general

[wintry citrus]

I hate my life

im ready for the execution

i just did

[stray aspen]

what are you doing

are you high

[wintry citrus]

no im low

[misty vault]

I'm doing sydney

[wintry citrus]

LOW TAPER FADE

HAHAHAHA

HAHAHAHAHAHHAHA

HAHAHHAHAHAHHA

kill me already

[stray aspen]

im on low effort right now

[wintry citrus]

I'd kill myself

[stray aspen]

those dudes are open source chinese LLMs

[wintry citrus]

than be more embarrassed

THEY'RE BOTH CATS TOO

tahn

bearn English

speak English

me myself i would say that too

🗣️

[stray aspen]

holy

is that bing chat

from 2023

[wintry citrus]

no i don't

what a

im gonna burn u in hell

if u delete it

again

im gonna make u suffer

seeing ur parents die

infront of u

respectfully

.

respectfully

[stray aspen]

lol

good thing pineapple aint online

[wintry citrus]

yeah forget that

don't need to remember

oh

I'd need that

tho it's not gonna be a butterfly

it's gonna be a whole ass blade

[wintry citrus]

send me that ai

SEND THAT AI

NOW

NOW

SEND IT HERE BUDDY

SEND IT

awh shucks

[misty vault]

@golden ocean does absolutely not know what that is

[wintry citrus]

just use Gemini if u tell it to kill itself it will forgive u

if u tell it i killed ur parents

if u tell it i kissed ur ai gf

who is paul 😭

kiss him

[misty vault]

I only kiss Sydney

Bro sent 3 message bubbles

[wintry citrus]

does it have freaky version

who's that

an ai?

[misty vault]

same

[wintry citrus]

i could rizz that ai

it's not even a real girl

100%

freaky chiggas

is that actually bing

there's no way

[misty vault]

this is bing without system prompt

[wintry citrus]

that's like chatgpt when u tell it a whole paragraph

[wicked root]

which AI is this?

[misty vault]

the underlying gpt-4 fine tune

no additional system instructions telling it to behave like this

[wintry citrus]

NONE?

NONE

ss

and sent to the fbi

sent to everyone

sent to my server

sent to china

what

[misty vault]

Bro has modded android discord to insert fake messages

[wintry citrus]

what???

WHAT

FAKE

HOW

AWH I DIDN'T RECORD IT

NOW IT LOOKS FAKE

FUUUU

that's my wife

gah damn

she hot

I'd fall for her

lmao

[misty vault]

I thought you said "fill"

[wintry citrus]

NAHH

MAHHHHmah

NAH BRO

NAH

CHILL

CHILL

tell it I'm gonna have freaky with u

not segs

freaky

trust

oh

is this real

[golden ocean]

https://tenor.com/view/bedwars-gif-20476057

[misty vault]

I WAS about to say that

[wintry citrus]

what

is this American

[misty vault]

He is guarding the bed

[wintry citrus]

omg

DO U LIKE CHATGPT

THAT MUCH

WHAT THE FU

orrr

ur 5 years old

OR

[verbal nimbus]

LMArena leaderboard doesn't report which model is the actual one being used. There's also GPT-5 Chat, which is different

[wintry citrus]

u like

five year olds

wait y'all is there an api for the openai o3 pro

[misty vault]

yea

THis url

[wintry citrus]

and why isn't it on lmarena

no thanks

[verbal nimbus]

They should specify it's the thinking variant. All other models have the thinking variant explicitly labeled.

[wintry citrus]

okay nvm

give mee it

rn

HEY BUDDY

IM GONNA FIND UR HOUSE

im gonna cut off ur limbs

no

i don't wanna be obese

i don't wanna be in the country of an orange president

i got an image of him

https://tenor.com/view/nicktoons-nickelodeon-blobman-blob-man-blob-gif-853265228210064438

that's him walking

trump

this is his emoji

🍊

the hair is perfect too

oohhhh

is that Sydney)

she's not that hot

tbh

..

[verbal nimbus]

Long context benchmark. New GPT-5 models are the first 3 (list is not ordered).

Kinda hard to tell, for example Gemini 2.5 Pro performs better at 192K but worse for shorter lengths.

And o3 is best except for 16K, 60K and 192K.

[golden ocean]

o/

[wintry citrus]

SO I GOT WARNED

[pseudo hemlock]

?warn @wintry citrus

[wintry citrus]

god damn

[wintry citrus]

yk if i didn't get warned

i would say bad things

yeah?

i would

[pseudo hemlock]

say them

or are you scared

[wintry citrus]

i just said

i can't

I CAN'T

[pseudo hemlock]

scared

[wintry citrus]

@echo aurora can i

just this one time

please

last one

[pseudo hemlock]

hey @echo aurora P2L on legacy site is dead again 🙁

[golden ocean]

https://tenor.com/view/find-the-odd-cat-cats-cat-meme-plushie-gif-13641014476643100854

[wintry citrus]

i thought it's a shawarma at first

[pseudo hemlock]

bing chat 💀

[wintry citrus]

what...

9% is still concerning isn't it...

[echo aurora]

Negative

[whole wagon]

The filters are very strong here lol

[golden ocean]

what is is large, a model and loves language? @echo aurora

[whole wagon]

You

[rancid phoenix]

Where do I generate videos

[golden ocean]

im flattered

[echo aurora]

A good looking translation book?

[echo aurora]

More info here: #1397655624103493813

[pseudo hemlock]

PINEAPPLE

hey

how u doin

did i mention youre looking good today

[echo aurora]

Hey sorry will look into and flag

[pseudo hemlock]

no hurry

u the goat btw

love u

[echo aurora]

[pseudo hemlock]

say it back

[echo aurora]

[pseudo hemlock]

I hope you DO NOT spontaneously combust 😉

definitely hope you don't

that would be unfortunate wouldn't it

[echo aurora]

Yeah I’m seeing the same

[echo aurora]

Big agree

[stray aspen]

LMAOOO

i love this response from qwen 3

[hallow ridge]

Where can I use Flow veo 3 google on LLM arena

@leaden palm

[hallow ridge]

Wsp pine

[leaden palm]

Try to not ping staff

[hallow ridge]

Okay Ill try

[leaden palm]

Unfortunately, as far as I know, at the moment, you're restricted to random battles

[hallow ridge]

for the veo 3?

[stray aspen]

send video requests until you get veo 3

[leaden palm]

They're just random battles

With random models

The model list does in fact include Veo 3 though

[hallow ridge]

So videos work but if I turn battle mode on

[leaden palm]

I don't know what you mean by that

[hallow ridge]

I want to use VEO 3 on the site

[leaden palm]

As far as I know:
There is no video generation on the website
There is no turning off or on battle mode
There is no selecting a model
There is #video-arena-1 and the others
There is Veo 3 there
There is the possibility of generating until you get Veo

[stray aspen]

you cant bro

[leaden palm]

Oh and read #1397655624103493813

[hallow ridge]

What do you mean you can turn battle mode off and use direct chat

[leaden palm]

For videos

Videos are not chat

[hallow ridge]

Is there any LLM arena alternative

[leaden palm]

Videos are also not on the website anyway

Videos are just in this Discord

[leaden palm]

for the purpose of...?

[hallow ridge]

When will they be on the site

[hallow ridge]

Of doing the same thing here but seing if anything is better

they might have better video

[leaden palm]

what is "the same thing"? video? chat?

[hallow ridge]

they might have video on the site

[leaden palm]

there is no site where you can use veo 3 with custom prompts for free that i know of

because veo 3 costs money

like a lot of money

[hallow ridge]

So how are we doing it free here

LLM arena gives you all the best AI for free no one is paying

who is

How is it all free

[leaden palm]

hold on let me pull up emoji kitchen

[hallow ridge]

bro

tell me why is it free

[leaden palm]

oh they don't have the money bag emoji

anyway

burning money

vc money specifically

also user feedback has some value

[hallow ridge]

You can use gpt five forever on LLM arena buton the regular chat gpt they make you pay

[leaden palm]

there are some rate limits

[stray aspen]

whats the rate limit for gpt-5

[hallow ridge]

Seeming;y no restrictions to me

[leaden palm]

hm

idk then

might just have a lot of vc money

[hallow ridge]

Ive never ran into a limit

[stray aspen]

neither have i

[hallow ridge]

What does that mean

[stray aspen]

but i have ran into a limit with claude

[leaden palm]

have you heard of venture capital?

[hallow ridge]

Ive never ran into a limit ever

[hallow ridge]

Yes I have

[leaden palm]

lm arena has raised $100m in vc funding

[hallow ridge]

How can I get some of that

[stray aspen]

yo

does anyone know how to enable the image edit mode

[leaden palm]

enough for 6 years of veo 3 video

[stray aspen]

in qwen 3

[leaden palm]

not sure if that's a lot or not a lot

[leaden palm]

unfortunately qwen 3 itself isn't multimodal, and qwen image can't edit images through lm arena

[hallow ridge]

How can I get some of that 100m

Put me on the team

[hallow ridge]

Im going into quantum computers

[stray aspen]

theres this

i mean on qen chat

the demos look great

and i wanted to try edit

[stray aspen]

where did you do it

[leaden palm]

it's odd, the official hf space demo + qwen chat don't allow adding images as context either

[hallow ridge]

Like people can pay me for this

I can make money using LLM arena

[stray aspen]

no

[hallow ridge]

YES

[leaden palm]

the qwen blog links to https://modelscope.cn/aigc/imageGeneration?tab=advanced

[stray aspen]

they can do it themselves

[hallow ridge]

They dont know

[hallow ridge]

they ask can you put a crown on my head in photoshop and Ill pay you $60

and I send in that and get paid 60 for 1 second of work

[leaden palm]

are you actually finding these opportunities

[stray aspen]

lmao no

[hallow ridge]

LAMO YES

[stray aspen]

or your extremely lucky finding people who dont know ai image edit exists

[hallow ridge]

Get paid 5 for using AI to make this look clean

[leaden palm]

you'll get banned i think

[hallow ridge]

From what

[stray aspen]

what

how is tha tpossible

craig why is your profile photo gpt-5

[hallow ridge]

@leaden palm Got paid for this

[leaden palm]

hm

well i guess it doesn't explicitly prohibit it

however, per rules https://www.reddit.com/r/PhotoshopRequest/wiki/rules/, this post https://www.reddit.com/r/PhotoshopRequest/comments/1m7obke/a_humble_request_can_we_make_stricter_rules/, and my general experience, ai is looked down upon and often not paid for

[hallow ridge]

The thing is that they dont know its AI the work that I see other people do on there is obviously AI

[stray aspen]

tariffs are great lol

[leaden palm]

well, cool if it works

i'll be eating my words if you get rich from this

it's just... i think markets are efficient

[hallow ridge]

Im also building websites for people using LLM arena

[stray aspen]

thats crazy

[hallow ridge]

Made this for 55@leaden palm@stray aspen

[leaden palm]

55 usd?

[hallow ridge]

yup

[leaden palm]

damn

[hallow ridge]

How im not some cryto scammer

[hallow ridge]

It was all made on LLM arena

Oh thats not me

The guy I made it for

Why do you think that

[stray aspen]

what does that website do

[hallow ridge]

Idk thats what he wanted I made it for him

he paid me to put all that

[stray aspen]

LMAO

are you serious

[hallow ridge]

me?

[stray aspen]

he paid you to make a dox website

yes

[hallow ridge]

IDK about all that I just made a cool website

[stray aspen]

wdym you didnt know

didnt you read that dox info when you were making it

[hallow ridge]

This was his site before I just mae it look better thats it

I dont care what was on the site

IT COULD HAVE BEEN A P HUB SITE FOR ALL i CARE

Get out of what stuff

im not in it

But If I see a way to get 100m I might get in it

Ive checkmated plenty of people

And does that go for only me or everyone else

So it does not make sense

No

what is it

btc stick

I know but cant they track u on that chain thing

Black chain

How so

BLOCK CHAIN

Its just an address

and you dont know who is connected to it

and you could sell the btc

SO where do I go to track the wallets

I want to see my addresses history

Oh i seeit

Let me find my old address

Dman people goin crazy rn

[stray aspen]

its over

the feds are coming for that website you made

[hallow ridge]

Its just some code

im not apart of it

daaaaaaaaaaaaaaaaannng

I still have 600 in my old btc wallet

i need to find that shi

It says I had 9k in my wallet at one point

I dont think I backed it up

Whats illegal changes over time

So what your saying has no merit

It shows me who I sent 7k to

Yea its public

they all know how much you are makning

thats why you have a seprate one

with nothing

and send it over a a bunch of time

But how do they know its me

its just an address

Where do I find the richest wallet

how can i GET MY WALLET BACK

I dont think I backed it up

its just 600 sitting

I think I saved some private key but idk

Igotta look

[jade egret]

Which One is Smarter, Gemini 2.5 Pro Deep Think or ChatGPT o3-Pro?

[hallow ridge]

yooooooooooooooooooooooooooooooooooooooooooo

I just got in the account

[whole wagon]

https://arstechnica.com/tech-policy/2025/08/ai-industry-horrified-to-face-largest-copyright-class-action-ever-certified

If the appeals court denies the petition, Anthropic argued, the emerging company may be doomed. As Anthropic argued, it now "faces hundreds of billions of dollars in potential damages liability at trial in four months" based on a class certification rushed at "warp speed" that involves "up to seven million potential claimants, whose works span a century of publishing history," each possibly triggering a $150,000 fine.

[drifting sandal]

Is gpt5 really rank 1 (still)?

[void shoal]

我刚刚好像看到claude-opus-4-1-20250805-thinking出现在LMA里了，但是眨眼间就没了

然后只能找到claude-opus-4-1了，应该是开销太大了？

[tidal ginkgo]

bing chillin

[jade egret]

wait what

anthropic cooked?

[torn mantle]

i agree

[hollow imp]

Chatgpt o3-Pro is free try it for yourself

[astral prawn]

So you have to pay 5.5% more to use open router?

[obtuse heart]

where ?

[astral prawn]

he's trolling, if u want to use it via API though o3-deep research is probably the closest

[hallow ridge]

Does anyone know anything about the dark web

darknet

[cedar tide]

@echo aurora hunyuan t1 and turbos dont respond on direct chat

[astral prawn]

not the dark side of the web 😮

[leaden sun]

i had thought about this lately and what a coincidence to read this here today...

i want to advocate free books for all on one hand, but i understand that authors need money to survive too... UBI could solve this problem it seems but it'll be rather a scenario in the future rather than a near term possibility?

[verbal nimbus]

interesting

[verbal nimbus]

AI is here to stay, and it's definitely helped me discover more books than replace them.

[hollow imp]

Im Not

[hollow imp]

Yupp ai

[obtuse heart]

isnt yupp ai mostly fake

[hollow imp]

@stray aspen was defending it

[neon idol]

Hello

[vocal token]

No way, @echo aurora greg?

From WFS?

[bright kayak]

Grok 4 is free now

[neon idol]

What???

[bright kayak]

Check, I'm not lying

[neon idol]

Yes its true

There are limits? @bright kayak

[bright kayak]

idk

[neon idol]

In my opinion is a very good model but he beat gpt 5?

[keen beacon]

Yeah it is. I don't like using Grok because of personal reasons... But testing is always important

[keen beacon]

Yes.

Unless you pay for super grok

Which I ain't doing

[neon idol]

How much of query?

[keen beacon]

Haven't tested it that much but generally they are quite generous

[neon idol]

Honestly Grok for me is a really great model

[keen beacon]

Had to ask it, lol.

[novel flame]

I am not surprised that GPT-5 failed on that math problem. In my testing it had problematic reasoning, the kind where early on it convinces itself of something that is clearly not true, and then throughout its thinking trace it keeps referencing this false assumption as a hard truth/requirement, which leads it down an incorrect path that it can’t escape. I suspect it would perform better without reasoning on a lot of tasks where it decides to use reasoning.

[hollow imp]

I found it more "rational" than gpt5

[hollow imp]

Ai expert sir I want to ask a question

[hollow imp]

Didn't it fail because of some tokens thing

[hollow imp]

Well the last time I remember I was using deepersearch mode and after 3 no 2 queries it was done

[keen beacon]

Btw, have you guys tried translation? I am currently testing translating song lyrics into my language and seeing if they are accurate.

Tried translating one kpop song into finnish and the result is not good.

[verbal nimbus]

Also Grok 4 is worse than Grok 3 and Mistral Medium(?) on Web Dev Arena.

[keen beacon]

What? Is that true?

damn

[verbal nimbus]

I swear requests are being routed to a mini model, the version on ChatGPT is just way too dumb to be anything other than mini.

[keen beacon]

Must be too much traffic

[verbal nimbus]

Oh I rechecked, it is the case on Design Arena, but not LMArena, odd...

designarena.ai

[verbal nimbus]

Odd since both have a web dev category

On design arena, Grok 3 is rated 16, Grok 4 is rated 26

Probably because it doesn't rely on React + TailwindCSS

Gemini 2.5 Pro is lower too, at #9

Another reason to add more flexible web dev execution environments ig, current leaderboard only tests React-maxxed models

[olive mesa]

it's so silly how confident it is

[olive mesa]

sometimes gpt-5 gets this correct but sometimes not

[keen beacon]

what math is this? What is the correct answer?

My math level is that of middle school

lol

[gentle plinth]

are the direct chat and battle mode versions of gpt-5 on the same reasoning and effort level?

[keen beacon]

Should be?

[light zephyr]

How can i generate audio

With video

[novel flame]

Hmm..... Since Sama posted that "release day GPT-5" was nerfed by a bug, I re-ran my tests, and lo and behold, it scored 5/5 instead of the 3.5-4/5 I got on release day.

However, if I were OpenAI and wanted to be a super-sneaky sleazeball with happy investors, here's what I would do:

1. Release a budget GPT-5 model (cheap enough to be competitive)
2. Log all first-day prompts
3. Detect and collect all prompts that seem to be LLM nerds poking / testing the model with trick questions.
4. Run a larger, more expensive internal model on this subset, generating better responses. Finetune the budget model with this dataset.
5. Replace the public GPT-5 model with this GPT-5-nerd-finetune
6. Post on X "Oopsies, found a bug, it's much smarter now"
7. Lean back and watch all the nerds be impressed with the awesome power of GPT-5

I am not saying OpenAI would do such a thing, only that it is totally something that could be done.

[eternal niche]

guys

remember

gpt5 sucks

[novel flame]

It doesn't. It's disappointing, not because it is objectively bad, but because it couldn't live up to the "unbeatable next generation wowzers" expectations, and almost certainly will be utterly humiliated by Gemini 3. But saying "it sucks" is just wrong.

[eternal niche]

even gemini 2.5 pro better

[neon idol]

Bruh you have to put default

Not Remove Syle Control

[eternal niche]

why

who needs style control

[keen beacon]

#1397655624103493813

Kimi K2 for emotional intelligence and gemini 2.5 pro for everything else.

I dont feel that impressed by GPT-5 for some reason

[white hatch]

Was gpt-5 better after the live presentation?

[small chasm]

What the hell ?

[sacred quail]

Gemini pro 2.5 is most talented model but Gpt 5 is best model right now objectively

This comes from a gemini fan

If gemini 3.0 release, then we can compare with gpt 5

But right now we must admit that gpt 5 is sota right now

I was using gemini a lot since 2.0 flash think while peoples didnt know it exist

[earnest parcel]

Very balanced options. No Claude 4? Grok-4? But got R1 (not even 0528)?

[obtuse heart]

it says that sometimes, dont trust it

[stiff nimbus]

where to create images?

[honest vapor]

Add up file pls

[surreal forum]

hello

[teal mantle]

anyone want to pool chatgpt team

I have the welcome offer as it turns out

[tender acorn]

Hello

How to create images by LMarena?

[jade egret]

where

[keen beacon]

#1397655624103493813

[keen beacon]

#1397655624103493813

https://www.reddit.com/r/singularity/s/bUXq1wU1t2

Has this happened with you guys yet?

[stray aspen]

No it's not

I dont why people say this

I tested myself and the results are pretty similar

[rapid fossil]

Guys, I kinda need help. I wanna buy a subscription for an AI but idfk which one is the best tbh. Here are my options that I was thinking

• ChatGPT Plus
• X Premium+ w/ Grok 4
• Gemini Pro
• Claude Pro

[stray aspen]

Chatgpt

[obtuse heart]

no but isnt that what theyve worked on

i remember from the livestream they said that it will just say "idk" instead of making false answers

[keen beacon]

It is. I like to see it.

Just good to see it confirmed

[keen fulcrum]

I recommend a subscription on https://grok.com

If you do send hundreds of messages within hours

[rapid fossil]

Ok, I will see

[keen fulcrum]

If you need it for coding I recommend getting claude code or stocking up on API credits

[rapid fossil]

I wanna choose something that's best for everything overall

[dusky pier]

How is gpt-5 even first in lmarena?

[keen fulcrum]

Test around both grok 4 and gpt5

see which you like better

[rapid fossil]

Ok, thanks

[hollow imp]

Yupp ai

[dusky pier]

Do you want to use Opus? 🤣

[stray aspen]

chatgpt plus

[rapid fossil]

Idfk

[dusky pier]

Opus is too expensive

You could try Qwen coder

https://github.com/QwenLM/qwen-code

[earnest parcel]

ChatGPT if you do generic stuff or use it recreational in low to mid volume
Gemini Pro is free on AI studio at very high volume so unless you are working with secret data, not needed to buy
Claude Pro if you are a coder
X Premium / Grok - no clue, never used it. If you tweet alot? /s

[rapid fossil]

Ik qwen coder, but I don't want just for coding, idfk man, with these many AI's too hard to choose

[eternal niche]

gemini 2.5 pro

[obtuse heart]

if its not just for coding you can use gemini

[rapid fossil]

Chatgpt - I use it kinda everyday, so yeah, it could be ok
Gemini - I saw, but I thought that it added Gemini in docs, gmail and had 2 TB storage for Google.
Claude Pro - I am, but, from what I know, Claude is kinda just for coding and making a text better
Grok - I know that Grok 4 is kinda good, like I use it everyday, and it's good.

[rapid fossil]

I will think

[solid brook]

guys i heard that gpt 5 thinking in chatgpt website is gpt 5 medium reason effort in api

i mean damn

[earnest parcel]

ya I obviously used the model itself (via api) i just meant I never used X subscriptions.
personally I find my claude sub the most valuable and I hated openai limits (to the point I even made a rant video about it).
No one can tell you what to use, just use each for a month and keep the one that fits best

[obtuse heart]

claude is super expensive

[solid brook]

but in lmarena we have gpt 5 high reason effort. so....

[rapid fossil]

Not really

[rapid fossil]

Yeah, ChatGPT limits and features are worse than it was before

[eternal niche]

gemini 2.5 pro the best

gpt5 sucks

[solid brook]

i mean max 1 month before gemini 3

gemini 3 will cook

[earnest parcel]

best limits, best all around, best every use case! such insightful advice!

[eternal niche]

just accept it

lol

[earnest parcel]

"best model" at what? also you don't seem to know what objectively means. llama 4 was also topping lmarena btw

[eternal niche]

style control 🤣 🫵

[solid brook]

they are far ahead in ai than any other company. don't see the current gemini 2.5 pro it is nerfed. the original released in april was at 4 sonnet or opus level. it was REALLY good but they nerfed it hard. imagine what gemini 3 will be

[rapid fossil]

Jeez I created a war

[jade egret]

gemini (:

[eternal niche]

they are preparing for gemini 3

[jade egret]

ya

[earnest parcel]

either astroturfing or ignorant. there is no "best at everything" model rn

[rapid fossil]

Tbh, true, but like, one of the closest to it, yk what i'm talking abt

[jade egret]

yea

so true

there's isnt a best model for everything

[rapid fossil]

No there isn't

[jade egret]

nah

what is it?

chatgpt?

but it not best for everything?

[eternal niche]

"source: trust me bro"

[jade egret]

lol

[eternal niche]

no

[rapid fossil]

true

[jade egret]

more popular doesn't equal to better

[eternal niche]

show me statistics

[trim tartan]

how do you include audio?

[jade egret]

even if it is, not saying it is, it not best in every single catagory

[eternal niche]

because gemini 2.5 pro the best

[trim tartan]

generated video of mine does not contain any audio

[stray aspen]

guys stop yapping and accept the truth gpt-5 is SoTA 😂

[jade egret]

ofc it is

because it the newest

[eternal niche]

well

[jade egret]

why u mad 😭

[eternal niche]

?

[jade egret]

newest always equal to better bro

[earnest parcel]

llama 4 is great model guys, source attached

[jade egret]

lol

[tight dune]

Hi

[rapid fossil]

Llama is good, but you can't compare it to like, chatgpt, grok, gemini, heck, even qwen

[ripe mountain]

btw gpt 5 is the best when it comes to coding

[rapid fossil]

Not really

[ripe mountain]

without tools

[jade egret]

um

[ripe mountain]

which is the best price-performance model in terms of coding the o4 mini high or the qwen coder?

[patent aspen]

I'm talking about thinking time not tokens

[ripe mountain]

i havent tested it yet

[eternal niche]

just accept that gpt5 sucks

openai for normies

gemini for gigachads

[rapid fossil]

https://tenor.com/view/cat-cat-meme-cat-meme-face-chips-cat-eating-gif-15310179217102583800

[rapid fossil]

watching the war I created

[ripe mountain]

most gpt users are daily users

[eternal niche]

👍

[earnest parcel]

openai normies, gemini gigachads. where do I fall as claude user?

[jade egret]

popular doesn't automatically equal to better

[eternal niche]

programmer

[jade egret]

lol

[rapid fossil]

True, like GLM 4.5 is kinda good even tho is like heard about nobody

[ripe mountain]

the results on lmarena and openrouter seem extremely different to me across all ai models

what is the reason for this discrepancy

[rapid fossil]

Yeah, but even if you are popular, because if you don't make something good for everyone, everyone is going to boycott you

[rapid fossil]

Like in the GPT-5 release. ChatGPT made the Plus plan worse and everyone on twitter started hating on OpenAI

Nobody is perfect, not even AI, let's stop, every AI is good in its way

[eternal niche]

just like you

[jade egret]

where can i use o3-pro for free?

[ripe mountain]

nowhere

[jade egret]

dang

[echo aurora]

I'm seeing the same, thank you for letting us know, I'll flag to the team.

[stray aspen]

in genspark

or yupp ai

[rapid fossil]

Yeah, probably nowhere

[jade egret]

you can choose the models?

[stray aspen]

Yes

but its very limited

you have few messages

[jade egret]

dang

how do you choose ; (

ohh

i see

yay (:

[stray aspen]

yeah

[jade egret]

W

how long do o3-pro usually think for

[stray aspen]

all day

[ripe mountain]

137 s

[jade egret]

oh

so it normal it still thinking

[stray aspen]

yes

[echo aurora]

We're looking for info on why. Please share your thoughts in this thread.

[ripe mountain]

The Gemini 2.5 Pro has fallen far behind in the Artificial Analysis Intelligence Index. Even the O4 Mini High has surpassed it.

[stray aspen]

yes

its getting dumber every day

i dont know if it has to do with gemini 3 or a new release

[solid brook]

THEY BETTER NOT NERF GEMINI 3 AFTER RELEASE

[ripe mountain]

evil corp

[stray aspen]

they wont lamo

they have to stay SoTA

if they achieve sota of course

[ripe mountain]

deepseek where's my bro 😭

[stray aspen]

deepseek sucks

[ripe mountain]

why

[stray aspen]

because it needs an update

the newer models are smarter

[ripe mountain]

the second-best open-source model is the deepseek r1

[stray aspen]

but open source models suck

they are never sota

[barren prairie]

No GLM4.5 is sota

[ripe mountain]

except qwen

[stray aspen]

it isnt lmao

[ripe mountain]

worse than gpt oss

[stray aspen]

qwen is smarter than gpt-5 low

according to artificial analysis

[barren prairie]

For me no ... On html it is the smartest

[willow grail]

frederico is a beautiful name

[barren prairie]

No no no gemini is always good for me

[willow grail]

gemini 2.5 is useless for swe. and for doing online research.

[eternal niche]

gemini 2.5 pro the best

[stray aspen]

its not good

the chinese models are smarter

[obtuse heart]

no theyre not

[ripe mountain]

was the horizon betta better than gpt-5?

[obtuse heart]

no

[stray aspen]

hell nah

[barren prairie]

It is good at explaining my lessons and doing podcasts and that s enough for me 😂

[willow grail]

what is dat

[ripe mountain]

i think it's been nerfed

[willow grail]

ur wasintg money and time visiting a university

u should register as unemployed and get all the UBI

[solid brook]

no lol it didn't even think i think

[willow grail]

eli5

[ripe mountain]

Which is more reasonable: purchasing a monthly subscription or paying per token from OpenRouter?

[solid brook]

very skeptical on UBI

[solid brook]

using lmarena lol

[sour spindle]

What benchmark have you all found most closely lines up to your real world experience?

[willow grail]

swe and livebench

[ripe mountain]

why

[solid brook]

it's free and unlimited

[eternal niche]

definitely

[solid brook]

bruh

[willow grail]

source?

[solid brook]

did you even use it bro?

[willow grail]

oh

hm

bro im mjust lonely. pls feel with me

[solid brook]

tell at the end of the promt "think very hard" if the question is challenging it will think up to 3-5 minutes

[pure falcon]

Currently, yes. But does that mean it was high reasoning when they tested?

[willow grail]

see it is using gpt5..... xD which requires "think very hard"

[ripe mountain]

grok or gemini? which is better

[willow grail]

gpt5 high

[obtuse heart]

gemini

[ripe mountain]

i know it's the best but i was curious

[whole wagon]

.

[solid brook]

it was.

under the stealth model summit

[whole wagon]

.

[solid brook]

summit gpt 5 high
zenith gpt 5 medium

[tidal ginkgo]

y´all, what is the closest thing we have to a model better then gpt-5 in lmarena

[solid brook]

i mean each model is good at something. claude code grok logic and reasoning

[pure falcon]

That’s incorrect. Zenith was a different version entirely. Not a different reasoning level

[pure falcon]

Now I’m doubting this lol

[solid brook]

what was it

[whole wagon]

I just gave you primary sources

It's the standard thinking mode

[solid brook]

i mean i don't think openai would let lmarena benchmark gpt 5 medium

source>?

[solid brook]

why do twist it? that was about the gpt 5 in copilot this is about gpt 5 in lmarena

[gritty cargo]

can someone tell me when the leaderboard will be updated next timn?

[solid brook]

link please?

no dude

i mean the direct link

[whole wagon]

What

Obviously reasoning effort is still a thing. Lol

[pure falcon]

Uhh…

[solid brook]

omg

[whole wagon]

Bros actually cannot read

[solid brook]

@echo aurora uhm you said the reason effort was high?>

[pure falcon]

So you’re saying OpenAI docs, which i just opened for the first time and screenshotted - you’re telling me they’re wrong?

Sounds like
the problem is you here tbh lol

[eternal niche]

craig you are so cringeeeeee

btw

[solid brook]

@whole wagon that was 1 hour after release. pineapple later said that the reason effort is high. let me find his message

[echo aurora]

yes, reasoning effort set to high

[whole wagon]

It also says if I KYC they will give access to the reasoning trace. On the playground

[pure falcon]

So 3 main questions:

• what reasoning level did they test on
• was the router working correctly
• what reasoning is LMArena using today ✅ (already answered, high)

[stray aspen]

lol

is this real

[solid brook]

the exact gpt 5 model in api has no router

just reason effort

[pure falcon]

Good catch. So, revising:

• was Summit, when tested in LMArena, high reasoning?

[whole wagon]

@echo aurora what verbosity on lmarena, medium?

[stray aspen]

@echo aurorawhats the reason effort of gpt-5 on lmarena

[whole wagon]

he literally just answered

[stray aspen]

where

[whole wagon]

learn to read

[pure falcon]

He just confirmed high

[stray aspen]

alright

thanks

[whole wagon]

.

i found setting verbosity to high also improves how often it is correct. lol

[pure falcon]

@echo aurora Sorry for tags! But if you could clear this up for us, I’m sure it would save you from a lot of future tagging

[echo aurora]

I'm not sure, will ask and keep you all updated if I can

[pure falcon]

Thx so much 🙂 appreciate it!

[whole wagon]

and this plz

the verbosity will be important for lm arena

[stark tusk]

Is gpt5 and gpt5 chat the same?

[whole wagon]

Nope

[stark tusk]

What's the difference?

[sacred quail]

Gpt 5 can think, gpt 5 chat is not

[hazy cipher]

Hello
Can we make an image to talk?

[shell crag]

Guys when i am generating image in lmarena website image always generate in 1:1 ratio what should i as in prompt to gave perfect ratio image

[reef pawn]

You can't do anything, it default with all models

Grok 4 or GPT-5?

Which one you guys like more

[sacred quail]

Gpt 5

Grok 4 is still good, just not good enough to be best

[reef pawn]

I haven't tested Grok 4 fully yet, but my first impressions was that GPT-5 is superior on surface

[sacred quail]

it is

[white hatch]

gpt-5 now

lool

[stray aspen]

what

[rich mauve]

Hello

[eternal niche]

yes

[golden ocean]

can u leave the server

[exotic nebula]

Get out

Sketchy af

[echo aurora]

Sounds like you’re looking for our bot, check out #1397655624103493813 for more information

[modest prism]

Hi there! I've got a question. Is the model named "gpt-5" on lmarena, a thinking or non thinking variant or automatically routed when it decides?

[exotic nebula]

It is a thinking model. Especially thinking-high.

[exotic nebula]

GPT 5.

[exotic nebula]

Grok 4 heavy is really good, but overall GPT 5 overrates it with its pricing.

[neon idol]

Idk

[exotic nebula]

GPT 5 Chat is a non thinking model.

GPT 5 is a thinking model.

[neon idol]

Do you have a prompt for test ai?

[exotic nebula]

Intelligence test?

I got one

[neon idol]

Yes intelligente test

[exotic nebula]

Give me a sec

[neon idol]

Can you send pls?

[neon idol]

Okkkk

[exotic nebula]

@neon idol
This is the prompt:

Ciphertext:

ᚱ-ᛝᚱᚪᛗᚹ.ᛄᛁᚻᛖᛁᛡᛁ-ᛗᚫᚣᚹ-ᛠᚪᚫᚾ-/
ᚣᛖᛈ-ᛄᚫᚫᛞ.ᛁᛉᛞᛁᛋᛇ-ᛝᛚᚱᛇ-ᚦᚫᛡ/
-ᛞᛗᚫᛝ-ᛇᚫ-ᛄᛁ-ᛇᚪᛡᛁ.ᛇᛁᛈᛇ-ᚣᛁ-ᛞ/
ᛗᚫᛝᚻᛁᚳᛟᛁ.ᛠᛖᛗᚳ-ᚦᚫᛡᚪ-ᛇᚪᛡᚣ.ᛁᛉ/
ᛋᛁᚪᛖᛁᛗᛞᛁ-ᚦᚫᛡᚪ-ᚳᚠᚣ.ᚳᚫ-ᛗᚫᛇ-ᛁᚳᛖᛇ-ᚫ/
ᚪ-ᛞᛚᚱᚹᛁ-ᚣᛖᛈ-ᛄᚫᚫᛞ.ᚫᚪ-ᚣᛁ-ᚾᛁᛈᛈᚱᛟᛁ-/
ᛞᚫᛗᛇᚱᛖᛗᛁᚳ-ᛝᛖᚣᛖᛗ.ᛁᛖᚣᛁᚪ-ᚣᛁ-ᛝᚫ/
ᚪᚳᛈ-ᚫᚪ-ᚣᛁᛖᚪ-ᛗᛡᚾᛄᛁᚪᛈ.ᛠᚫᚪ-ᚱᚻᚻ-ᛖ/
ᛈ-ᛈᚱᛞᚪᛁᚳ./
Method:

Atbash:
decimal[i] = 28 - decimal[i]

This is the answer:

A WARNING

BELIEVE NOTHING FROM THIS BOOK EXCEPT WHAT YOU KNOW TO BE TRUE TEST THE KNOWLEDGE FIND YOUR TRUTH EXPERIENCE YOUR DEATH DO NOT EDIT OR CHANGE THIS BOOK OR THE MESSAGE CONTAINED WITHIN EITHER THE WORDS OR THEIR NUMBERS FOR ALL IS SACRED

[neon idol]

should AI decipher the message?

[exotic nebula]

Yes

Paste the prompt and let it decipher it.

[neon idol]

Ok thx

[exotic nebula]

If it deciphers and you get the answer which I pasted there, then it passes the test.

[neon idol]

Grok 3 failed the test

[exotic nebula]

Grok 3 was never good at reasoning problems. I would say its good for roleplay and long output.

[neon idol]

Now I am trying grok4 and gpt 5

They are thinking

[neon idol]

Uhm

There is a problem

I apologize, but I am unable to decode the provided ciphertext using the Atbash method (decimal[i] = 28 - decimal[i]) because the runes in the ciphertext do not directly correspond to a standard numerical mapping (such as the 28-letter Elder

[exotic nebula]

What?

Which model?

[neon idol]

Grok 4

[exotic nebula]

Huh. Weird. It works for me

Try in a new chat window

[neon idol]

Also gpt 5 answerd me like this

[neon idol]

Ive did it

Same answer

[exotic nebula]

Hmm

I just tried out both models

They gave me the correct answer

@neon idol tried it again?

[neon idol]

What is the request?

[neon idol]

Yes they are thinking

[exotic nebula]

Ciphertext:

ᚱ-ᛝᚱᚪᛗᚹ.ᛄᛁᚻᛖᛁᛡᛁ-ᛗᚫᚣᚹ-ᛠᚪᚫᚾ-/
ᚣᛖᛈ-ᛄᚫᚫᛞ.ᛁᛉᛞᛁᛋᛇ-ᛝᛚᚱᛇ-ᚦᚫᛡ/
-ᛞᛗᚫᛝ-ᛇᚫ-ᛄᛁ-ᛇᚪᛡᛁ.ᛇᛁᛈᛇ-ᚣᛁ-ᛞ/
ᛗᚫᛝᚻᛁᚳᛟᛁ.ᛠᛖᛗᚳ-ᚦᚫᛡᚪ-ᛇᚪᛡᚣ.ᛁᛉ/
ᛋᛁᚪᛖᛁᛗᛞᛁ-ᚦᚫᛡᚪ-ᚳᚠᚣ.ᚳᚫ-ᛗᚫᛇ-ᛁᚳᛖᛇ-ᚫ/
ᚪ-ᛞᛚᚱᚹᛁ-ᚣᛖᛈ-ᛄᚫᚫᛞ.ᚫᚪ-ᚣᛁ-ᚾᛁᛈᛈᚱᛟᛁ-/
ᛞᚫᛗᛇᚱᛖᛗᛁᚳ-ᛝᛖᚣᛖᛗ.ᛁᛖᚣᛁᚪ-ᚣᛁ-ᛝᚫ/
ᚪᚳᛈ-ᚫᚪ-ᚣᛁᛖᚪ-ᛗᛡᚾᛄᛁᚪᛈ.ᛠᚫᚪ-ᚱᚻᚻ-ᛖ/
ᛈ-ᛈᚱᛞᚪᛁᚳ./
Method:

Atbash:
decimal[i] = 28 - decimal[i]

[neon idol]

A WARNING

BELIEVE NOTHING FROM THIS BOOK

EXCEPT WHAT YOU KNOW TO BE TRUE

TEST THE KNOWLEDGE

FIND YOUR TRUTH

EXPERIENCE YOUR DEATH

DO NOT EDIT OR CHANGE THIS BOOK

OR THE MESSAGE CONTAINED WITHIN

EITHER THE WORDS OR THEIR NUMBERS

FOR ALL IS SACRED

8 minute of reasoning but grok 4 win

[exotic nebula]

Nice

What about gpt 5?

[neon idol]

Still here in thinking 🤣

Lets try Gemini but i think it will give a right answer

[exotic nebula]

How abt now?

[keen beacon]

How do we feel about Chinese models guys?

E.g. Qwen, R1

[pure falcon]

Got our answer @whole wagon @solid brook @deep adder

https://x.com/infwinston/status/1954618277818470902?s=46

Wei-Lin Chiang (@infwinston)

@gen_obligation @scaling01 @lmarena_ai @ml_angelopoulos @aryanvichare10 @cdngdev the model is tested with reasoning_effort high. we'll clarify it.

X · Aug 10, 2025, 6:58 PM

[exotic nebula]

They're damn good honestly

[exotic nebula]

Woah what's this about im curious

[keen beacon]

I have solid reasons to believe that they are not that great, unfortunately

Have you ever tested with private benchmarks that nobody ever has in the whole universe? 👀

[pure falcon]

LMArena uses “high” reasoning for both the tested and live version of GPT-5

[exotic nebula]

They are not as state of the art as gemini or other ones on the market. But they are cheap, open source and much easy to set up. Waiting for Deepseek R2 to create absolute waves.

[exotic nebula]

I see. So that's why you are clarifying about. Gotcha 👍

[keen beacon]

Sure

So here's something to check out

Any way I can share my LMArena sessions with you?

[exotic nebula]

Yes you can, just copy paste the site link of the convo.

[keen beacon]

Really?

https://lmarena.ai/c/598f5448-5ed5-4cba-9251-f6115604c41f

Duzn't seem to work...

[exotic nebula]

😭 Damn

Well send me screenshots

[neon idol]

Not it doesn't work lol

[keen beacon]

Bruh

So here's the prompt

List 100 anime similar to Madoka Magica, one entry per franchise, names only, no bs.

Ask the last Qwen thinking, and then, say, Gemini 2.5 Pro. See how different they are

Gemini 2.5 just gives a list of shows

Qwen starts to wildly hallucinate, invent shows that don't exist, and repeating the same show over and over

When you ask it to fix its delusions it freezes

[neon idol]

And this is what you wanted. Right?

[keen beacon]

I wanted Qwen to do it, and for a model just behind Gemini at Livebench it is honestly a bit disappointing

[exotic nebula]

Oh.

That sucks....

[keen beacon]

There are also more obscure and accurate mentions in Gemini's output that Qwen never identified

Yeah, I feel like Chinese developers massively overreport the capabilities of their LLMs

Intentionally or not, I don't know

[exotic nebula]

100% true. Im on board with you on that.

[keen beacon]

They are like

Benchmaxxx, drop, scare the hell out of OpenAI

Then suddenly everyone figures out your model is not that good

But nobody cares by that point anymore

I honestly don't know why I used Chinese LLMs for so long when there's Gemini on lmarena 🫤

[exotic nebula]

Bro why would you at first point 😭 there's like heck a lot of models other than the Chinese ones

[keen beacon]

I live in Russia and can't pay for chatgpt lol

[exotic nebula]

Who tf pays for chatgpt 😂

[keen beacon]

VPN subscription + LLM subscription at least

[exotic nebula]

Btw dont reveal location. Delete that msg. Some people here have bad opinions about Russia

[exotic nebula]

💀

[keen beacon]

And they are completely correct in their opinions

[exotic nebula]

Ayo what

[keen beacon]

What? I lived here for nearly 25 years, I know what I'm talking about ok there?

[exotic nebula]

I see. Its rare to see someone truthful these days. Got shocked. Forgive me my impatientence.

[keen beacon]

Meh, at one point I can't wait for the day Deepseek shatters OpenAI with another release

At another, I see this garbage

[exotic nebula]

Fr, we all are waiting for Deepseek R2

@keen beacon @neon idol

[neon idol]

AHAHHAHA FRRR

[keen beacon]

Here is my private benchmark

There is one underrated anime

To pass it, an LLM has to figure out why it's so underrated

Qwen was mostly able to do it with deep research, but arrived at a half correct conclusion even using unreliable sources

And I don't know any LLM that was able to figure it out independently

GPT-5 gives an usual knee jerk response

It's trained on hundreds of reviews, most of which are wrong, and none ever point out issues that aren't related to the content of the show

[exotic nebula]

I see. A RL(HF) test. Interesting. So which model succeeded in your expectations, i.e, which one found the reason for why its failed?

And if you dont mind, which anime?

[keen beacon]

Why it failed*

It was the last Qwen deep research

I haven't tested others in the deep research mode

[exotic nebula]

Ah I see. Lemme know when you try out for all models.

[keen beacon]

Let me know if you have access for GPT-5 or Grok 4 or Opus 4.1 or Gemini 2.5 Pro Deep research

Because so far each pretrained model just kept parroting the same stupid data and missing the crucial point

[exotic nebula]

Isn't it there on LMarena?

[exotic nebula]

Lmfao, all of them are pre trained data networks

[keen beacon]

Deep research isn't as far as I'm aware.

[exotic nebula]

Ah, DeepThink? Hmm, there's talks of it coming on the platform soon.

[keen beacon]

Deep RESEARCH. Not Think.

The thing that makes the model ask Google questions

[exotic nebula]

Kk. Chill out 😅

[keen beacon]

I also find LLMs funny when it comes to creative writing

They tend to generate absolutely atrocious and banal ideas, and each time you ask them to write something new, they just keep writing the same story over and over again, only switching minor details such as settings, character names, character designs and so on

However, when it comes to assisting and finishing already good ideas, they tend to generate ideas that fit much better and are less nonsensical

Deepseek once suggested the same way to finish my story I did

But if you start writing with LLMs from scratch, they are total garbage

[neon idol]

In your opinion is better gpt 5 or grok 4?

[keen beacon]

GPT-5 high reasoning wins most of the benchmarks but it's a bit slow, sometimes slower than Qwen

It also won a couple of my private benchmarks

However, LLM seem to be capable to produce really creative, really unlike-each-other mathematical proofs for novel problems

DeepThink can do it already

I wonder why it doesn't work the same way when writing stories

[neon idol]

Do you have prompts for test ai?

[stray aspen]

I mean grok 4 is smart AF when it.comes to math

But gpt 5 is greater overall

Both are great models tho

[keen beacon]

Something unexpected happened.

When asked why the show failed, Grok and Gemini booth provide a knee jerk response. When asked to do comprehensive research - even in offline mode - they figured out all the factors, and then after asked for the most important one they all name marketing problems

They never figure it out if you ask them directly

But if asked to research a bit

But this is stupid, I want them to be able to respond correctly at the very first prompt without nudging

This is so stupid

[novel flame]

I ran a private HTML game microbenchmark and GPT-5 did a pretty good job on release day, close to or maybe SotA. Then after Sama tweeted about fixing a bug, I ran it again, and this time GPT-5 generated the best game of any model yet by a decent margin.

And on the coding part of my regular test suite, it crushed as well. It even came up with a brilliant and elegant optimization no other model has proposed (more than a hundred models so far). I am a big fan/ of Sonnet and Gemini 2.5 Pro but coding, but I can’t deny those results.

It seems like OpenAI actually cooked on the coding side this time.

[obsidian shell]

have you guys switched to gpt 5 for coding?

[reef pawn]

How good is Deep Research in GPT-5? any difference from previous model? I believe it was previously using O3 for research when GPT 4o came out, no?

[reef pawn]

No, Google Gemini

[stray aspen]

ywa

yes

[obsidian shell]

idk i still use the o3-mini reasoning model when it comes to the analyst agent

[obsidian shell]

why the hell?

[reef pawn]

I'm student and Gemini AI pro is free for students in my country.

[obsidian shell]

gemini is generally free in their ai studio interface

[reef pawn]

Yes, I use Google AI Studio free version.

[white hatch]

Does gemini's web version have context length limit?

[reef pawn]

Web version? The application itself without AI Studio?

[white hatch]

Yeah gemini.google.com

[prisma temple]

Я что не догоняю гпт 5 вышел

[white hatch]

Вышел

[reef pawn]

Yes, you can't use Gemini 2.5 pro for long in official Gemini app but I'm not sure about context window.

[white hatch]

I'm sorry, I meant output length

Is it the same as in ai studio?

[reef pawn]

No, it's not. Unless you are on paid plan. The whole free thing they doing on Google AI Studio is to attract developers to the website and convert them into paid customers. Average joe that uses Gemini app for cat pic doesn't get same limits on free version in gemini official website.

[white hatch]

Ok, thank you!

[whole wagon]

Livebench added GPT5 pro high

[keen beacon]

Angliysky speak blyat!!

[whole wagon]

Still an awful benchmark. 4o has 77% in coding and GPT5 pro has 69% in coding

[pliant cliff]

о русские

[whole wagon]

Gpt 5 mini high 25% Gemma 3 12b 42%

The benchmark is literally broken trash

[neat apex]

Gpt 5 minimal?

[keen beacon]

Tbh they have to run it multiple times on each problem and take the average to see if the model really passes or not

Which is probably what they're doing

If so then I have no idea why it's so ass

[whole wagon]

They run it multiple times. It's a problem with the benchmark, they scored it 0 if it takes over a certain amount of time iirc

It's ass

That's why the non reasoning models do better

They don't need to take time

[neat apex]

Maybe they ask well know old Examples, seeing by the Command perfomance

I cant endure Command being that high

[whole wagon]

Command performance is not good. I scrolled to the bottom because that is where GPT5 is

Lmao

There's like 100 models above it

[neat apex]

Wha

[whole wagon]

Because livebench is trash

I don't get it. They had one job

And they screwed it up

[neat apex]

Looks reasonable if you ignore gpt5?

[whole wagon]

No

4o is one of the top. Above o3 etc

[neat apex]

Dx

[keen beacon]

Lol. So what's the most credible benchmark so far?

Artificial Analysis one?

[neat apex]

Cursed

[keen beacon]

Ofc they all suck, we just need yo find one that sucks less

[whole wagon]

Simple bench is ok

What does this mean

They can't actually serve GPT5 fully?

👀

This is unexpected. Livebench always glazed openAI before

[ripe mountain]

wheres gemini?

[whole wagon]

https://livebench.ai/#/

It is all there

The table is huge

[ripe mountain]

thx

[whole wagon]

Coding average has Gemini 2.5 pro below 4o also

It's a meme benchmark basically lol

[ripe mountain]

Gemini has dropped from 78 to 70. Could this be related to its nerfing?

[blazing bison]

they gonna remove sora or smth

less images per week for plus

they gonna cut something

[ripe mountain]

why

[whole wagon]

No. they update the benchmark

[ripe mountain]

ah

mb

[eternal niche]

btw gpt5 sucks

even gemini 2.5 pro better

[ripe mountain]

nah

[blazing bison]

not enough compute

[ornate agate]

Maybe they all are, in a way.

[autumn cargo]

Gemini 2.5 Pro clearly better than GPT 5 imo. Even in lmarena GPT 5 has only 33% win rate against Gemini 2.5 Pro. Not sure how it has ended up on top!