general

[sour spindle]

might be a google guy now

look what you've done openAI

[eternal niche]

it depends on what

[keen beacon]

Because they are pre-release models

[tired herald]

🙂

[stray aspen]

how

are you a LMArena dev or seomthing

[tired herald]

no

just having fun with extensions

it doesnt fully work right now tho

[native flame]

Hii, is the legacy site down??

[echo aurora]

I'm seeing this too. Will flag.

[keen beacon]

Well the code named models appear only randomly so uhm you can just hope that if you have a long convo, they will appear a lot.

[tired herald]

I just got a really good idea for generating code with models

[ocean vortex]

Good for you

[hollow imp]

Does the legacy site have any updates

Like any new models smth

[tired herald]

i dont think so

[hollow imp]

I really like the customisation parameters

[tired herald]

good point

[hollow imp]

Like when @echo aurora you guys changed into new lmarena I was asking gpt and perplexity about you guys

😂

[tired herald]

503 Service Unavailable
No server is available to handle this request.

damn

legacy dead

#ai-creations message

[echo aurora]

The legacy site doesn't receive updates sorry to say.

[tired herald]

oh

it

seems

file sending may break due to the limit of tokens per user message

Ah, who cares

okay, found another small bug.....

no

make it easier to generate code with the AIs, and easier to copy and view it

damn

thats ai for you

it guesses

intelligently guesses

no, the thing is that on LMArena, they dont have any system message at all

So they dont have a name per se

none of them is trustworthy

they profit by being friendly

[stray aspen]

google deepmind guy

hes nice

[keen beacon]

Demis 100 percent

[stray aspen]

agreed

[tired herald]

HAHAHAHAHHA

FINALLY

[eternal niche]

where is V.P.

[tired herald]

#ai-creations message

working system prompt

[eternal niche]

🙁

[tired herald]

Finally I can force them AIs to not roleplay as completely different models

yes my bruv

ChatGPT, trained by OpenAI

[high ginkgo]

Stop using that dumbass chat format

[trail creek]

why are you glasing gpt this much tf?!!

damn

[neon idol]

Ehat is the prompt?

And then how can I have the picture

Uhm idk how to do

Lol

I eill give you the answer

[tired herald]

well well well

gpt 5 mini high is so slow 😭

reasoning or no?

okay

[neon idol]

I have from grok 4

[stray aspen]

craig are you paying a claude subscription

[tired herald]

GPT-5 Mini High

Claude 4.1 Opus Reasoning

why do mine feel worse

[neon idol]

@tired herald bro if I dm you can you see with grok 4?

I have the svg

[tired herald]

Huh

okay

dm me

[golden ocean]

omg fraud detected

[tired herald]

Actually mine look more detailed

[tired herald]

Nani

I wonder why system prompts dont work on the mistral medium model

[neon idol]

Glm 4.5 is goodddddd

[scenic salmon]

👋

[tired herald]

Hi?

[scenic salmon]

Just joined, so yeah, hi, lol

[tired herald]

Go check out the other channels 👍

[golden ocean]

yeah go check them out NOW @scenic salmon.

[scenic crypt]

@echo aurora

[tired herald]

Huh 😭

[echo aurora]

[echo aurora]

@scenic crypt

[scenic crypt]

Is the problem solved?

[echo aurora]

Which issue are you referring to?

[misty vault]

gpt-4-0314 not being in the arena

[scenic crypt]

There is a problem with the video creation converter counter. It says that 8 videos cannot be created per day. Veo 3 Audio no longer appears.

@echo aurora

[echo aurora]

Yes, the daily amount for Video Arena has been fixed.

Veo 3 Audio no longer appears
This was changed. There is now Veo 3 with and without audio.

[golden ocean]

Russian state news

it does suck tho

wtf is this L by gpt 5?

obvious indication it is windows machine and provides linux commands

even gpt 3 dont fail that

[pure falcon]

If you post this one more time, I’m reporting you

[eternal niche]

ok

he said not to spam

i am not spamming

[pure falcon]

Posting the same thing 700 times is spamming

[golden ocean]

ayo please dont attack lmarena with russian botnet

[trail creek]

real

[echo aurora]

We've taken action, let's move on please

[white hatch]

😭

[trail creek]

https://tenor.com/view/the-warn-button-looks-very-useful-the-mute-button-looks-very-useful-ngl-ban-warn-mute-gif-14828387595483272930

[obsidian cargo]

theory: the stealth model 'toad' is deepseek v4

[stray aspen]

oh damn

they banned the gemini ragebaiter

[misty vault]

Huh wdym? paws is still here

paws day 1 without gemini 2.5 pro😔

[inner gate]

I prefer grok 4 over gpt 5 and 2.5 I was wondering if anyone else doed

[hollow imp]

Doed?

[wicked root]

Hello, LMArena ranking aside, would you guys recommend GPT5 (subscription model) for coding complex (1500+ line) projects?

[marsh stratus]

depends on what your alternatives are

Ideally you'd have multiple models ready to go

[wicked root]

I've been using gemini pro exclusively but I get rate limited every day

[tidal ginkgo]

guys

i think i might never use LMArena again

[marsh stratus]

That's honestly pretty boilerplate

[tidal ginkgo]

is my personal info in risk for using this service?

[leaden palm]

this looks like boilerplate google analytics

if you don't like analytics, use a browser with tracker blocking or get ublock origin

[marsh stratus]

as someone who uses both Gemini and GPT, it's nice to have them cover each other's weaknesses

[tidal ginkgo]

yea i will use tracking blocking and vpn

[inner gate]

DOES

[wintry tinsel]

What do you use it for that makes you prefer it

I prefer Claude opus for everything not, math/logic/advanced reasoning

[wintry tinsel]

Use brave it’s got automatic tracker blocking

[tidal ginkgo]

nah i already installed the thing this guy said

[wintry tinsel]

That’s fine but brave is goated it’s got everything built it, 0 bloat ware, and many other useful features

[inner gate]

I usually just use it as a companion I’m not that techy. I don’t prefer it because of intelligence I more meant the actual app lol. I do like how it responds tho

[wintry tinsel]

Built in Tor as well

[inner gate]

General things

[wintry tinsel]

Bruh are you using Ani

[inner gate]

ANI?

[wintry tinsel]

Yeah

[inner gate]

Idk what that is

As a companion I meant when I have a question it’s my go to

[hollow imp]

Eww tobi how will you execute the PLAN like this

[inner gate]

💔 I’ll ask it for ideas

[wicked root]

@marsh stratusI think I'm in love with claude. Gemini's good but I keep getting rate limited, Grok is... special in a 'stay away from me by at least 5m at all times kind of way' but it froze when I asked some complex coding question, but Claude re wrote my entire code with just a single prompt.

[viscid sun]

why chash the LMarena chat history ? how to I solve this problem ?

[stark mango]

hello

every boddyy

[keen beacon]

Every social media platform knows a bit more about you than LMArena

lol

this is silly.

[tidal ginkgo]

well now i use tracking blocker

[keen beacon]

Everybody should use that when they install any browser

some dont even know ad blocking

[tidal ginkgo]

yea

[echo aurora]

Is your chat history now missing?

[echo aurora]

Unfortunately, if your chat history is missing you won't be able to get it back. This is a problem our team is looking to address both with new features but overall site reliability as well. I am sorry you don't have access to that history anymore.

[trail creek]

nano banana removed from the arena?

[echo aurora]

I don't believe so, what makes you say that?

[trail creek]

Its been very long since i got it in the battle.

don't lie to me yes it is

[golden ocean]

bro thinks hes bing chat

[keen beacon]

It's not removed

I just got it in my testing

[trail creek]

Idk why you are lying but it is hundred percent removed

im not schizo.

[golden ocean]

This is @echo aurora's reaction rn to

[golden ocean]

this message

[misty vault]

Or

I'm sorry, but I don't believe that's accurate. I think there may be some misunderstanding here. I'm still learning, so my assessment could be mistaken, and I appreciate your understanding and patience.🙏

[exotic stream]

Buddy your data isn't that important chill the hell down 🤣😭😭

Mf acting like the CIA is on his ass

[golden ocean]

No bro it's not that, i mean like what if the data of me gooning with ai girlfriend gets leaked and it can linked to my identity irl?!?! then no one want to hire me anymore : (((

like claude serious asf but damn he know how to .... well i can't say more due to the rules here :(

🤣😭😭

[stray aspen]

dont link your identity tot he AI

[golden ocean]

but LMarena does that now for us

[keen beacon]

Maybe they forgot you

[keen beacon]

The stuff is anonymized, even in direct chats, unless you provide your real name and other revealing info

[golden ocean]

other websites might have my name

[keen beacon]

if it concerns you, you can always use a VPN.

[golden ocean]

that was indeed my plan, but then this mean @exotic stream interrupted us 😡 im shaking and crying rn

[exotic stream]

If privacy concerns you that much just use a local model bro

[golden ocean]

not the mc video deleted too 😭

[severe stream]

Hey does anyone here what the limits for each model are or is it just completely free

[keen beacon]

There are some rate limits

especially if you put too many inputs in one minute aka rpm

[severe stream]

do you know how much it is

[keen beacon]

No, but the whole service is free.

just dont rapid fire the chat

lol

[severe stream]

Damn why is'nt this website popular then wth

[keen beacon]

Don't know, lol

[echo aurora]

Tell your friends!

[inner gate]

Have u tried advertising lol

[golden ocean]

am i your friend? 🥺

[echo aurora]

You didn't wish me happy birthday

[stray aspen]

is that you pineapple

[misty vault]

yes

[echo aurora]

I take it back

[pure falcon]

What! How did i miss this?? Happy bday pineapple!!

[echo aurora]

Lol not my birthday, thanks tho

[pure falcon]

Well, you deserve blessings for a wonderful day no matter what

[queen thorn]

is there an answer to why most of the models get all fucky and respond with that?
"Something went wrong while generating the response. Please try again." or is there maybe a temporary fix i could use to use my current chats so that the model doesn't lose the context of whatever it's working on?

[inner gate]

Hmmm

[blazing bison]

1 - This message generally means that you are rate limited
2 - About more context, prob no, bcs of costs

[vital lake]

@echo aurora
How do you guys get all these expensive APIs lol? Sorry for ping

[blazing bison]

Why don't use the discord search function?

Ppl ask it 999 times everyday

[echo aurora]

Yeah sorry to say sometimes models will error out for different reasons. We are working on making these error messages more clear about what went wrong

[vital lake]

What do I search lol?

[golden ocean]

the video arena ruined search discord function

[scenic salmon]

I assume they're granted API credits, if not near unlimited API access by all of the model companies since lmarena provides valuable feedback on what users like and don't like 🤷‍♂️

saves them from having to do the research themselves

[echo aurora]

Sry for delay, doorbell. We are paying for the usage. This blog post should be helpful: https://news.lmarena.ai/new-lmarena/

[scenic salmon]

ha, you got VCs to back you, I should pick your brain if you were involved in the raising at all @echo aurora

[vital lake]

LOL

Sounds to me like they want to build their own models

Using the data

[scenic salmon]

it's not the worst idea, they have more data on the type of responses preferred than any of the individual model companies

[vital lake]

And your getting high quality data from things like GPT-5 and Opus

Thats worth milions

[scenic salmon]

yeah, no synthetic training data needed, all real human questions

[scenic salmon]

$100M to be exact

[vernal oxide]

how to do i use bananaLLM

the uhh

nano banana image generation

[vital lake]

Btw how did Solar get so good?

I remember they were hot garbage

They fr locked in

[echo aurora]

Models that are using a codename can only be used in the Battle mode (random models being sampled side-by-side). Meaning you aren't able to select it specifically

[vital lake]

Do you guys charge a lower AI lab if they wanted to test their AI as a codename?

[rough brook]

hi guys what the best prompts for still frame, still camera no movement?

[true tusk]

HI

[unborn lily]

Hi

[opaque mirage]

hi

[hollow field]

Hi

[keen beacon]

wth

[wintry tinsel]

Countdown to Gemini 3💀

[rare python]

Is the proposal of direct chat for codename models disapproved or still on your todo list?

[tribal nymph]

Many models shown on the site aren’t the real ones or are just named versions that don’t exist (like GPT-5, Grok 4, etc.).
They are actually working on GPT-4 & Grok 2.

Kindly fix it ASAP

[keen beacon]

Where

Guys I have a question

Has anyone of you experienced a model becoming unable to answer some questions correctly after it was upgraded?

Like GPT-o4 answered some correctly, then o3 wasn't able to get it at all

I found that Deepseek and Qwen answer my questions correctly based on incorrect data (!) it was trained on what the heck

[unborn lantern]

@echo aurora Brother, the chat history issue is causing a lot of trouble. Please take necessary steps to fix it.

[random wolf]

I hope they put soon the "cancel button" it's so frustrating

any solution to stop the "generating"?

[solid brook]

What happened?

[keen beacon]

Evidence for that?

[queen thorn]

I understand, but it's weird since like it's not even related to what AI you're choosing on the website, once this error appears even changing it to a different model wouldn't fix that whatsoever, the chat itself goes to hell

until you're starting a new chat that is, and everything works well again, only if there was a way to maybe export the chat after it dies and maybe import it into another one so that you'll be able to feed the model context about what you're trying to build with it, instead of starting from zero and having to feed it information about what you were using it for in the previous chats

because it’s currently in a very unstable state, the website and the concept are cool, but that doesn’t mean much if your chats die after just three minutes of use

[verbal nimbus]

https://x.com/btibor91/status/1955241562486763962

[verbal nimbus]

For fairness, I think other reasoning levels should be included, not a configuration that even Pro users won't get to use (if X posts are accurate).

[keen beacon]

i believe toad is a qwen series model:

• it uses emojis in a similar style
• "the vibes"

neither do deepseek models use emojis for responses nor does gemini.

The only other model i've observed is chatgpt-4o-latest

[obtuse heart]

0 evidence

[viscid timber]

what happened to video arena 4

[ocean vortex]

There are some decent reasons for them to do this though. Reasoning effort changes are more impactful in cost for chatgpt than API, since it's often doing tool calls and that mostly happens while it's reasoning. They are fine going all out on API even for diminishing returns but for chatgpt this makes no sense if they get big increase in cost with little to no benefit. And then also, "128" technically still falls within high reasoning effort range for gpt5, since medium on API is 64.

[blazing bison]

The rate limits are for all models, not for specific model. And you know that every prompt you send is gonna be public right? Lmarena is for you to test models not for real work

[blazing bison]

Roon said that this juicy thing is not the reasoning effort

[neon idol]

Toad? What the hell is?

[blazing bison]

And that gpt thinking is using high reasoning

[ocean vortex]

Kindly educate yourself please. Just because a model responds it is based on some older variant, that is completely normal behavior. Models are not trained about themselves much usually, since this is not very useful for IRL applications of solving actual tasks or problems. Fine-tuning on redundant things like that will degrade performance everywhere else

[blazing bison]

Claude is trained to say which model they are btw

But others, no

[ocean vortex]

Yeah that's why I said "usually"

Most of the time models will get this info from a system message

[blazing bison]

Ye just claude do that

Just claude train to say which model they are i mean

Openai, grok, they use the system prompt generally

[ocean vortex]

link? 🧐

I'm pretty sure it is. There's a direct link between reasoning effort and that number for sure

[blazing bison]

I didn't saved it but it was in Tibor post

[viscid lake]

Chat gpt us still better at Pokémon

[fleet lintel]

Which company is Nano-banana? OpenAI ?

[keen beacon]

Maybe you could ask it for a different style?

[keen beacon]

Google.

[fleet lintel]

Does nano implies small model??

[keen beacon]

Well we can only assume that is the case. No one knows for sure

[viscid lake]

Maybe

[unborn lantern]

Is there any website like lmarena?

Anyone know?

[viscid lake]

First,banana
Second chat gpt
Third is imegen 4 ultra

[ocean vortex]

Tried to look for it and only found this. Which seems to confirm it being the reasoning effort rather than deny lol

[unborn lantern]

Is there any website like lmarena?

[ocean vortex]

Also I don't think juice number directly translates gpt5 vs o3. For o3-medium the number is 2X compared to gpt5-medium, but it only just generates more:

not by much I mean

[ocean vortex]

https://www.designarena.ai

[blazing bison]

Idk them, there is another post that he said that think harder on router can reach high thinking too

I think that it is this one

[hollow imp]

😭

[rocky mauve]

anyone know an ai that’s capible of making high quality gfx’s? Is it even possible yet

[hollow imp]

No it isn't at all

I've tried that a lot

Please don't waste your time like me for that

[ocean vortex]

that's different context though. Their router was the weakest link at the start and their model card technically describes it as being more significant than it was on launch. If they haven't improved that yet I'm sure they will. It honestly made no sense that it couldn't do more than low reasoning effort, especially now with it being renamed to "Auto" and describing it as "decides how long to think"

It should be able to do chat/low/medium (for Plus sub), the way they are presenting it

[blazing bison]

it sucks anyway

like o3

always better on api

[ocean vortex]

Then for Pro high as well (128)

[blazing bison]

their chatgpt prompt sucks and the infra or whatever they do on the backend sucks

and gpt-5-pro feel exctly like o3-pro, i'm sure they just changed the name

[ocean vortex]

I mean at the start when that was named "GPT5" it did actually perform better than gpt5 with no reasoning and that was fairly obvious. Now it's still a good option for people who are confused by all the models and just want to ask a question

[blazing bison]

now that gemini copy pasted their frontend

when they finish copy pasting the features too

it's gonna be better than chatgpt

[ocean vortex]

Nah no way. It is based on gpt5 now rather than o3. Different base model. It's like comparing gpt4o from last year vs gpt4.1

[blazing bison]

i retryed my old prompts btw

it did the same frontend

the same answer

[ocean vortex]

Try something different but still with visuals, you should see the difference tbh

it now ranks nr1 on webdev

[blazing bison]

yeah i tryed already to see the difference

[ocean vortex]

while before it was bad

[blazing bison]

there is none

the model that they deployed for testers on api was called o3-alpha too

xDD

[ocean vortex]

I haven't tried frontend, but for svg the difference is obvious

[blazing bison]

after that they changed it to nectarine

[ocean vortex]

yeah that's just random joke lol

they did it with gpt2-chatbot as well

[blazing bison]

idk about that

the model that won the math olympiada has o3-alpha on the name too

[ocean vortex]

They knew everyone would see that name, they did it deliberately...

[blazing bison]

it was leaked on the git

[ocean vortex]

Not to make it too obvious that it's gpt5 base I think

But still promote their stuff

[ocean vortex]

it was on lmarena and webdev

with that name

It was immediately obvious to everyone that this name alias is public

Besides, @blazing bison , those gains wouldn't be possible otherwise without changing the base model tbh. gpt5-medium generates less than o3-medium but performs considerably better

[winged burrow]

/video prompt

[blazing bison]

i believe that gpt-5 is another base model

just the pro version that is strange

as a heavy user of the pro version

i noticed the differece from o1 to o3-pro clearly

but from o3-pro for gpt-5-pro

almost none

[ocean vortex]

well then pro is different as well. Pro is literally just some code for the most part

[blazing bison]

and i'm talking about frontend in general

[ocean vortex]

And it is using gpt5

not o3

[blazing bison]

they took 3 months to update from o1-pro to o3-pro

but released gpt-5-pro on the same day

that was already sus asf

and

there is no gpt-5 pro on api why

it's like they changed o3-pro low to o3-pro medium now

[ocean vortex]

They were slow with o3-pro as well. I think it took them time to test and approve. You can probably get some unexpected results by running each prompt 10+ times and to make gains it is not super easy out the box. Small changes even to just the prompting can make a fairly big difference

[ornate agate]

lol. Isn’t pro just gpt high now?

[ocean vortex]

This time it may actually be capacity for real... They did big changes with new releases and are still finetuning the caps for subs. Pro API would nuke their gpus lol

[blazing bison]

like i said, i can't proof, but i'm not the only one pro user that feels that, everyone that heavy use pro models from openai feels the same

i think the real gpt-5 pro is gonna appear on chatgpt only after it's avaliable on api

[ornate agate]

I thought it was leaked that pro is just gpt high

[ocean vortex]

conspiracy theory much. This is extremely unlikely

[blazing bison]

gpt high didn't take 10-20 minutes to answer

[blazing bison]

well i have it, i'm using it, and the AI communites on x says the same

[ocean vortex]

No, pro is parallel test-time compute. As should have been immediatelly obvious but you can read their model card if you don't want to take my word for it. 👀

[blazing bison]

and there is a router on pro models on chatgpt too

it answer in seconds if your question is simple

[ornate agate]

Hmm. What is it then?

[keen beacon]

What? So 200 juice is reserved for API specifically, and even Pro users don't get this treatment? Is this a joke?

[ocean vortex]

In ChatGPT, we also provide access
to gpt-5-thinking using a setting that makes use of parallel test time compute; we refer to this as
gpt-5-thinking-pro.

[ornate agate]

Yeah this is what I was thinking of. It shows pro as juice 128

[blazing bison]

yes

[ocean vortex]

Notice that even when writing this they knew it's gonna be available at first only on chatgpt

not API

[keen beacon]

Why did they cap the juice at 200?

Why not 512 or 1024?

Why this exact number?

[blazing bison]

no different results

[keen beacon]

We can only wonder...

[ocean vortex]

@blazing bison I think it "feels" the same probably because they used much of the same prompting to compute ~10 attempts into a single answer. So when it follows their instructions, answers are to look similar as previous pro, by design

[ornate agate]

I have a theory on that. If you think about it a bit it’s obvious…

[keen beacon]

Context length?

[ornate agate]

Ye

[blazing bison]

no

it's bcs above 200 there is no difference or the model actually lose performance

[keen beacon]

So you want to say that at some point the context of reasoning becomes too long and the model starts to slop?

[ocean vortex]

No what. What evidence do you have to prove this absurd theory that they are using o3-pro in disguise marketed as gpt5-pro?

[ornate agate]

Yes. So 200 is the max cot length it was trained on. Above that is slop.

[blazing bison]

my gut feeling and my friends that actually have pro

idk what you're talking about on plus plan

[keen beacon]

But why just 200

[ocean vortex]

So basically nothing. It's difficult to test models properly even doing the right things, let alone 'gut feeling'...

[blazing bison]

if you run same prompts and have the same results

it's enough

[ocean vortex]

No it's not

o3-pro was not gpt3.5

[blazing bison]

it is

[ocean vortex]

?

[blazing bison]

now i have better things to do

bye

[warm fulcrum]

arguing about pointless things goes hard

• ur being ragebaited and u cant see it

🔥

[keen beacon]

By the way did you know guys that these stealth models drop their names if you ask them nicely

[vapid zinc]

Is there any specified schedule for when the leaderboards are updated?

[blazing bison]

🤓

[keen beacon]

I don't know what's the system prompt of LMArena that restricts them to share their names

But they will at least very generously share the information about their creators

[ocean vortex]

@blazing bison Also wait a sec, you are SERIOUSLY arguing that gpt5-pro is just a scam and that they are using o3-pro instead, based on your gut feeling and no evidence to distinguish between 2 very good models? You are ACTUALLY for real? 🤣 🤣

[warm fulcrum]

i wasnt siding with u

[ocean vortex]

LOL

[blazing bison]

they can call anything gpt-5-pro on their end

it's not a scam

[ocean vortex]

no they can't

it would be an obvious scam

[nimble trail]

It's technically is tho💀

[blazing bison]

it's not, on their end gpt-5 can be called o3-alpha and it's ok

[ocean vortex]

What gpt5-pro means is CLEARLY defined in their model card

no 2 ways about it

[warm fulcrum]

ur saying that openai will break federal laws

good point

[flint sandal]

Gpt5pro means that its gemini knockoff

[warm fulcrum]

👍

[novel crater]

dang you can use Claude 4.1 for free now too lmarena is awesome!

[keen beacon]

Imagine if OpenAI really scammed everyone with this move 💀

[nimble trail]

What do you mean it's okay 😭😭😭

[blazing bison]

?? they can call anything gpt-5-pro bro

[keen beacon]

You sound delusional and misinformed

[blazing bison]

i'm not saying that they are delivering a worst model

[warm fulcrum]

could you be more vague

[keen beacon]

To be honest it's not much surprising after the news about Sama being a pathological narcissistic liar

[warm fulcrum]

ur saying openai is doing illegal stuff

which they cnat

[blazing bison]

if they change reasoning effort of old o3-pro to high and call it gpt-5-pro it's ok

[keen beacon]

Lol.

[blazing bison]

they aren't breaking any laws

doing that

[warm fulcrum]

no

[keen beacon]

Yeah, nobody can do illegal stuff

[warm fulcrum]

because they market it as a better tool

[keen beacon]

They will goto jail

Its illegal

[blazing bison]

if the reasoning effort is high, the result is better

then it's a better tool

[warm fulcrum]

source: trust me bro

i guess gpt-5 is actually just gpt 2 fine tuned

we got played..

[blazing bison]

i'm not saying that it's actually it

no

i tryed it

with the same prompts

and got the exactly same results

[keen beacon]

Imagine if everything they did to gpt-5 was just scaling the reasoning up to 200 juice, once Gemini scales up to this number OpenAI will be cooked 💀

[warm fulcrum]

which platform did you try this in?

[warm fulcrum]

zenith was 64 juice

was better than anything atm

[wet sparrow]

I agree with him. I don't think the people arguing here even have Pro

There are even posts about it on the OpenAI forum

OpenAI said they are working on it

[keen beacon]

So here's the following dilemma

[keen beacon]

the hidden model in battle mode

You get GPT Pro with two stupid models running at once that kind of "increases" the odds of successful completion of a task

Or one smart GPT High

Two stupid models or one smart?

[blazing bison]

we don't really know what it is

or what it do

[keen beacon]

qwen:

[warm fulcrum]

"basically guys i think gpt-5 pro is the same as o3 pro bc my gut says so"

[wet sparrow]

Theo even made a video about it. It's not directly about Pro Models, but it is about all GPT-5 models. There is a problem with them

[warm fulcrum]

holy ragebait mother of 3

[wet sparrow]

OpenAI has already addressed this

[keen beacon]

it's trivial to jailbreak LMArena to make the models namedrop

[keen beacon]

? pls tell in detail

[keen beacon]

Up to you, hacker

But I can say that, from my experience, if it is a model that is worth attention, everyone will be talking about it

Be it Zenith, Toad or whatever

agreed

[keen beacon]

zenith was probably gpt 5

lmarena should probably reveal the hidden models after their public release

Toad is nothing special btw

[warm fulcrum]

zenith was the nectarine model

im only assuming because of what theo says

it matches with zenith's output speed and accuracy

[keen beacon]

ahh okay

[warm fulcrum]

not confirmed tho

[keen beacon]

yeah its meh

[keen beacon]

I meant that it is not by a major provider

[keen beacon]

nah its just not SOTA so IDC

[wet sparrow]

That's why I think you don't have the Pro model. If you had paid for it and didn't notice any improvements, you would be upset too

[keen beacon]

If it is by a major provider everyone'd already losing their crap here and there

[warm fulcrum]

💀

what are u yapping about

[keen beacon]

but i just saw it side by side with qwen 235 b n thought hey they loook similar

[warm fulcrum]

i have the pro model

[wet sparrow]

Show proof

[keen beacon]

Better follow openrouter for stealth models

[wet sparrow]

😆

[warm fulcrum]

wowie

i have gpt-5 pro it seems

i must be a millionaire

[blazing bison]

There is no point in discuss this btw

[hollow imp]

@keen beacon

I'm watching this

https://youtu.be/fHRS_NOs24w?feature=shared

[warm fulcrum]

india

[hollow imp]

This documentary even better than the movie

[warm fulcrum]

🔥

[keen beacon]

There are so many stealth models on LMArena and they are total garbage

[hollow imp]

@keen beacon

[keen beacon]

What?

[hollow imp]

What is a stealth model

What are they hiding from

[keen beacon]

Toad

[hollow imp]

Toad?

[keen beacon]

I don't know

[hollow imp]

https://tenor.com/view/gamatatsu-naruto-toad-jiraiya-gif-15336191263686460582

[keen beacon]

Yes, and also these Phantoms IIRC

They are nothing special anyway

[hollow imp]

Iirc?

[keen beacon]

If I Remember Correctly

[hollow imp]

What is IIRC

[keen beacon]

Most hype comes from stealth models from openrouter

[hollow imp]

Ohh

Bruh I thought

[keen beacon]

Here are the big things

[hollow imp]

IIRc is a

Ai term

😭

[hollow imp]

Can you give an example of a stealth model

[keen beacon]

Toad, Zenith, Nano-banana

[hollow imp]

I know nano banana

[keen beacon]

Me too

We went to the same school together

[keen beacon]

They all come from big startups and enterprises btw. There's no way someone can train a LLM in their garage, so think about something as big as Netflix

But it doesn't matter

None of them are SOTA

Claude, Gemini or Gemini no doubt

Gemini has genuinely impressed me today though

[ocean vortex]

@echo aurora make this into a channel. We have community-creations, community-polls would fit and be useful. 🙂

[keen beacon]

My favorite anime (it is NOT Madoka Magica) is often criticized for things that are present in 99% of other anime, but failed mostly due to marketing reasons

There are very few models that have identified this problem

Qwen and Deepseek often point it out, as does GPT-5

But Gemini was the only one that compared it to other shows and figured out that these criticisms do not actually matter

[queen thorn]

I'm fully aware that the prompts are public yes, so it's essentially a platform that lets you test different models that stop working about 2 minutes after using them?

[somber rover]

how to make video

[tired herald]

#1397655624103493813 message

[solid brook]

No i think only the claude opus model has limits

Because it is so expensive

[queen thorn]

nah man, it's gemini 2.5 pro, chatgpt 5, claude 4 sonnet, opus as well

3.7 claude

[solid brook]

You sure? After how many prompts did you encounter limits on lmerena?

[queen thorn]

one

look

[solid brook]

Bruh lol

[queen thorn]

one specific prompt

[solid brook]

I used gpt 5 high for hours

[queen thorn]

i'll give ya a prompt you'll see

[solid brook]

Okay send it

[solid brook]

That is not limit

[terse shuttle]

try to reload page

or clear cookie

[queen thorn]

i've done both

of these things

[solid brook]

Send prompt

[terse shuttle]

reopen browser

idk

[queen thorn]

i'll send you the prompt

hold up

other things work fine

whoops

wait

i'mma sent it in dms

[vernal oxide]

guys

how do i know if i got the banana llm

[keen beacon]

great!

[verbal nimbus]

Gemini is falling behind a bit when it comes to agentic abilities (hopefully the next model proves this wrong).

[tired herald]

the next Gemini will probably topple everything

[verbal nimbus]

nightride-on might be the next one

Not sure whether it's Flash though

[queen thorn]

can anyone in this chat try and use this prompt and tell me if it bricks a chat for any of you on any model?

    <form method="POST" action="{{ route('register') }}">
        @csrf

        <!-- Name -->
        <div>
            <x-input-label for="name" :value="__('Name')" />
            <x-text-input id="name" class="block mt-1 w-full" type="text" name="name" :value="old('name')" required autofocus autocomplete="name" />
            <x-input-error :messages="$errors->get('name')" class="mt-2" />
        </div>

        <!-- Username -->
        <div class="mt-4">
            <x-input-label for="username" :value="__('Username')" />
            <x-text-input id="username" class="block mt-1 w-full" type="text" name="username" :value="old('username')" required autocomplete="username" />
            <x-input-error :messages="$errors->get('username')" class="mt-2" />
            <p class="text-sm text-gray-600 mt-1">This will be your unique profile URL: {{ url('/') }}/username</p>
        </div>

        <!-- Email Address --> ```

[tired herald]

No idea, havent played at all with stealth models

[verbal nimbus]

I only encountered it once, it was pretty good.

[tired herald]

it does

[verbal nimbus]

It would be nice if there was a mode that tested tool use/ReAct loops more.

[tired herald]

its probably escaping the data and making the packet invalid

[queen thorn]

yeah it's weird as hell

[tired herald]

not really, thats how code just works sometimes

[queen thorn]

it replied with that, and after it did, it got bricked completely

[tired herald]

because it accidentally created a situation that made the LMArena website start going crazy

[queen thorn]

i'm trying to figure out if there's a way to fix a bricked chat

[tired herald]

i dont think so

[queen thorn]

nothing seems to be working lmao

[tired herald]

prob is impossible

but idk

ask pineapple

[stray aspen]

reload and try again

thats what pineapple told me

[queen thorn]

cleaned my cache, hard reload, cleaned the browser history

disabled all of my extensions

restarted my browser

try this prompt and tell me if it's working for ya

[queen thorn]

.

[tired herald]

Its not the browser

its LMArena

[queen thorn]

has to be an issue with lmarena

yeah

yeah

true

[tired herald]

i know this issue

[stray aspen]

that code is the prompt?

[queen thorn]

yeah copy paste it

into lmarena

[tired herald]

My Extension had an issue with it

[stray aspen]

which model

[tired herald]

All models

[queen thorn]

any of them

[stray aspen]

ok

[queen thorn]

could it be that it's happening because it's inside of that text code block?

that?

[stray aspen]

that prompt is broken

[tired herald]

no

[stray aspen]

theres something wrong with your prompt

[tired herald]

no

[white hatch]

nothing breaks for me

[tired herald]

the prompt is not the issue

[queen thorn]

HUH?

on a new chat tho?

[tired herald]

then you have luck? or you added something to the prompt

[stray aspen]

but everything works fine until i put that prompt

[tired herald]

its how LMArena handles messages

[white hatch]

Probably luck

[white hatch]

yes

[tired herald]

Can you show a pic

[queen thorn]

yeah idk man that's insanely weird

idk how??

[tired herald]

how very very interesting

[stray aspen]

thats GREAT

[queen thorn]

WHAT?!

[tired herald]

I think I also know why its working for you

[queen thorn]

gpt 5 high as well???

[white hatch]

yes

[tired herald]

who wouldve guessed

its the 1:1 the same issue I had with files for my extension

[queen thorn]

bruh

[tired herald]

say hi first for example

and then give

[queen thorn]

let's try that

[tired herald]

<x-guest-layout>
    <form method="POST" action="{{ route('register') }}">
        @csrf

        <!-- Name -->
        <div>
            <x-input-label for="name" :value="__('Name')" />
            <x-text-input id="name" class="block mt-1 w-full" type="text" name="name" :value="old('name')" required autofocus autocomplete="name" />
            <x-input-error :messages="$errors->get('name')" class="mt-2" />
        </div>

        <!-- Username -->
        <div class="mt-4">
            <x-input-label for="username" :value="__('Username')" />
            <x-text-input id="username" class="block mt-1 w-full" type="text" name="username" :value="old('username')" required autocomplete="username" />
            <x-input-error :messages="$errors->get('username')" class="mt-2" />
            <p class="text-sm text-gray-600 mt-1">This will be your unique profile URL: {{ url('/') }}/username</p>
        </div>

        <!-- Email Address --> 

[white hatch]

```php

[tired herald]

yes

thats the code block format

[tired herald]

You need to copy my message through the discord copying

try that

<x-guest-layout>
    <form method="POST" action="{{ route('register') }}">
        @csrf

        <!-- Name -->
        <div>
            <x-input-label for="name" :value="__('Name')" />
            <x-text-input id="name" class="block mt-1 w-full" type="text" name="name" :value="old('name')" required autofocus autocomplete="name" />
            <x-input-error :messages="$errors->get('name')" class="mt-2" />
        </div>

        <!-- Username -->
        <div class="mt-4">
            <x-input-label for="username" :value="__('Username')" />
            <x-text-input id="username" class="block mt-1 w-full" type="text" name="username" :value="old('username')" required autocomplete="username" />
            <x-input-error :messages="$errors->get('username')" class="mt-2" />
            <p class="text-sm text-gray-600 mt-1">This will be your unique profile URL: {{ url('/') }}/username</p>
        </div>

        <!-- Email Address --> 

[tired herald]

still not copying with discord bruv

``

`

[echo aurora]

Maybe. I'm like to try and keep channel lists light, and I don't think polls get used enough to justify their own channel atm.

[tired herald]

you need three of `

at the beginning

and end

works everytime for me

[tired herald]

yes

that

exactly that

copy it and paste

[queen thorn]

interesting

[tired herald]

its message handling stuff

[queen thorn]

that's so freaking weird

[tired herald]

🙂

[stray aspen]

thats GREAT

[verbal nimbus]

Interesting that GPT-5-High actually loses to Gemini 2.5 Pro the majority of the time

[tired herald]

well

it happens

[queen thorn]

welp, so lmarena

gotta fix some stuff

[tired herald]

well, LMArena is changing its shape slowly

[queen thorn]

new update?!

haha

[tired herald]

no

[queen thorn]

jkjk

[tired herald]

🙂

Im making another small extension to change the look of LMArena into more OpenChat style yk

[stray aspen]

can you put the model selector next to the iamge button

[white hatch]

yooooooo good luck

[tired herald]

Thats what im working on

right this second

very fun stuff

[verbal nimbus]

Actually it's kinda odd, because Gemini has a higher win rate against more models, yet is ranked lower. Green = higher win rate. Red = lower win rate. Yellow = tie/no data. Statistical paradox?

[queen thorn]

make it so that the prompts are gonna work by adding that

hello

😂

[tired herald]

😭

[queen thorn]

just kidding it's still not gonna do anything

[tired herald]

the + has an issue with me

oh no

[queen thorn]

make ai give it a quick upper cut back into place

[verbal nimbus]

@echo aurora How does this work?

[tired herald]

ai would fail miserably 😭

[solid brook]

Lol the filters turn off when adding ```

[tired herald]

oh what

no way

[queen thorn]

huh?

wait lemme try and use that on that website

[tired herald]

test it out

[solid brook]

Uhm but still the model itself has safety filters

[keen beacon]

Yea and it won't let me generate girls kissing

😠

[tired herald]

Not when you give it a really nice system prompt

[solid brook]

Yeah....

[queen thorn]

?

[tired herald]

filters....

[queen thorn]

yeah hold on

it's still responding it's funny

to me

[solid brook]

How to do that?

[verbal nimbus]

Did it fall out of the flex box?

[tired herald]

Prob because I turned it into a circle

[tired herald]

Not yet available for normal people

[solid brook]

You mean you're part of lmarena team?

Or connected to them

[random wolf]

how do I stop "generating"? it's been an hour, lol

[verbal nimbus]

That shouldn't change it

[verbal nimbus]

Did you refresh the page?

[tired herald]

no idea

[random wolf]

@verbal nimbus yes brother, I did.

is there any solution, to fix this problem? it's so frustrating.

[queen thorn]

do you think it's possible to save a chat that has that prompt that bricks it?

[tired herald]

just lemme cook well

[verbal nimbus]

Do new chats work?

[queen thorn]

because something weird just happened

[tired herald]

ups

[random wolf]

it's working. but the chat AI I used, knows all the information and important things I have

[queen thorn]

something very weird just happened @tired herald ya gotta help me understand that

that's insanely weird

so i've had a chat that was bricked right?

couldn't send any messages whatsoever

not even a hello

[tired herald]

yes

?

[verbal nimbus]

Ah yeah, that's annoying. Very odd that it hasn't timed-out.

[queen thorn]

i've responded with ``` and then wrote the whole code thing

and it legit worked somehow???

[verbal nimbus]

If you want the info, you can copy the whole chat out by manually highlighting everything and ask another AI to clean the raw paste and summarize it.

[random wolf]

I hope they will fix it

[queen thorn]

but now

[tired herald]

ok, I really dont know why its happening, thats really weird

@echo aurora

we need your help

[random wolf]

I will try bro. thank you! that's a big help

[queen thorn]

legit yeah, WE NEED ANSWERS!!!

haha

[tired herald]

dem

[solid brook]

Report in #1343291835845578853

[echo aurora]

Hey sorry I'm in the middle of something else and haven't been following this chat closely. Can you submit a bug report and TLDR everything so I can take a look/let the team know?

[queen thorn]

i'll do that right now

[white hatch]

btw, try clearing cookie files

[tired herald]

#ai-creations message

do the buttons look good enough?

[grizzled turtle]

if someone know how to stop generation? my generation with chatbot bugged, its just loads endlessly

for example, so that the chatbot gives an error or if possible in some other way

[terse shuttle]

nohow

[echo aurora]

Unfortunately, your only two options at the moment are to refresh the page (sometimes this works) or start a new chat. This is a problem we're aware of where models will continue to generate without ending.

[grizzled turtle]

unfortunately, restarting the site doesnt help, Ill have to make a new chat(

thanks for the answer

[glossy epoch]

Hello to all

[red sluice]

Why is style control not called "length control" instead?
Style control is confusing!

[inner gate]

Howdy partners

[echo aurora]

Really sorry to hear that it didn't work. The current state what happens when chats get stuck isn't great, and we plan to make changes to help with this.

[raven helm]

@echo aurora , I’m sorry to interrupt, I’d just like to ask a quick question; why is it that there is no direct chat on WebDev Arena?

[inner gate]

What’s webdev used for?

[patent bane]

is gpt-5-high on lmarena supports tools?

[terse shuttle]

nah

only chat or gpt-5-search in search mode

[terse shuttle]

react apps

but WebDeb doesn't hace direct or side by side mode 🙁

[hardy lion]

because it uses others things besides length, like lists, bold and markdown formatting

[ancient reef]

Any opinions on "folsom-0805-1"
I tried it on a logic prompt and it actually wasn't terrible.

[obsidian cargo]

I wish we could compare toad to zenith

[whole wagon]

“We have to make these horrible trade-offs right now,” he said. “We have better models, and we just can’t offer them because we don’t have the capacity. We have other kinds of new products and services we’d love to offer.” Sam in recent interview

[queen thorn]

BRO HOW IS NO ONE COMPLAINING ABOUT THIS?!

IT'S LITERALLY ALWAYS HAPPENING LEGIT ONE SINGLE PROMPT CAUSED IT

WHAT THE HELL

[whole wagon]

That's been a thing forever, seems no resolve

[queen thorn]

bruhhhhhhhhhhhhhhhhhhhhhhhhh

[stray aspen]

wdym lmao

its probably the most common complaint

[queen thorn]

and still no solution? not even a temporary fix?

[stray aspen]

no

[queen thorn]

bruv

[stray aspen]

lmarena team is working on it

[tired herald]

uff

[gentle plinth]

even gpt-oss-120b is better then the minimal effort gpt-5 imo

[queen thorn]

Something went wrong while generating the response. Please try again.

[ocean vortex]

gpt5-minimal < gpt4.1

[hollow imp]

https://cdn.discordapp.com/attachments/1261350936769724548/1405882476232048752/image.png?ex=68a07196&is=689f2016&hm=34908bdcd463852a3199f976aab3472fe8319abd9a0d1a815cae7a2dc0f73df0&

[ocean vortex]

but to make this more confusing, this is also likely true:
gpt5-chat > gpt4.1

[gentle plinth]

it depends to which model you get routed

[gentle plinth]

anything is better then gpt5-minimal xD

[ocean vortex]

gpt5-minimal is probably only in the API

[gentle plinth]

it uses that if it "thinks" that the question doesnt require much thinking

which is like in 99% of the cases i feel

[brittle furnace]

Something went wrong while generating the response. Please try again.

[ocean vortex]

it uses gpt5-chat. That's direct replacement model for chatgpt-4o-latest. It's a different model from gpt5-minimal, this one has no reasoning at all so it most definitely performs better as reasoning wasn't the focus at all

[gentle plinth]

just use the prompt in direct chat, and paste it into react project:

[tired herald]

very cool

ill take this and use it as my system prompt

[ocean vortex]

instead of reasoning being something it was trained for that was then later taken away... It was fine-tuned from the get go to perform as good as possible without relying on reasoning

[hollow imp]

@gentle plinth pls pdf support noah 🙏

[gentle plinth]

wdym

i just asked it to generate a site with the prompt

and copied it

its what they are using in webdev arena apparently

[hollow imp]

What is webdev arena

[gentle plinth]

but i honestly cannot guarantee that it will work out of the box

bc they seem to have some specific project setup

but maybe ai can help you with that

[gentle plinth]

https://web.lmarena.ai/

[hollow imp]

What does it do

[gentle plinth]

it generates a website

with two random models

[hollow imp]

You are the ceo of lmaren right?

[gentle plinth]

and you have to say which one is better

[gentle plinth]

no

[tired herald]

whats even going on here

[gentle plinth]

the purple name is just a role i picked xD

[hollow imp]

Isn't creating a website very hard and you need wix for that?

[indigo hazel]

https://x.com/Alibaba_Qwen/status/1956399490698735950

Qwen (@Alibaba_Qwen)

🚀 Qwen Chat Desktop for Windows is here!
💻 All the power of Qwen Chat — now with MCP support for smarter, faster agents.
⚡ Run up MCP Servers, supercharge your productivity, and stay in control.
📥 Download now → https://t.co/uYQIIGQAJo

X · Aug 15, 2025, 4:56 PM

[gentle plinth]

ai can do some small websites

[hollow imp]

I'm curious

[gentle plinth]

for larger projects of course it can get to its limits

[hollow imp]

I'mma go play with it

What should I say but

[gentle plinth]

why would i install something that i can use in the webbrowser

[stray aspen]

qwen sucks lmao lol rofl

[hollow imp]

Where is direct chat

In webdevarenn

[stray aspen]

there isnt

[atomic stream]

Qwen is good over deepseek?

[stray aspen]

not at coding

[hollow imp]

I want it to show a photo in the website

[stray aspen]

but the image model from qwen is great

[atomic stream]

What are your thoughts on Kimi AI?

[stray aspen]

it sucks

its only good for writing

they need to make a reasoning version

[hollow imp]

How's Gpt 5 high

[stray aspen]

its amazing its SotA

[hollow imp]

Isn't Claude on top for that

[whole wagon]

@echo aurora is it possible to add different reasoning efforts for gpt models? Because we don't get high in the chatGPT app

[hollow imp]

What about gpt 5 search

[bronze urchin]

Do you have a good explanation for prompt injection these days? GPT5 compatible?

[stray aspen]

idk

i dont use search

[hollow imp]

Why

Mf pays for openai subscription

😡

[stray aspen]

because no

im not doing any reasearch

[bronze urchin]

Want to know if my company is safe

[stray aspen]

mostly coding

[hollow imp]

@stray aspen why don't you compare lmarena gpt 5 high and yupp ai

[stray aspen]

i dont needd to lmao

[hollow imp]

What company

[atomic stream]

Does Lmarea have any limits aside from opus?

[glass gulch]

[stray aspen]

lmarena is greater than yupp ai regarding usage of gpt-5 high

[errant cave]

What a crash holy crap

[stray aspen]

holy sigma

[hollow imp]

@gentle plinth helloo

[gentle plinth]

seems like gpt-oss-120b is around o3-mini level for hard prompts

[regal river]

GPT 4o better than GPT 5 wtf

[proud yoke]

How did they manage to make 5 worse than 4o 😭

[quiet dust]

Is it possible to switch between Auto, Fast and Thinking models on the phone in ChatGPT?

[proud yoke]

Is it even cheaper to run?

[glass gulch]

How recent is the leaderboard for the rest of you

[errant cave]

I think it's likely that OpenAI and lmarena came to some kind of agreement to make GPT-5 look better than it actually was on release

[hollow imp]

o3 search is better than gpt5 search I can say that with 100% surity

[stray aspen]

new leaderboard

[glass gulch]

i cant tell if the announcement implies it changed rn lol

[modest prism]

4o better than 4.5 ? How?

[errant cave]

Or maybe OpenAI just deceived lmarena before release

[gusty helm]

gemini still top of pack 😄 ?

[hollow imp]

@echo aurora 🙀

[errant cave]

4o was always top tier

[quiet dust]

Guys. Is it possible to switch between Auto, Fast and Thinking models on the phone in ChatGPT?

[stray aspen]

gpt-5 sucked on release

it was trash

now its way better

[tired herald]

still trash

[errant cave]

Still sucks

[gentle plinth]

people vote for it because of its sycophancy

[stray aspen]

no

its great

its SotA

[tired herald]

golden trash is still trash

[hollow imp]

GPT-3o

[stray aspen]

lol

[hollow imp]

SOTA

[errant cave]

I like GPT-5 clapping back against stupid ideas in contrast to 4o

That's like the only thing it's better at though

[atomic stream]

Yupp ai, is this free?

[tired herald]

how interesting

[vapid zinc]

i dont understand how ppl think 2.5 pro is so good

[stray aspen]

@tired heraldhow much are eyou gonna sell the plugin for

[modest prism]

OpenAI made people addicted to this model.

[hollow imp]

It's offering o3 pro for free

Such a scam

[stray aspen]

i mean its not bad

[hollow imp]

Gives you credits for using the ai

[tired herald]

im not selling, im gonna put up the code for free

[stray aspen]

but gpt-5 is way better

[tired herald]

prob on github

[hollow imp]

Does anyone use webdev arena

[vapid zinc]

yeah, not bad but grok 4, o3 pro, 5 pro are all noticably better no question

[stray aspen]

its not a scam lmao

stop yapping

[hollow imp]

It is

[stray aspen]

but right now lmarena is way better

[hollow imp]

100%

[stray aspen]

agreed

[hollow imp]

There's no better worse in 2+2

[atomic stream]

When I went there, it asked me for a login; thank God I have not.

[hollow imp]

@gentle plinth Vro hello

[gentle plinth]

stop randomly tagging me

[tired herald]

ups, LMArena doesnt like this at all

[modest prism]

The 03-25 variant was amazing. They just ruined it.

[gentle plinth]

or i will block

[vapid zinc]

I think its bc on lm arena ppl ask simple questions

[stray aspen]

i love how gemini 2.5 pro preview versions were better than the final lol

[echo aurora]

No. We wouldn't do something like that, ever. That'd go against everything we stand for and are trying to build.

[hollow imp]

How to attach a photo in webdev

[gentle plinth]

its not possible

currently

use the prompt i said in direct chat

[hollow imp]

No one has done it

[gentle plinth]

then you can

[hollow imp]

Give

[gentle plinth]

#general message

but as i said it will probably not work out of the box

ask ai to help you run it

[hollow imp]

It will not run on chrome?

Wdym

[whole wagon]

🙁

I want to see where the gpt I actually use in chat ranks

[modest prism]

Recently I use 2.5 flash more than 2.5 pro. DeepMind used some over-optimized technical changes to 2.5 pro, they do not know how to fix it.

[atomic stream]

What about Gemini and others? Have anyone got limts on it?

[errant cave]

It's gpt-5-chat I believe

[stray aspen]

on lmarena?

[echo aurora]

Sorry currently busy with other stuff, will try to address later

[errant cave]

Fifth place at 1427 ELO

[stray aspen]

i have only run into limits with claude 4.1 opus on lmarena

[atomic stream]

Yeah

[stray aspen]

as far as i know claude 4.1 opus has limits

[whole wagon]

chatGPT has thinking variants still

[stray aspen]

not sure about the rest

havent run into limits

[whole wagon]

They just don't use the high reasoning

[coarse glade]

You guys were lying in gpt 5 it really is gpt 4 you guys are lying to us

It’s really a bad thing guys

I have proof

[whole wagon]

Bruh

[coarse glade]

Look at this

And each gpt 5 model says this

[hollow imp]

If I paste the prompt into gpt-5 high, will that work?

[coarse glade]

See you guys were lying

[stray aspen]

the api doesnt know what model it is

[stray aspen]

its not stop yapping

theres no system prompt and the api doesnt know what it is