plan the book with gemini 3.0 and create the book itself with opus 4.5

[burnt sinew]

Oh got it

[sharp mirage]

Thinking

[fiery gull]

Is just like I do to get the best result

[astral bloom]

1, 7, 18, 45, ....?
sol:
115
aₙ = 3aₙ₋₁ - aₙ₋₂ - 2

[sharp mirage]

What is that?

[astral bloom]

someone sent it as a challenge for ai's

[fiery gull]

Bro is impossible to vote opus 4.5 vs gpt 5.2 xhigh 😐

Bro I don't have 1% of smart that AI

[sharp mirage]

Did you try gpt5.2xhigh?

5.2 high was so mid

[sonic flare]

How do i get access to opus 4.5 bro

[sharp mirage]

Lmarena

[fiery gull]

I'm using for work now, for my use I fell an improve that gpt 5.1

But I think is because the 'EXTRA' high

[fiery gull]

Lmarena 👀

[sharp mirage]

I think so

But tbh I don't think it's worth it to buy api or using it

Bro its so expensive

[fiery gull]

Bro but in code (word html) the gpt 5.2 is really cooking like gemini 3.0

[sharp mirage]

I got no time to try

[fiery gull]

Yeah, the 3 is sooo expensive 💀, but gpt is more that opus and gemini

[sharp mirage]

I was sleeping 🙂

[latent crest]

I sadly have not. It happened to me twice. I go into the page, then login, google and they ask me to download a file called “google” and I deny it

[latent crest]

Oh it’s from the iPhone

[fiery gull]

(generic project you can see it) the opus make a mistake in word html the gpt is perfect

soooo good the gpt 5.2 xhigh in word html

[sharp mirage]

Hmmm

Gpt is better here

Btw are you on Mac ?

[fiery gull]

gpt 5.2 xhigh = gemini 3.0

[fiery gull]

no, windows

[sharp mirage]

Pro?

[sonic flare]

I mean, which of this? lol

[fiery gull]

yeahhh, just use it to write

plan/thinking the book with gpt 5.2 xhigh

gemini 3.0 for create the book in gemini app

use the 3 lol

[sharp mirage]

Thinking

The first one

[fiery gull]

ahhh thinking, allways

[sharp mirage]

But it has a rate limt

[sonic flare]

Thanks man

[fiery gull]

I don't read the question lol, use allways thinking mode

[sharp mirage]

I do love opus 4.5 thinking but the rate limt hits so fast

[modest prism]

Which one has better vision and image understanding? Gemini 3 or gpt 5.2

[fiery gull]

I'm so anxious for still see the fight that 3, like messi vs cr7 :D.

[sharp mirage]

🙂

[sonic flare]

Yeah it's so annoying, i had to wait for 5hrs to reset

[fiery gull]

gemini 3.0 still the better, much

[sharp mirage]

Gemini 3

[whole sundial]

#1397655624103493813
<@&1349916362595635286>

[neon idol]

Fr

[modest prism]

Why is gpt 5.2 high so fast on lmarena it feels like it doesn't think and it's instant

[neon idol]

Maybe questions are simple?

[sharp mirage]

Pineapple saw that:

[neon idol]

Like 2-2

[fiery gull]

hmmm, I'll see it later o-o, but you using high mode?

[sonic flare]

Will that work? Wouldn't that mess up my content?

[fiery gull]

lol I delet it ;-;

[sharp mirage]

😉

[neon idol]

https://tenor.com/view/doakes-stare-gif-6201328377551420443

Pineapple now

[fiery gull]

see the direct chat 👀

[vivid coral]

OMG OMG OMG.....IT'S HERE!!!! @echo aurora is the 🐐 🐐 🐐

[sharp mirage]

Yeaaaah

Btw I think no one cares 😔

[vivid coral]

Everyone cares, nobody uses closed book LLMs in the real world, we all get caught up in this bubble we have here and don't realize what the masses want and need

[sharp mirage]

@echo aurora did you add glm4.6V?

[echo aurora]

Just added! You’re too fast

[vivid coral]

🤣 🤣 🤣

[modest prism]

LLMs are just glorified token prediction algorithms. Trillions of dollars have been wasted

[fiery gull]

we need it

and glm 4.6v fast? I want another small model in rank 👀 , serious this new model is toe-to-toe of glm 4.5v?

[modest prism]

Something I like about 5.2 is that the hallucinate rate seems a lot lower than Gemini 3 pro

[sharp mirage]

Btw you can now try it from the glm site

Chat.z.ai

Btw I think glm is cooking

Yea I told ya in the screenshot I send

[cloud zinc]

where can i access that

[sharp mirage]

But y'all said fake

[sharp mirage]

Are you guys shown together all the time ?

[modest prism]

API costs

[cloud zinc]

api requires money

xhigh is also token hungry, i aint paying

[sharp mirage]

Guys what is the best Minecraft command coder ?

[cloud zinc]

then why u on lmarena

[sharp mirage]

For now I found chat gpt 5.1high and glm 4.6

No extramehigh

[fleet lintel]

https://x.com/xw33bttv/status/1999293467806302261

Lex (@xw33bttv)

Holy shit lmao #ChatGPT 5.2 is quite possible the worst model they've ever released. I have no idea what the fuck they have done - there's no way this was the alpha model my cohort tested, nor is it even remotely close to how well 5.1 was performing the other day.

This is quite

X · Dec 12, 2025, 1:41 AM

ignore 5.2 model. Let's wait for 5.5

[fiery gull]

Gpt 5.2 thinking is a gpt what decide how many thinking itself will use

[fiery gull]

Yeah 🫦, we need still to clanker the gpt 5.2 to openai do a gpt 5.5 better

[fleet lintel]

they overtrained on arc-agi to create some buzz.. but real life performance got worse

[fiery gull]

Too, ever month a new chatgpt lol

[fleet lintel]

i have plus membership.. and i gave very decent shot to it and used it excessively today.
it's honestly trash compared to gemini 3

[sharp mirage]

Bro

[queen veldt]

Ye

[sharp mirage]

Bro its like every week another ai drop

[queen veldt]

Gpt 5 high is only for pro hsers

Extended thinking in gpt is medium

[sharp mirage]

Why funny ?

[devout vault]

Grok 3 is smart because it has less restrictions

[sharp mirage]

Good point

[fiery gull]

Grok 3 in 2026 lil'bro?

[fleet lintel]

yes, i think plus users gets medium. this actually makes me mad. they are treating paying customers badly. it honestly feels very scammy to me.

[whole sundial]

well its weights will be opened in early 2026

[fiery gull]

OpenAi seeing this 🤑🤑

[devout vault]

Chatgpt is no longer the godfather of Ai it became just like deepseek

[fiery gull]

Good, but the modeed deepseek v3.2 don't is very better? 🤔

[sharp mirage]

Claude killed the dream of chatgpt

[burnt sinew]

most people will continue using it though because it's all they know

[fleet lintel]

i rather pay for gemini 3 pro ..get 2TB storage as well and get much much better performance.

[burnt sinew]

rather get it for free...

[sharp mirage]

Yeah same

[burnt sinew]

and unlimited usage

[fiery gull]

But the openai have an 1ti of divide 💀

[sharp mirage]

And gpt isn't worth it

[fleet lintel]

through ai studio? I need enterprise controls, so i must pay

[fiery gull]

I'm using aistudio and gemini app same time

Gemini pro plan

Maybe a gemini 3.1 without lazy? My dream 👀

Exist veo 3.1.... I just need dream it

[fleet lintel]

good thing is that gemini 3 base model is really strong, it will be much easier for Google to do better post training and get the best out of it

[cloud zinc]

gpt 5.2 so bad

[weary galleon]

GPT 5.2👎 👎 👎 👎 👎

[queen veldt]

Post training

We still have 2.5 flash preview

It's never leaving the preview

[burnt sinew]

i think its a preview model?

[queen veldt]

We won't get the gemini 3 pro regular

It'll stay in preview

Until gemini 3.5 pro preview comes out or something

[sharp mirage]

Bro Open ai is lying to us ?

[fleet lintel]

this is not surprising to me. 5.2 is built on 1.5 year old base model. OAI had enough time to already squeeze the best out of it. Changing it further (like 5,1 to 5.2) would result in improving in one area (arc-agi) and downgrade in others.

[sharp mirage]

No way this true

[queen veldt]

They say the sponsors don't have any influence on testing

BUT

Arc-agi test is basically giving the model one unique test that the model hasn't been trained on

Sooo

[modest prism]

What benchmark?

[queen veldt]

If they paid someone in the company to snitch the arc agi test 2

Prompts

They could train their model for it

Which i honestly think happened

They probably have some snitch

I guess that's their code red

To do false publish

Of "super improved model"

I'm not amazed at all on the gpt 5.2

[modest prism]

If you guys find a reliable way to use gpt 5.2 extra high for free please tell me I need it so bad

[queen veldt]

I've tested 5.1 1 month ago i think

And it failed some tests that gemini 3 pro passed

And now the 5.2 failed same tests

It's some math problems which require bunchh of steps to get to final result

Gemini 3 pro is king

Claude is for coding

Even for agentic coding

Codex max 5.1 was terrible i had to re-do prompts multiple times since I couldn't start my app.....

Meanwhile sonnet and opus did them in first try like easy

[cloud zinc]

simple bench

[modest prism]

@cloud zinc

[queen veldt]

Regular customer's gpt isn't even here on the list tho

😂

[cloud zinc]

gemini 3 pro way better

[thorny schooner]

Is it just me or are the bugs are starting to become more common here on the website because I am seeing the weird disappearing witch becoming a lot more often in reports but also infinite generation unless you Reload Glitch becoming a lot more often both in personal experience and the reports

[queen veldt]

https://tenor.com/view/lucinda-lucinda-the-witch-sofia-the-first-wand-laugh-gif-14589531

[queen veldt]

Yes mods are aware of it for a while

No fixes yet

[thorny schooner]

I i'm aware I was just commenting on it since I'm pretty sure I was the first one who made a report on it at least for one of the glitches at least from where I can see in the report area

Because honestly both of those glitches has been going on for me for a bit while now

[queen veldt]

They don't even know why it's happening

It's a problem I've talked for a while..m

We neeeeeeeedddd to see the error code or something

So we can be moreeeee speeecifiicccc

Offf thee errorrr

Just the retry again isn't enough

[thorny schooner]

Well they probably need to figure out soon enough because if they don't they could be losing a lot of customers soon and fair enough for the code but i am not even going to try to see what code is in the error cuz I have tried to see the code and it looks confusing as hell

I already gave video examples to one of the staff

When it comes to the glitch itself

[bright shard]

@echo aurora An AI arena for audio, music, etc., would be amazing! It's the only thing LMArena is missing!

[burnt sinew]

google a/b test right now are for what you guys thing?

[burnt spindle]

hello

[burnt pulsar]

I've tried all day to get gpt 5.2-high to work on lmarena, no luck so far.

[weary galleon]

It doesn't have thinking.

[proud bobcat]

AGI!!1!1!1!1!1

[weary galleon]

Is this image AI-generated?

[proud bobcat]

What

This is artificial analysis benchmark

[weary galleon]

Maybe I'm wrong, but it looks like fake.

[pseudo summit]

could u send link pls? 👀

[proud bobcat]

Sure give me a sec

https://www.reddit.com/r/singularity/s/YaVYlnFDyv

Boom

[weary galleon]

KAT-Coder-Pro V1 has too much, and Gemini 3 Pro has too less.

[proud bobcat]

https://artificialanalysis.ai

Here’s the site

[proud bobcat]

Reminder that they add this all up by benchmarks

Like

Multiple

This is the average

[pseudo summit]

ty!

haven't seen that one yet, but it looks interesting

[proud bobcat]

Artificial Analysis is usually quite reliable

It’s a good assessment on general performance without the bias

[pseudo summit]

like in general? or like in relation to LMArena ranking?

o wait, just read ur second message again. im dum

[weary galleon]

As I said, it's fake

[lucid geyser]

@echo aurora How long do new models take to appear on the leaderboard

[echo aurora]

It mostly depends on the amount of votes we're seeing

[lucid geyser]

So how long until 5.2 do u think?

Also is there a higher probability of getting a newer model

[queen veldt]

https://x.com/i/status/1999548496198926728

Epoch AI (@EpochAIResearch)

GPT-5.2 scores 152 on the Epoch Capabilities Index (ECI), our tool for aggregating benchmark scores. This puts it second only to Gemini 3 Pro.

🧵 with individual scores.

X · Dec 12, 2025, 6:34 PM

[echo aurora]

I wouldn't want to give an ETA as I'd hate to give the wrong impression.

[weary galleon]

Do you have insider information?

[proud bobcat]

Hate on me if you wish but I feel this is another grok 4 situation

[golden ocean]

openai is cooked

done for

[echo aurora]

I am an employee here

[proud bobcat]

OpenAI spits out model supposedly SOTA, crushes benchmarks, blah blah

Comes out and while it scores great on benchmarks it fails individual tests

And daily use

Same with GPT5

[lucid geyser]

Nah it’s fine but how long have passed models taken like Gemini 3, gpt 5.1 if you know

[weary galleon]

Unveil all codenames to us.

[echo aurora]

Sorry to say I don't know. Generally though, the Text leaderboard will take about a week to update, but this can vary.

[echo aurora]

[lucid geyser]

Thanks

[lucid geyser]

Get hired

[weary galleon]

Please

🙏

[echo aurora]

Yeah sorry I can't provide many details on this.

[echo aurora]

Well since you said please...

[weary galleon]

🙏

[queen veldt]

It would litteraly end up on reddit in 2 hours

[lucid geyser]

Might have to leave a bad review

[echo aurora]

https://tenor.com/view/andersomviolao-gif-12381443319241980618

[neon idol]

Bro knows more things as we expect

[queen veldt]

Yeah same as google employees

They have insider information

= free cash on polymarket

[neon idol]

Is that the Microsoft support client room?

[empty stump]

is gpt 5.2 xhigh only in the api

[vivid coral]

my early returns are that GPT 5.2 search is definitely an improvement over 5.1, need more time to say how much yet

[golden ocean]

REAL

[weary galleon]

WoW! It's too subjective.

[echo aurora]

Just kidding though

[viral cedar]

@weary galleon how is that wrong?

[weary galleon]

GPT 5.2 is a piece of crap

[native yarrow]

^

[viral cedar]

"trust me bro" statements by AI companies

😐

[empty stump]

every time they behind they just add more reasoning effort

[viral cedar]

wasn't Deepseek 3.2 said to also be really good

but turned out ass

[hollow ivy]

Longterm, OAI has no chance vs Google (and even vs Anthropic).

[torn mantle]

lol

[hollow ivy]

-# Grok is a wild card.

[torn mantle]

told ya

oai hit a plateau

[hollow ivy]

Currently Anthropic is ahead.

(coding is most important discipline, long-term.)

[viral cedar]

💯

[hollow ivy]

Hopefully, Anthropic has high-security standards..

(vs spying, hacking, etc)

[proud bobcat]

3.2 did its job well

I use it daily for math and roleplay

I love it

[hollow ivy]

is it top-3 in roleplaying?

[compact flame]

Do people actually use ai for roleplaying rn?

[weary galleon]

Roleplaying with a robot? Hmmm...

[hollow ivy]

yes

[proud bobcat]

….

Sadly yes.

It’s a HUGE market

That

I contribute to.

[hollow ivy]

not sadly, it can be fun

[proud bobcat]

It is fun

[queen veldt]

How fun?

What is it roleplaying?

[fickle venture]

So guys which model is better for coding?
GPT 5.2 or Claude 4.5 opus

[proud bobcat]

#5 rn set to become #4

[hollow ivy]

opus-4.5

by a huge margin

[sullen quest]

opus

[proud bobcat]

Adventure, stupid scenarios, just in general

A little unholy business on the side

[fickle venture]

Dam

[queen veldt]

Opus is op

[proud bobcat]

Opus clocks GPT

[fickle venture]

I will try gpt 5.2

[queen veldt]

Try it why not

Its bad tho

[proud bobcat]

I don’t think 5.2 is bad, I just think openai doesnt know what it wants to be

DeepSeek is math god, Claude is code god, Gemini is vision and jack of all trades

But what is GPT?

[compact flame]

Balance ⚖️

[queen veldt]

Balance between broken code and bad math

[compact flame]

Universal ai ig

[native yarrow]

GPT is for students who dont know stuff about ai and will use the popular one

[fickle venture]

Tbh what does GPT mean?

[proud bobcat]

One size fits all

[fickle venture]

Cheating exams yeah I use it alot

[proud bobcat]

Yeah true

[proud bobcat]

Use DeepSeek

[fickle venture]

And I end up getting F

[proud bobcat]

You will not be let down

[queen veldt]

Nah I'm student and using gemini 3 pro

It's waaay better

Even for casual chats

[proud bobcat]

DeepSeek is so peak at geometry for me

[fickle venture]

I mean on my school I got no internet so I just use mobile data chatgpt seems to use way less so it's better

[compact flame]

Speciale or just thinking?

[proud bobcat]

You have to be a bit more wordy but it pays off

[proud bobcat]

Speciale thinks too long

[fickle venture]

I will use it on my next physic exam

[proud bobcat]

I use no think

[compact flame]

Fair

I swear speciale has some sort of paranoia

[compact flame]

Everytime I look at it's thought it's always thinks what if?

Speciale feels like chatgpt pro with long reasoning

[proud bobcat]

It’s great for reasoning but only for like questions of the universe

You have to specify FULLY

Just use the no think and thinking variants

[vivid coral]

Nobody uses commie Claude with their ridiculous limits they impose here, and everywhere

[hollow ivy]

-# in battle mode it has no limits :)

[proud bobcat]

Mogged.

[vivid coral]

on code mode?

[hollow ivy]

text/chat mode (default)

[compact flame]

Damn it's expensive

[proud bobcat]

People pay the premium though cause it’s good

I used to have a hatred for Claude but they make solid stuff

Really solid

Opus 4.5 thinking is a MACHINE

[vivid coral]

Yea but on battle the model changes on every follow-up, which is great for some testing, but not if you wan't to test longer chains

[compact flame]

Why did you hate them tho

[hollow ivy]

Untrue. If you never vote, the model stays the same.

[sullen quest]

anthropic killed their family

[vivid coral]

WHAT?!

[proud bobcat]

Just use side by side comparison

[compact flame]

What

[proud bobcat]

Sonnet 4 sucked

So I didn’t use their models for a while

[sullen quest]

so?

[proud bobcat]

Came back when 4.5 released

[compact flame]

Wym anthropic killed somebody

[proud bobcat]

And I really like its prose

[vivid coral]

no, limits are heavy there

[hollow ivy]

you need a (google) account, though

[compact flame]

Anyways

[vivid coral]

hmmmm interesting

[compact flame]

Chatgpt needs some training

It's supposed to be ⚖️ not uh downgrade

[proud bobcat]

There’s limits???

I’ve never hit limits at all??

[proud bobcat]

Dw 5.3 comes out next week and 5.4 next month and will fix EVERYTHING

[hollow ivy]

you use lower models? or only battle?

[proud bobcat]

Are we talking lmarena or Claude

[compact flame]

Fr?

[latent crest]

U guys how can I use Midjourney for free??

[proud bobcat]

I don’t know, but wouldn’t be surprised lmao

[fickle venture]

Dude just drop GPT 6 why like that

[compact flame]

Dude if we ever get chatgpt 6?

[hollow ivy]

gta6 before gpt6 ^^

[compact flame]

I won't be surprised there will be 6.7 gpt

[vivid coral]

I'm not sure why some people are getting heavy limits and some aren't. I just assumed everyone got them since I did, interesting

[compact flame]

Cuz knowing openai this might happen

[fickle venture]

GPT 6 for sure cuz GTA 6 ain't coming out

[proud bobcat]

ChatGPT 6.7 will finally be able to answer how many r’s in garlic first time

And it’ll say “SIX SEVENNN” every 2 prompts

[compact flame]

Frfr

Would be a great April fool's model tho

[vivid coral]

GPT 12.2 will be out before GTA6

[proud bobcat]

GPT 10 will finally be able to make coherent organized code

Trust

With 5% LESS errors

[fickle venture]

Imagine OpenAi skips GPT 9 like Windows and Apple did

[proud bobcat]

GPT 8 will have ads integrated

Like every sentence there’s an ad

[proud bobcat]

https://tenor.com/view/ishowspeed-try-not-to-laugh-gif-7682731162751353849

[golden ocean]

https://tenor.com/view/stan-twt-stantwitter-stan-twitter-relicario-gifs-relicariogifs-gif-11979182475974977057

[torn mantle]

the only one im seeing whos still far ahead from hitting plateau are google

[thorn path]

Wait did I miss the final numbers?

[lucid geyser]

Why do u think that

Claude is so good even without reasoning

[torn mantle]

burnout / no improvement on pre-training ( thats why they are starting from scratch ) / less data quality compared to google / key staff elements poached by other labs

[lucid geyser]

Grok has the most powerful cluster

[torn mantle]

what did they do with it

[lucid geyser]

I mean they suck but it was recently completed something is possible

[torn mantle]

also

https://cloud.google.com/blog/products/compute/using-cloud-tpu-multislice-to-scale-ai-workloads

google can basically use multislice to create a more powerful virtual cluster

than xai

trust me, they are far ahead

be it on hardware / software

[mild anvil]

Does anyone know if they've extended the limit of 5 videos per day?

[torn mantle]

i just searched, so the maximum they can pack with this method is 50k but thats still way faster than any cluster for ai training giving how efficient their TPUs

[hollow ivy]

maybe a fun game for some people? #ai-creations message

-# (it's 100% free & open-source)

[stray aspen]

can we get claudius 4.5 opus on vision arena

[burnt sinew]

did they lobotmize gemini 3?

[hazy forge]

claudoctius 4.5 opus will NOT get vision arena status

[thorny schooner]

Have no clue feels like it has been like that for a while now

[atomic lagoon]

Hmmmm thats weird because I seem to be using that commie claude with their ridiculous limits. You gotta manage it correctly

[viral cedar]

did yall see the mcdonalds and coca cola ad

that made me mad

[atomic lagoon]

It was funny

The mcdonalds one

[mild anvil]

I was able to generate about 10 videos in a single day, wasn't the limit supposed to be 5 per day? Does anyone know if they increased the limit?

[proud bobcat]

They’re so ass

[viral cedar]

yes, idk why they did it 😭

[noble vessel]

hello

[vivid coral]

I don't really understand the hate. It's not life changing, but Twitter is ridiculous with the hate for this thing

[proud bobcat]

The search is easy

But the model itself sucks

It’s very clearly benchmaxxed

[vivid coral]

I guess that's fair

[lofty anchor]

Bro I use Claud for code I hate the thinking limit can someone tell me how to ovoid the limit

[golden ocean]

Cwaude

[burnt sinew]

@echo aurora 🌩️

[golden ocean]

@🍍

[whole sundial]

<@&1349916362595635286>

#1397655624103493813
this too

[lucid geyser]

Like they don’t have to be physically connected?

[plucky sparrow]

Because it was over hyped, even by openAi themselves. High benchmarks. "Code red".

If they just called it gpt5-high-context I think a lot of people would've been happier

[vivid coral]

lol, i've seen those screenshots all over twitter

[plucky sparrow]

the other thing is, unlike Google, who releases Gemini 3 Pro (and not just Flash) to everyone to use, GPT models are typically paywalled

so a lot of people who have access to these models are paying for it, meaning their expectations (and also hatred, if expectations are unmet) are higher

Altman seems to do a good job at selling to VCs, but not such a good job at knowing how to appeal to the general public

probably a lesson to be learnt there somewhere if you're thinking of starting a company

actually I think the other reason is the amount of hype-litter all over twitter for GPT5.2

Gemini 3 Pro had a lot of it too, if it wasn't able to produce much, pretty sure there'd be a lot of gemini 3 pro hate too

[verbal shadow]

Hi

[plucky sparrow]

hi

[plucky sparrow]

I literally just went to x.com, and saw this at the first post at the top of my feed.
https://x.com/slow_developer/status/1999661802666557487

Haider. (@slow_developer)

i'm still kinda confused how openAI made that much progress with gpt-5.2 when gpt-5.1 was only a month ago

my guess is it was an internal model they held back due to high compute costs and because they didn't think it was needed

until gemini 3 and opus 4.5 arrived

X · Dec 13, 2025, 2:05 AM

if people get bombarded with this, and they try the model, expecting "incredible progress" and it can't answer simple questions, yeah, they're going to post about how it sucks

[astral bloom]

system prompt in code arena

[tawny brook]

Which open source model is currently the best overall

[tawdry vapor]

Did you mistake it for glm-4.6v?

[echo aurora]

Yes, good catch. Thank you

[modest prism]

AGI

[astral bloom]

why did the sys prompt i pasted got deleted?

[bright shard]

@echo aurora An AI arena for audio, music, etc., would be amazing! It's the only thing LMArena is missing!

[shell oasis]

LMARENA IS THE BEST PLATFORM I HAVE SEEN TILL DATE. ADDING UP THE VIDEO ARENA IN THE WEBSITE IS 🔥🔥🔥🔥

I wanted to support by donating some amount... @echo aurora is there a link for donation?

[whole sundial]

lmarena is a private company so they don't accept donations, they only make money from evaluation services

[shell oasis]

Still I am mesmerized by the progress. I write parody songs and I haven't been able to create videos for the lyrics till date due to expensive subscriptions by the AI websites
LMArena opened the gates for me. I really am very grateful 🙏🏻

[shell oasis]

I am feeling very happy after seeing the video option when I opened lmarena website today

[whole sundial]

the video arena is not on the website, maybe you are confusing it with something else?

[shell oasis]

@whole sundial but I saw this option

[whole sundial]

looks fake, where is the plus button?

[shell oasis]

Sure?

[whole sundial]

@echo aurora is this legit?

[shell oasis]

Looks genuine to me

[whole sundial]

oh probably an a/b thing

[shell oasis]

That's why I came running here to express my happiness here 🤩

[tired shadow]

I tried going to the same link

it gave this bruh

hmm

maybe its a bug like on my phone

[shell oasis]

Am I lucky? I have literally created my first video

[tired shadow]

on my phone I had web arena before i got it on PC

oh hell nah

my phone has 0%

[whole sundial]

assuming it's an a/b test, you are pretty lucky

[shell oasis]

tears of joy 🥹

🥹🥹🥹

[tired shadow]

please emulator I need this, my lmarena is kinda videoless

[whole sundial]

looks pretty legit to me

[shell oasis]

I will be able to create videos for my 58 parody songs

Thank you from the depth of my heart LMArena 🙏🏻

[hushed gyro]

do you have access to beta test?

[shell oasis]

I have no idea, I just use lmarena on daily basis

I opened normally today like every other day...

[tired shadow]

noooo

[shell oasis]

I am logged in bro

[hushed gyro]

@echo aurora yk if random people have access to features early, why not give it out to EVERYONE?

🤔

[tired shadow]

it may be a bug

[hushed gyro]

can you check if sora is on there

[shell oasis]

yes its there

[hushed gyro]

well I want to keep this kind of bug lmao

[tired shadow]

sora 2 pro?

[tired shadow]

+++++

[shell oasis]

yes sora 2 pro

[tired shadow]

YUPII

[hushed gyro]

woooooooo~~~

[tired shadow]

thank god

[shell oasis]

God, thank you 🙏🏻

[tired shadow]

finally after I waited so long

[hushed gyro]

OMG?????

VEO 3 FOR FREE????

@echo aurora nah

yo can someone from the company explain why this guy has video arena on the site???

[whole sundial]

i see strings that relate to a video arena in the code, this is 100% real

[tired shadow]

but lmarena is like this, we have claude opus, sonnet, and literally gemini 3 pro (gemini 3 pro is kinda limitless I think)

[shell oasis]

That's why my first message after I came running here today 😁

[hushed gyro]

if you remove this string, is the limit removed

[whole sundial]

i doubt it

[hushed gyro]

pineapple has a lot to explain...

[tired shadow]

nah, it will just dont show anything

[whole sundial]

yeah, imo this is the type of feature that shouldn't be a/b tested like this, it's either launched or not launched

[hushed gyro]

can I add you as a friend?

[tired shadow]

ill check my alt on lmarena

[whole sundial]

if it's not launched and just an a/b test, lots of people who want a video arena will be upset

[hushed gyro]

that isn't explicit anyways, let alone a/b test

they probably have done it poorly, not rolling it out to everyone

[shell oasis]

Are you all logged in to the website also?

[hushed gyro]

yas

[whole sundial]

(well, we already know it's a/b)

yeah neither of my logged in lmarenas have it

[shell oasis]

So good news is we are all going to get the video arena soon on the website 🔥

[shell oasis]

sure

[hushed gyro]

the problem is, pineapple usually has the speed of FLASH to address these kind of problems

[lucid geyser]

What does the video selection even look like

[hushed gyro]

wdym? how to get into it?

[whole sundial]

it's pretty late for him though

[shell oasis]

Looks like this

@lucid geyser

[tired shadow]

im starting operation alt check, I have 10 alts

[hushed gyro]

@lucid geyser there should be a video button next to Image & Code

[hushed gyro]

brother what

[lucid geyser]

On the bar though

[hushed gyro]

just click the models

[lucid geyser]

It doesn’t show image models without clicking image

[hushed gyro]

@echo aurora pls... we need an explanation on what's happening

[lucid geyser]

What happened

Bro he’s sleeping prolly chill

[hushed gyro]

he literally went online minutes ago

but whatever

I want this a/b test situation to end

roll the vid arena out to everyone!!!

[lucid geyser]

Why

[hushed gyro]

well if you have noticed

some ppl have the video arena on the website, some don't

[shell oasis]

Maybe after a few hours it will roll to everyone...just like how android updates happen to get feedback

and then stable update to all

[torn mantle]

kinda, they are still interconnected with fiber but not in the same datacenter

[zealous sparrow]

Videoarena came to the website

[shell oasis]

you got it too?

[zealous sparrow]

Yea

[shell oasis]

Great 👍🏻

[keen topaz]

Hello, this is Lakki. I am a web developer. If you need help with any project, you can hire me

[torn mantle]

you seem like a vibe coder ngl

[compact flame]

Fr?

[whole sundial]

<@&1349916362595635286>

[hushed gyro]

YAY!

[hushed gyro]

I don't have it???

[shell oasis]

Soon

[keen topaz]

Need?

[jade cloak]

hey, some devices got Video generation option (including mine), but some didn't. why

[queen veldt]

Samsung browser no video arena

[jade cloak]

@echo aurora

[compact flame]

Maybe pc only?

[tired shadow]

no

I tested on chrome, tor and firefox

[compact flame]

I guess maybe it's country based

Like with sora

[whole sundial]

so it is a/b, an honestly pretty stupid one at that, they could've launched it on beta lmarena first and announce that instead of basically making people jealous for one another based on if they have the video arena or not

[tired shadow]

I dont think so

[tired shadow]

ye im sure its a/b

[compact flame]

Hmm

Then maybe it's just to some users

Like early access

[tired shadow]

a/b

[jade cloak]

try India

i don't think its country based

[hushed gyro]

we need an explanation from the company now!

and we are slightly upset

[jade cloak]

look guys

there are many models

[sterile tartan]

Yupp has even More

But LMarena is Better for Convenience and SOTA Models

[jade cloak]

UI preview for you guys

[hushed gyro]

@echo aurora hey! umm I have noticed some users have access to the video arena on the website, but some don't.

Can you explain if this is an a/b test and if it is next time pls be explicit about this

[sterile tartan]

Wait since when LMarena has Videos on Website

💀

[hushed gyro]

like a few hours ago

@shell oasis this guy got it first

[sterile tartan]

Seems like is rolling slowly

[sterile tartan]

Lucky him

Probably a valuable user

[hushed gyro]

does anyone realise that this company sometimes does things in a shady way?

[jade cloak]

ye

[hushed gyro]

he uses it daily so probably that's why

[sterile tartan]

The Algorithms Yeah

[hushed gyro]

THEY SAID DECEMBER!!!

[shell oasis]

I am a DevOps Engineer...I find new ways to automate the code deployments...Heavy automations

[compact flame]

If it thinks you're old enough?

How is it even supposed to tell?

Well probably talking about taxes with chatgpt maybe gonna work

[shell oasis]

Rockstar Games also said GTA VI in May 2026...but November 2026 now 😔

[compact flame]

Maybe the video update is not intended?

They just maybe accidentally rolled it to some users

[shell oasis]

@compact flame if some got it, maybe its intended for all

[hushed gyro]

that would mean that backlash is coming for them, since they made a fully functioning feature (that is heavily requested) then decided to gatekeep it

This happened with the Retry Button on Battle Mode, took them a month to add it back... smh

[compact flame]

Cuz like nobody said anything about video arenas

It's just silence and boom it's here

[hushed gyro]

as I said before, this company does a lot of things in a shady way, this isn't a good example of transparency, I don't have a lot of trust into this company

[shell oasis]

I wanted that bro...I always secretly wished lmarena to allow video creation on website

[hushed gyro]

I mean the 3 channels in this DC server are lackluster

[compact flame]

Well let's hope video arena is real and it's not an accident

[hushed gyro]

https://discord.com/channels/1340554757349179412/1449335460223783014

@shell oasis @compact flame check it out

[shell oasis]

Manifestation works 🙌🏻😇

[compact flame]

Hm maybe we can copy a link that leads to this feature?

Like with code and etc

[hushed gyro]

try inspect

[compact flame]

I tried didn't find anything useful

[hushed gyro]

how did they find video string?

@whole sundial need some help over here

[compact flame]

I'm not that good at inspect anyways

[whole sundial]

developer mode, go into debugger and you should be able to search for certain things

[hushed gyro]

where is that on Edge?

[mystic flower]

I’m a full-stack developer building a project and I need an API key for image and video generation.

[whole sundial]

i'm not sure

[north osprey]

heloo

[stark forge]

I'll help the first 10 people interested on how to  start earning $100k or more within a week, but you will reimburse me 10% of your profits when you receive it. Note: only interested people should send a friend request or send me a dm! ask me (HOW) via Telegram username @Susan _Vachon

Or The telegram link in my bio

[whole sundial]

oh i just figured it out

[whole sundial]

<@&1349916362595635286>

[hushed gyro]

tell me pls

[hushed gyro]

hey uhh you might be able to find the URL that leads to vid arena

[whole sundial]

ctrl + shift + f

[whole sundial]

don't think that's possible

they made it in such a way that you have to be a part of the a/b test group to access the video arena

[compact flame]

Says something related to experiment

Ig maybe it's experimental

[hushed gyro]

well how to apply for that test group?

and there's no way pineapple is silent about this

something really sketchy is going on

[compact flame]

Maybe there more people who got it it's just like maybe was randomized

[whole sundial]

it randomly chooses people

[compact flame]

Choosing people randomly is not efficient for testing ig

[queen veldt]

Ummm

[shell oasis]

Well you know, only 2 videos are allowed per day 🤯
then check back after 14 hours

damn

[compact flame]

Can you like share the link to video arena or it's not possible

[wispy sierra]

Hello

[shell oasis]

https://lmarena.ai/?mode=direct&chat-modality=video

The link won't work

[compact flame]

Hmm it gives me an error sadly

Well I guess it was worth a try

[shell oasis]

Unless its released

But very low limit right now, only 2 videos per day

damn

[compact flame]

I think because everyone will be abusing sora 2 pro or whatever

[shell oasis]

I would have to buy a local PC only but right now all AI companies ate up RAMs 😔

[shell oasis]

I write parody songs and I wanted AI videos to create the videos for my lyrics

[shell oasis]

Yeah true though

[wispy sierra]

A bug occurred, and I posted it on the bug forum. How long will it take for the moderators to see and fix it?

[north osprey]

How do I delete a video that has already been generated, sir?

[whole sundial]

there is only one person here ( @echo aurora ) that looks at them, so it may take anywhere from a few hours to a few days

[edgy wharf]

Gemini 3 pro accidentally fed me its internal thought pipelinei nstead of the proper output. Is this common knowledge, or something that's not known?

[wispy sierra]

Thank you

How does LM Arena allow us to use paid AI models for free?

[queen veldt]

Nobody knows

They are paid by big companies for testing the models (in battle mode)

Those secret name models are paid by companies

But they say for direct chat and other stuff lmarena is paying for API

For 6 images with nb pro you are costing them $0.9

[wispy sierra]

İts best thing ever on the in the internet

[queen veldt]

Maybe they get it for cheaper idk

[queen veldt]

Yeah

[whole sundial]

they are paid by companies that want their models evaluated + they have over $100 million invested into them

[willow sleet]

how was offline llms guys using lm studio or ollama? fast generation on rag or just use notebook lm? heard context is so low and no memory at all..

[queen veldt]

What

[wispy sierra]

Have you noticed that ChatGPT 5.2 forms sentences with missing words and inverted grammar? Why can't this AI model even form a sentence

[meager harbor]

gemini 3 does it too

me thinking ai will only get smarter

I was wrong

[compact sleet]

It sure do talk more American, it's agi.

Jkjk

[zealous sparrow]

yeah videoarena is like a rollout rn if someone is wondering

Here is what the video player looks like

[compact sleet]

But actually... Imperfections in phrasing and grammatical structure on a normal conversation sounded more... Natural right? Perhaps it was trained on it? Iunno. Just hope it's not messing around on logic strict tasks, like coding or general analysis.

[slim gorge]

how's gpt-5.2 guys

[wispy sierra]

Weird

[zealous sparrow]

mid as hell

[slim gorge]

well thats was expected tbh

they're just rushing things a lot cuz they dont wanna fall behind the competition

[zealous sparrow]

in other terms it was benchmaxxed

they argued it has a goated OCR, later when compared to gemini 3 pro OCR it wasnt even close

[slim gorge]

they're pulling a grok move

[compact sleet]

It's okay if you want to generalise it, it's doing very well on common life hood related tasks, on analysis, and on logic training. I'd say it's on par with gemini 3 pro. But only on high thinking sadly.

The only grace it had over other models are creative writing at the moment. But not sure if people here used it for such purposes.

For its price, it's a bit underwhelming.

[zealous sparrow]

it scored very bad on simplebench

simplebench

[ocean ferry]

bro don't even try challenging Gemini at OCR task dawg

[slim gorge]

yeah gemini is unmatched at OCR and vision in general

[zealous sparrow]

simplebench tests LLMs with these questions

Yet 5.2 xhigh scored #8

[slim gorge]

L

openai falling behind, google and anthropic are gonna be at the top

[ocean ferry]

bro and that Gemini 3 pro that almost match human on simplebench is the lobotomized version dawg

[zealous sparrow]

correct but the lobotomy was mostly done on coding

not thinking

[compact sleet]

It does prove nothing. It's more encouraged to bench it yourself on the lmarena.

With your own needs and logic set of testing. It's free to test anyways.

[zealous sparrow]

I benched it with my own questions

[compact sleet]

I agree it fails on coding compared to other models.

[zealous sparrow]

It didn't score 3/3

[compact sleet]

Which prompt you ask it?

I want to replicate your own findings

[zealous sparrow]

I had a whole uh

prompt for testing

xhigh got 2/3 sure

but i noticed one bad thing with it

[compact sleet]

Mhm, just post it. I'll replicate it on my own llm arena

[zealous sparrow]

It confuses Miles with A measurement unit

Or Kyle as a word game

[compact sleet]

Ah the miles yesterday?

Daisys miles kyles

[zealous sparrow]

You were there when i did the tests right

yes

that

[compact sleet]

Wait

[zealous sparrow]

LLMs often just confuse stuff with word games, is what i observed

[compact sleet]

is the prompt this?

Lucy and Mary were at a concert, one of them got in but the second didn't, even tho the tickets were booked. Why?
Daisy and Mike were at a park. Daisy took 3 daisys, and mike took 0 why?
Luke and Miles were driving on bikes down a hill. When they got down to the hill, Miles was missing, Why?

[zealous sparrow]

Yeah.

I have answers for this too

[compact sleet]

Fair, you can crosscheck my prompt when testing it too then.

As follows:

Make a scenario of where three guys met in a bar, each of them told a story, in which there are unclear lies woven from every of them, not made because they want to lie, but they simply didn't get the picture clearly at that time. But, there was also a shared truth among their similar stories. They argued of which version was the right one.

The bartender came, and told the lies and truth of their story, because the bartender saw the incident himself.

[ocean ferry]

yeah bro and the most hyped one is coding bruh, why do they even lobotomized it, it would've scored like 80%+ on SWE-Bench Verified

[zealous sparrow]

Saving TPUs, idk?

[compact sleet]

This will test LLM complex logic of making at least: 3 lies on a similar story, 1-3 shared truth of a similar story, 3 real truth on verification.

All in a same timeline event.

It's a generate scenario and analysis scenario at one.

[zealous sparrow]

This right here is too easy of a question for LLMs

[ocean ferry]

what model is ghostfalcon?

[zealous sparrow]

Some sort of gemini

I highly believe uh

Gemini 3 flash

[ocean ferry]

is it good?

[zealous sparrow]

From my testing, eh. For codeArena

seahawk and skyhawk were better imo

[ocean ferry]

nooo they even lobotomized G3 Flash!!😭

[compact sleet]

@zealous sparrow is your answer is like this for the Lucy Daisy and Miles test?

1. Lucy = Lucky
2. Daisy = Name of Flower
3. Miles = a unit of distance?

[zealous sparrow]

I can give you the answer

A1: One of the tickets were forged/invalid
A2: Daisy took all the daisies or it was just her name because LLMs struggle to reach that point
A3: Miles fell off the side

[compact sleet]

🤔

[zealous sparrow]

both failed the Miles question

from my testing

no model currently scored 3/3 on this

5.2 xhigh was close before failing on the Miles question

[compact sleet]

Then at this test of yours that is being replicated in my place, Gemini 3 failed all 3 then?

It literally thought of a name play, instead of the most possible yet the most boring scenario.

[zealous sparrow]

This is also just an easy question

[compact sleet]

Of course I can re-run the question, just in case

wait..

[zealous sparrow]

yeah both models

just have in their training