general

[blazing bison]

even today models can do like 20% of my job

[leaden palm]

some say that when you spend enough time with a model you begin to imitate it, even when you aren't using it

(see _opencv_ on x for instances of this)

[misty vault]

I can confirm

1 year after talking with sydney

[leaden palm]

have you been a good bing?

[misty vault]

yes

[bright kayak]

Why would they put a " symbol at the end?

[blazing bison]

wow i'm gonna follow this account

it's interesting

[leaden palm]

ah fair point, guess it's just 5 then

[inner gate]

What’s gpt oss?? How does it compare to gpt5?

[blazing bison]

i don't know now if my experience with sydney was so good like in memory

[bright kayak]

Lobotomized

[keen beacon]

Open source models

[echo aurora]

more info here: https://openai.com/index/introducing-gpt-oss/

[inner gate]

https://tenor.com/view/what-more-do-you-want-from-me-crying-emotional-what-do-you-want-gif-14843331

[inner gate]

Thanks

[whole wagon]

Those benchmarks are not representative of real world performance

Be warned lol

[leaden palm]

oh i saw an interesting thread about gpt oss, let me pull it up

[keen beacon]

And I heard it was only good in English too primarily. Absolutely awful in my native language.

[leaden palm]

it's truly fried

(full thread https://x.com/jxmnop/status/1953899426075816164 / https://xcancel.com/jxmnop/status/1953899426075816164)

jack morris (@jxmnop)

curious about the training data of OpenAI's new gpt-oss models? i was too.

so i generated 10M examples from gpt-oss-20b, ran some analysis, and the results were... pretty bizarre

time for a deep dive 🧵

X · Aug 8, 2025, 7:21 PM

[patent aspen]

Talking about OpenAI as some poor little startup that we shouldn't be so hard on for needing a small context window and long thinking times to benchmaxx is the weirdest complex

[bright kayak]

What i mean is, way too much safety, and the model itself isn't good for an os model

[inner gate]

Thought you said it to ME lol

Thanks

[whole wagon]

Simple bench guy had a funny story. When he posted gpt oss result openAI guys contacted him to rerun it and then they got an even worse result

He talked about it in latest vid

Titled gpt5

[tulip kindle]

Where GPT-5?

[whole wagon]

It is there

It is just not the top

So it's hard to see

[bright kayak]

Better question is wen gpt6

[cosmic salmon]

They've nuked it 😭

[bright kayak]

???

[tulip kindle]

No. They delete GPT-5 main

[whole wagon]

Pay the $20 for a non transparent router and 32k context window

Great deal

[bright kayak]

Same for me

Gpt 5 too expensive but grok 4 is fine

[tulip kindle]

Just strange if one time they get GPT-5 and after that delete

[whole wagon]

Well they couldn't keep gpt5 there because it's literally better than chatGPT plus for free then

Because you are always on the full model when selected in LM arena

Feels like plus tier regressed so much....

At least you could select o3

[cosmic salmon]

Although on medium reasoning by default as far as i know

[echo aurora]

Hmm, okay flagging, thank you!

[patent aspen]

Could also just be capacity issues

[clear spear]

wait where did gpt-5 go?

[patent aspen]

Oh I thought it was chatgpt5

Nvm

[clear spear]

WHY'D THEY DELETE IT!?

[echo aurora]

It's coming back

[clear spear]

oh okay 😊

[cosmic salmon]

Maybe GPT-5 were the LLMs we chatted with along the way...

[thin creek]

Okay

[misty vault]

Maybe the GPT-5 was the friends we made along the way

[misty vault]

😊

[whole wagon]

But why buy plus tier when GPT5 on lmarena is better than the router

[thin creek]

Is this Claude 4.1?

[whole wagon]

Yes

[clear spear]

Did you read it? 😄

kinda upset to why the models can't use search actively

[gentle plinth]

Bc all your data is going to be published eventually if posting it on lmarena, also no tools etc.

[clear spear]

all my data?

https://tenor.com/view/sweats-gif-25346666

[cosmic salmon]

lmarena actively spies on you confirmed

[whole wagon]

Ngl. People are still going to be putting personal info into LM arena they should at least set up a filter

[clear spear]

https://tenor.com/view/squidward-squidward-tentacles-spongebob-look-around-awkward-gif-496727432057798195

[whole wagon]

Before publishing the data

[blazing bison]

they do

[clear spear]

so even deleted chats?

[cosmic salmon]

Nah, you can have it, guys 😽

[blazing bison]

yes

[clear spear]

https://tenor.com/view/dead-my-honest-reaction-damn-skull-skeleton-gif-15885937969029652188

[clear spear]

I ain't got no money to pay for router

will someone give me their card info pretty please

https://tenor.com/view/smiling-cat-happy-cat-smile-gif-3555983607747631247

[cosmic salmon]

Just find one in the lmarena dataset, i guess 🚬

[echo aurora]

can confirm

[clear spear]

And how exactly would I do that 😄

[cosmic salmon]

Manually, i guess

[clear spear]

hey can you find me a credit card

[echo aurora]

hey no

[blazing bison]

lmao

[cosmic salmon]

welp, an attempt was made

[clear spear]

no fair

[blazing bison]

gpt-5 killed

rip

[echo aurora]

this was compelling

[echo aurora]

yeah there was an issue, working on bringing it back asap

[clear spear]

https://tenor.com/view/happy-happy-dog-dog-happiest-dog-super-happy-gif-17885812

[cosmic salmon]

It even disappeared from every single leaderboard for a second. Now it's back tho

[floral comet]

What does it mean?

[clear spear]

you dare to delete my message inferior being

[stray aspen]

no

its llama 1.1

[keen beacon]

@clear spear ay

Why is your pfp a

[echo aurora]

it's a dog

[cosmic salmon]

it's a dog

[clear spear]

furrypepe

OMG

IT'S HUGE

[echo aurora]

stap

[keen beacon]

It’s Ronaldo

suii

[clear spear]

do you enjoy watching men kick balls

cause I sure do 😄

[keen beacon]

Yes

[echo aurora]

kk back to AI topics please!

[clear spear]

No back to you secretly watching me

[keen beacon]

GPT-5 moment.

[cosmic salmon]

The consequences of not having access to GPT-5 on lmarena for a few minutes

[clear spear]

yeah I want it back NOW!

please

https://tenor.com/view/evil-smile-gif-1974600105454603239

[keen beacon]

It’s bad

[clear spear]

so far it does what I expect

[keen beacon]

4o was a better waifu

[cosmic salmon]

I like Chinese-provided waifus more

[clear spear]

wait does gpt-5 on lmarena have the ability to use specific models for different tasks?

@echo aurora

[cosmic salmon]

It's not a router, so probably not

[clear spear]

https://tenor.com/view/lewissaunderson38-lewissaunderson-cry-crying-cry-meme-gif-11499408157122608476

[keen beacon]

It’s the ultimate benchmark

[cosmic salmon]

Unless you mean different experts, but we don't know if GPT-5 is MoE or not

[keen beacon]

We all know it

Why is GPT-5 gone from LMArena

[cosmic salmon]

They killed it

Trying to revive it now

[clear spear]

It's taking a vacation

currently in the bahamas

[keen beacon]

Bruh

[clear spear]

what

is

that

about

[cosmic salmon]

Angy Copilot 😠

[maiden fulcrum]

where is gpt-5

i cant see it on direct chat

[clear spear]

CURRENTLY ON VACATION

[cosmic salmon]

I ate it

[clear spear]

WAIT 5 WEEKS

[echo aurora]

It's having an issue atm, team is working on a fix

[maiden fulcrum]

ok thank you very much.

[clear spear]

u gotta lie to them 🙄

[inner gate]

I think there’s problems with it

[echo aurora]

hey no

[inner gate]

Must be why

[misty vault]

In the meantime u guys will have to use @deep adder as replacement for gpt-5

temporary

[maiden fulcrum]

I dont like trolls

[clear spear]

well luckily I'm not a small rainbow rat living in a mushroom

[inner gate]

Is that what trolls are

🤔

[keen beacon]

Is there a limit to LMArena models? I’m a new to these kind of stuff

[cosmic salmon]

@deep adder

5rk1/4bppp/3R4/4p3/1p4P1/1N3P2/PPP4P/1K3B1R w - - 1 22

Your move

[maiden fulcrum]

@echo aurora is gpt-5 on lmarena using (high)?

[echo aurora]

Yup

[cosmic salmon]

Probably medium

[inner gate]

I don’t think so

You should ask pineapple for more details

[cosmic salmon]

Welp, i was instantly proven wrong

[stray aspen]

wheres gpt-5 normal

[keen beacon]

@echo aurora Please

[maiden fulcrum]

@echo aurora I cant delete my history chat. It always reappears

[echo aurora]

There are rate limits

[cosmic salmon]

Just delete your cookies completely lol

[echo aurora]

This was a bug; however, it was fixed yesterday, are you still seeing this?

[keen beacon]

Where can I read about them

[maiden fulcrum]

yep

[inner gate]

I thought I hit the rate limit for gpt5 when it stopped working lol

Then it got removed so I thought it was an error

[cosmic salmon]

Same

[echo aurora]

I don't believe exact info is shared somewhere

But I will raise to the team if that's something we want to start doing

[misty vault]

crack bench

[keen beacon]

Guess I’ll have to find out myself then

[inner gate]

There was one question I had. When an ai takes a while to reply is it Because of reasoning?

[echo aurora]

Nope! there should be error message

[cosmic salmon]

Yup

[stray aspen]

craigbench

[inner gate]

I see

[echo aurora]

I think depends, overall lag is an issue we're aware of

[inner gate]

Would it tell us directly if we hit the limit?

Or is it the general error message

[maiden fulcrum]

@echo aurora I am still seeing the bug

[echo aurora]

I believe we updated it so it says rate limit error

[inner gate]

Try deleting your history

That’s what I did

And whatever chats get removed from the history

[echo aurora]

can you try a different browser?

[inner gate]

Also gets removed on the website

[echo aurora]

let me know if it's still there

[inner gate]

Ah I see Thabks

[keen beacon]

I really wonder how the rate limit resets if there’s no account needed

[maiden fulcrum]

@echo aurora I cant see it on Edge

[cosmic salmon]

Probably IP-based or directed by your browser cookies

[keen beacon]

It would be stupid funny to be based on cookies

[echo aurora]

as in can't see the chats after deleted (meaning it's not a bug)?

[pseudo magnet]

wheres gpt-5 😞

[inner gate]

Removed for some fixes

[pseudo magnet]

it will be back??

[echo aurora]

sorry about that!

[inner gate]

Pineapple and his team wants you to have a great experience

[echo aurora]

yes! it will be back

[inner gate]

Yes it will

😉

[keen beacon]

I will not abuse it of course, I’m just curious

[maiden fulcrum]

no, it is still there in chrome

it wont go

[inner gate]

Delete ur history

[pseudo magnet]

thanks @inner gate and @echo aurora

[inner gate]

Today’s history

Try after that

After deleting ur history also close ur tab

So it refreshes

Then check again

[keen beacon]

What’s the point in deleting the chats the data is shared anyway

[maiden fulcrum]

???

[inner gate]

It might be because of OCD

[maiden fulcrum]

shared?

[keen beacon]

???

[inner gate]

I usually delete the chats I don’t use (I don’t have OCD) it just looks cleaner

I like it that way

[cosmic salmon]

I mean, on the legacy website your can just overwrite the model you're using by intercepting a join web request, since they aren't encrypted, so i wouldn't be surprised

[inner gate]

That’s how it improves I think

[keen beacon]

This stuff is free

[misty vault]

Can I buy you

[inner gate]

Isn’t that why ur data is shared?

Cus it’s free but not

[keen beacon]

Yep

[inner gate]

I thought so

[cosmic salmon]

If the product is free, then maybe we're the product... 😔

[inner gate]

Our data is the oroduct

Lol

[keen beacon]

Who cares.

[inner gate]

I don’t to be honest

[keen beacon]

It’s not like I’m gonna use more than 5-10 messages per chat.

[inner gate]

Lol

I sometimes ask them for car tunes and compare answers

[keen beacon]

I mostly use Opus for help when codex is too dumb.

[whole wagon]

4o coming back confirmed

[inner gate]

My Claude 4 opus usually stops working after a few messages

I haven’t used it recently

Maybe it’s been fixed

[keen beacon]

Elon said Grok 5 is coming before the end of the year

[cosmic salmon]

Opus is nice for agentic coding imo, but for fine-grade adjustments o3 is better

[inner gate]

I heard gemeni 3 is also coming this year

Gemini

[keen beacon]

Yeah, maybe.

[whole wagon]

At the end of the day what the people want is just a model to endlessly glaze them

All this benchmarks don't matter kek

[inner gate]

I read somewhere that ai models are made to sugar coat things to you and always think you’re in the right

[cosmic salmon]

I hope they increase the context window even more, because that's one of the best features Gemini has over competitors

[inner gate]

U can tell it u murdered someone and it’ll give you excuses on why you had no choice

[keen beacon]

Yeah.. I want to switch to Claude Code so bad but OpenAI is giving me free tokens to have fun and Opus is too expensive

[cosmic salmon]

Man, i wish i had free money from OpenAI and didn't have to use lmarena and openrouter's free models 😔

[inner gate]

What grok4 model is used on LM arena?

[keen beacon]

I share my data for free tokens.

Who cares, it’s not like they don’t already know all my weaknesses

[inner gate]

You get free tokens from openly sharing ur data?

[keen beacon]

Yeah.

[inner gate]

I did not know that

[keen beacon]

Look, I’m gonna show you.

[inner gate]

Go ahead

[atomic adder]

I swear gpt 5 just disappeared

[cosmic salmon]

Please don't swear

[keen beacon]

When I press that i it should’ve show me the limits but it’s stupid on mobile, so here are the limits.

[atomic adder]

im betting gpt 5 just disappeared lol

[keen beacon]

Ignore its gpt4.

The real models are here.

[atomic adder]

bro how do you do that lol

[keen beacon]

Do what

[atomic adder]

sacrifice your personal data for free tokens lol

[keen beacon]

Who cares

[jade egret]

lowkey it kinda fun talking with gemini 2.5 pro

[cosmic salmon]

Why

[stray aspen]

i love gpt-5

[atomic adder]

you clearly because you're the one that does it

[keen beacon]

I share my data, that means I don’t care about it

[misty vault]

girlfriend simulator

oh wait... that's sydney

[cosmic salmon]

Where's Four-Door Dodge Charger Daytona as a selection option?

[atomic adder]

well true at least you get something out of it lol

[keen beacon]

Can you show the cars?

Or at least one, I’m curious about the models

[jade egret]

idk

[misty vault]

no

[unborn lantern]

where is gpt 5? Lol

[keen beacon]

Here

ASI

Be amazed

[unborn lantern]

Not in lmarena

[jade egret]

gemini tweaking

[keen beacon]

Ah

[keen beacon]

What are you trying to do

[jade egret]

idk

prob role playing lol

[stray aspen]

why is gemini hallucinating apache 2.0 licenses and copyright crap in its python scripts

[jade egret]

his prob

[keen beacon]

You gave him PTSD

[jade egret]

LOL

and btw

this is the rest of the message

[vale estuary]

What happened with GPT-5 ?

[jade egret]

wdym

[keen beacon]

Is this through LMA?

[stray aspen]

lol its so high

[vale estuary]

Its not there in the site anymore

[jade egret]

no gemini official website

[jade egret]

lol

[sacred quail]

if you have a pro plan(not max) it happens sadly. Claude wants from you to buy max or pay api

[keen beacon]

You know about the ai studio right?

[jade egret]

ye

[cosmic salmon]

https://tenor.com/view/fnaf-memes-connection-terminated-fnaf-security-breach-fnaf4-fnaf-gif-25323747

[jade egret]

my gemini is high

i will name the chat "high gemini"

[cosmic salmon]

that's what happens when your reasoning effort is too high lol

[inner gate]

𝖮𝗁 𝗐𝗈𝗐 𝗈𝗄𝖺𝗒

[jade egret]

lol

[keen beacon]

Anyone played with verbosity parameter?

From my quick research it’s just a yap meter

[sacred quail]

4o lovers would love the current gemini btw. Praising you for doing breathe

[keen beacon]

Wonder if it makes any difference in code

Beside the explanations

[stray aspen]

@keen beacon

[vale estuary]

GPT-5 isn't there anymore?

[cosmic salmon]

Yeah, it drowned in the amount of requests per second

[keen beacon]

Bro I thought GPT-5 made you a browser game with those cars

[stray aspen]

its roblocks

[vale estuary]

It will be fixed?

[keen beacon]

Roadblocks

[inner gate]

𝖧𝗈𝗐 𝖽𝗂𝖽 𝗎 𝗀𝖾𝗍 𝗍𝗁𝖾𝗋𝖾 𝗂𝖽 𝗅𝗈𝗏𝖾 𝗍𝗈 𝗌𝗁𝖺𝗋𝖾 𝗅𝗈𝗅

[cosmic salmon]

Most likely, yeah

[wheat onyx]

nah fake, already did this. someone else did with nano and it worked too

[stray aspen]

the system routed it to gpt 1.5

[wheat onyx]

so it's beyond using the wrong model?

ah

[buoyant dust]

where is gpt5

[keen beacon]

Search for OpenAI Platform, Settings, Data Control and Sharing, you need to verify ID if I remember right and it’s not for everyone free, you have to be lucky.

[cosmic salmon]

Broke all the bones

[inner gate]

𝖨 𝗌𝖾𝖾

[wheat onyx]

but that was pervasive. it works for me, and it worked for others using nano

[keen beacon]

Most people get it, so try your luck, you have nothing to lose.

[inner gate]

𝖳𝗁𝖺𝗇𝗄𝗌𝗇𝗆𝖺𝗇

[keen beacon]

What

New strawberry problem

Same I get a satisfaction

[stray aspen]

lmao

[inner gate]

𝖨 𝗀𝖾𝗍 𝟩 𝖿𝗋𝖾𝖾 𝗐𝖾𝖾𝗄𝗅𝗒 𝖾𝗏𝖺𝗅𝗌

🥲

[keen beacon]

Unlucky

[stray aspen]

if it got it right it means this is all a government conspiracy

[keen beacon]

I don’t know what that 7 free evals does.

[wheat onyx]

i've heard that not adding the extra 0 decimal can cause issues (but shouldnt)

5.90-5.11 works, even though 5.9-5.11 should always work too

[inner gate]

No idea either

[keen beacon]

The X is shy

[stray aspen]

got it wrong

[inner gate]

Mahbe it means I get gpt pro for free

7 days a week

[keen beacon]

lol no way.

[wheat onyx]

could be model selector

[inner gate]

Wait

[wheat onyx]

theres nano, mini, normal

[inner gate]

If it lets me evaluate models for free

7 times

Idk

[keen beacon]

@wheat onyx see

5 gets it wrong.

[wheat onyx]

no. gpt 5 is multiple models

and you don't know which it is

[stray aspen]

however if you ask this

it does it right

[keen beacon]

Is this Claude?

[wheat onyx]

no

[keen beacon]

Or Copilot

[wheat onyx]

web gpt

[stray aspen]

its copilot

[keen beacon]

I see

[stray aspen]

yeah it works

[wheat onyx]

they really have to fix the autorouter

[keen beacon]

AGI when

You know what

I want ASI

I want to focus on myself more

Let the AI do everything

[prime mulch]

I also want asi

[stray aspen]

wdym autorouter

[wheat onyx]

they choose the "correct" model

also something they have to fix - if you have a REALLY long conversation, it basically crashes

[clear spear]

WHY IS MY GPT 5 NOT BACK YET

[tidal ginkgo]

why is gpt-5 gone

[frail ibex]

Same here. No GPT-5. Just mini and nano

[tidal ginkgo]

maybe bcz of costs

[slow sail]

too much use?

[tidal ginkgo]

maybe

used it a lot here

[slow sail]

i guess pineapple can say

[storm needle]

@echo aurora

[tidal ginkgo]

GUYZ WHY IS MY BEST FRE MODAL NOT HERE YATTTTT!!1!1111111!!!!! I WANT FRE STUF!!!11!!1

[keen ferry]

too much hype around gpt 5 and too many people use it without actually using the actual vote feature id say

[echo aurora]

Yeah GPT-5 isn't available atm

[slow sail]

dont just ping them man

[echo aurora]

team is looking into

[tidal ginkgo]

ok

[tidal ginkgo]

is it bcz of the hype?

[floral comet]

There's o3 now

I noticed it's a lot faster now, do the performance degrade?

[vital lake]

😭

[mental fern]

It takes too long to delete a chat. The deletion doesn't seem to complete until the page redirects to the homepage. Also, if I try to delete multiple chats at once, the process often fails and none of them get deleted. Is anyone else experiencing this?

[solid brook]

They are trying to fix it

[proper spindle]

Hey guys. Where is GPT-5?

[wet sparrow]

people burning tokens

[tidal ginkgo]

lol

btw it back

[woven turret]

we have gpt-5 and gpt-5 chat?

[stray aspen]

gpt-5-chat is live

[stray aspen]

they are different

yeah

[woven turret]

im confused that what are they used for

[random wolf]

everyone I need help. why is it always saying "Something went wrong with this response, please try again."???

[blazing bison]

rate limited prob

try again some hours later

[tidal ginkgo]

gpt-5 chat???

what is even the difference?

[blazing bison]

this is the gpt-5 in chatgpt

this model is finetuned for chat

[tidal ginkgo]

oh so the gpt-5 is more "pro"

[blazing bison]

yes

no

actually

[tidal ginkgo]

?

[blazing bison]

one is more nicer with the answers the other no

[tidal ginkgo]

ok? still don´t understand lol

[blazing bison]

one gonna say, nice, you are smart, good job

the other not

[tidal ginkgo]

ok

so chat is nicer lol

[blazing bison]

yes

[tidal ginkgo]

or does it only have more temp?

[tidal ginkgo]

oh ok

[blazing bison]

no it's completly different model

[tidal ginkgo]

ok

[blazing bison]

finetuned for chat

[errant thorn]

which is smarter

[tidal ginkgo]

same ig

[blazing bison]

the difference is clear

[tidal ginkgo]

oh

[blazing bison]

same intellect

just the format of the answer change

[tidal ginkgo]

just dumbed down

that

[blazing bison]

i would use gpt -5-chat more btw

for day to day things

[tidal ginkgo]

for coding im using gpt-5

[blazing bison]

yes

for coding gpt-5

[tidal ginkgo]

for everyday chat i use gpt-5-chat

[blazing bison]

yeah

exactly

[tidal ginkgo]

cool

[blazing bison]

but if you are studying code

gpt-5-chat is better

[tidal ginkgo]

yeah

haven´t used it much so

just testing it´s coding skills

wait

[blazing bison]

claude sonnet still better coder btw

[tidal ginkgo]

is there a gpt-5-thinking?

[tidal ginkgo]

is opus better than sonnet?

[blazing bison]

no

[tidal ginkgo]

oh

ok

[blazing bison]

wait

the no is for gpt-5 thinking

[tidal ginkgo]

oh

ok

[blazing bison]

about opus being better than sonnet, in benchs it is

test for yourself

[tidal ginkgo]

ok

lemme check

[woven turret]

is this gpt-5 still (high)

[tidal ginkgo]

idk

[solid brook]

Gpt 5 chat does not reason i think

[woven turret]

oh, my simple technical question show that:

1. gpt-5(lmarena) = gpt-5 thinking (chatgpt+) >= new gpt-5 (lmarena)
2. gpt-5-chat = gpt-5(plus)
3. gpt-5 in copilot has limit context

[tidal ginkgo]

@blazing bison told you

there is a thinking

[solid brook]

Guys the gpt 5 thinking is only in chatgpt website

In api there is reason effort

[tidal ginkgo]

oh ok

[woven turret]

i dont know, but gpt-5 on lmarena have great outputs, better than gpt-5 (plus user) and copilot, oc in my case

[tidal ginkgo]

wait

copilot has gpt-5?

[woven turret]

yeah sure

it randomly use thinking and quick answer

[solid brook]

Yeah the api is better

[tidal ginkgo]

oh cool

[woven turret]

lmarena team did well

[tidal ginkgo]

is there a limit?

[tidal ginkgo]

thanks staff

[woven turret]

idk, i just tesh 4-5 question

[tidal ginkgo]

btw someone do a bench gpt 5 and gpt 5 chat

[woven turret]

the limit is its output, maybe

[tidal ginkgo]

oh ok

[solid brook]

No i don't think so

[tidal ginkgo]

ok

[solid brook]

@echo aurora Context window limit is 400k for gpt 5 right?

[tidal ginkgo]

god yall pinging admins lol

[mental fern]

Is GPT-5 always “Generating…”?

[woven turret]

it is thinking, maybe, lol

[mental fern]

maybe my question is quite hard, thx

[blazing bison]

both answer at the same speed

[woven turret]

Yeah, my gpt-5 on plus account didnt think in my simple question test

[mental fern]

official description

[blazing bison]

atleast on lmarena

2 seconds for first token for both

no matter what the prompt

so are both thinking o both non thinking

[mental fern]

it looks like the gpt-5 on chatgpt.com automatically selects thinking or no-thinking

[ocean fulcrum]

Don't be confused, start super grok for sometime, then turned to gpt 5 you can clearly see what others can't

[solid brook]

Grok overpriced af

[ocean fulcrum]

Is it takes millions for subscription

[solid brook]

It is 300 doolers a month

[jade egret]

google stock doing pretty good today lol

[blazing bison]

tested on api playground, both reason

[solid brook]

Gemini gemini gemini

[tidal ginkgo]

takes it´s time

[mental fern]

Mmm i refreshed the page and it appeared.

[fallow trench]

how to use veo3?

[tidal ginkgo]

#video-arena-1 idk maybe here

but it´s veo 3 fast

[fallow trench]

ok

thanks

[tidal ginkgo]

hey, just now this is happening to me

maybe lots of people using it at a time?

[stray aspen]

copilot is also great

just use "think extremely hard"

at the start of you rprompt

yupp.ai is also a great website for trying gpt-5

but its limited

but overall microsoft copilot has hbecome a great option for gpt-5

in my opinion

[mental fern]

I don't think gpt-5 and gpt-5-chat are equally smart.

[jade egret]

why does gemini keep talking aboitu the current time

[jade egret]

wait what the real answer?

pi + e?

[mental fern]

TREE(2) is actually 3

[jade egret]

ya

i googled it 🙃

[stray aspen]

isnt gpt 5 chat a router

[mental fern]

I tried switching it to Thinking, but failed.

[inner gate]

Whays the difference between gpt5 and gpt5 chat?

Ahhhhh that makes more sense thanks lol

[stray aspen]

They arent

The open ai website itself says it

[solid brook]

It can think a lot

[tidal ginkgo]

oh ok

[haughty tangle]

Wow, so many models fail at this

[solid brook]

If you want to think less use "think low" in promt

[hoary elbow]

What is the difference between GPT five and GPT five chat?

[solid brook]

Gpt 5 is better

Use that

[tidal ginkgo]

oh, didn´t know that u could say in prompt how much to reason

[tidal ginkgo]

chat dumbed down

[hoary elbow]

All right

[tidal ginkgo]

gpt-5 better at any reasoning

[hoary elbow]

Ok

[woven turret]

oh, thank you, old but good tip, it works

[tidal ginkgo]

hey yall

how much does 5 chat hallucinate?

[hoary elbow]

Whenever I chat with him only in one conversation, he hallucinated and I’ve had over 20 conversations with it

How about you guys?

Oh, you’re talking about the chat version

I never really talk to it

[tidal ginkgo]

yeah

u shouldn´t lol

[hoary elbow]

[tidal ginkgo]

nobody uses it in here

[haughty tangle]

Does anyone have questions like this that models confidently fail at?

[hoary elbow]

Do you use it?

Mr. Discor

[tidal ginkgo]

nah

just test it

[solid brook]

It is bad

[hoary elbow]

Then how would you know if it’s worse?

[tidal ginkgo]

for something i want a quick answer 2

[solid brook]

The chat version

[tidal ginkgo]

because it doesn´t reason as much

dumber answers

[hoary elbow]

All right

But is it at least faster?

[tidal ginkgo]

yes

lots faster

use it depending on your situation

[hoary elbow]

OK, so I’m gonna use GPT-5 for coding and essays etc. GPT-5-chat for simple stuff

[tidal ginkgo]

exactly

[hoary elbow]

Because you said it was way faster

[tidal ginkgo]

yes

[hoary elbow]

Right now I’m getting it

[solid brook]

Actually

[hoary elbow]

So GPT five chat is like the direct answer version of GPT five

[solid brook]

I think gpt 5 mini is better than gpt 5 chat

[tidal ginkgo]

i hear you

[solid brook]

Gpt 5 mini reasons

[tidal ginkgo]

mmm

lemme test them

[hoary elbow]

And I never tested the chat version so

[tidal ginkgo]

dude

chat is way faster

[hoary elbow]

Yeah, but it’s worse

[tidal ginkgo]

yeah, i mean

[hoary elbow]

Someone said it was even worse than GPT five mini

[solid brook]

Yeah but gpt 5 mini is a good balance of power and speed

[tidal ginkgo]

good point

like 2 sec of difference

the problem with chat is

it´s dumbed down for people that only use chatgpt on their phones for a recipe on a tuesday

[hoary elbow]

People keep saying that GPT five is a one step closer to AGI but is it really I just wanna hear y’all’s opinions and I’m not talking about the chat version

[tidal ginkgo]

ehhh

getting there

[haughty tangle]

I mean ig

[solid brook]

It is. But the people that are trashing on it they were way too hyped and expected AGI

[tidal ginkgo]

guys

[haughty tangle]

Before GPT 5 I thought we would have AGI 2026, now I’m thinking 2027-2029

[tidal ginkgo]

point

[hoary elbow]

All right, can’t wait till 26-27-28 or 29 to see how AI gets

[tidal ginkgo]

lol

[haughty tangle]

true

[tidal ginkgo]

just wait until gpt-6

that hopefully is an AGI

[hoary elbow]

If Gemini three pro comes out, will it be better than GPT five

[tidal ginkgo]

yes

[hoary elbow]

Will it be one step closer to AGI?

[tidal ginkgo]

but not announcments

[tidal ginkgo]

slow steps

[haughty tangle]

before that we will have Gemini 3.0, maybe 3.5 and grok 5

[tidal ginkgo]

more like baby crawling

[tidal ginkgo]

grok 5

will be even closer than gemini 3

[haughty tangle]

Yes

[tidal ginkgo]

by a lot

[hoary elbow]

When baby grok releases, what will it teach kids?

[tidal ginkgo]

ehhhh

not my field lol

[hoary elbow]

All right

That’s OK

[solid brook]

I mean the gpt 5 we have now i a very good tech. If you showed it to someone 10 years ago they would have a heart attack lol

[tidal ginkgo]

no chatbots in that era lol

[hoary elbow]

What happens if you show GPT five to a Victorian child

[tidal ginkgo]

even if u show it to someone 2 years ago

[haughty tangle]

sigh we still wait for AGI so we can finally have an infinite self-improving feedback loop and have a dystopian future with unaligned ASI

[tidal ginkgo]

won´t comprehend

[hoary elbow]

All right

[tidal ginkgo]

lol

[hoary elbow]

People are saying GPT five sucks

[haughty tangle]

I am confident we will have a version of AM in the 2040s

[hoary elbow]

Is that true?

[haughty tangle]

its good

[tidal ginkgo]

no

[haughty tangle]

It’s like when grok 4 released

[hoary elbow]

All right

[tidal ginkgo]

very good

best ai yet

[hoary elbow]

Whenever I used Grok for for coding, it was damn near Gemini 2.0 pro level

[tidal ginkgo]

for now...

until gemini 3 or grok 5

[solid brook]

Majority of people are dumb and have way too high expectations

[hoary elbow]

So I don’t know all about all the hype in for Glock four when it came out

[haughty tangle]

yep

[tidal ginkgo]

yeah

they all want an AGI by 2026

which for now is NOT happening

[solid brook]

I feel 2027 or 2028

[haughty tangle]

me too

[tidal ginkgo]

closer to 2028 than 2027

sadly

[haughty tangle]

ASI though? Won’t take long once we have AGI

not gonna take long for obvious reasons

[tidal ginkgo]

lol

[tidal ginkgo]

mmmm

[solid brook]

Grok is overpriced af it needs to change.

[tidal ginkgo]

maybe 2030

[haughty tangle]

I’ve seen people say ASI will take decades after AGI

[tidal ginkgo]

yh

[solid brook]

I think most people buy supergrok for ani not the actuall ai

[hoary elbow]

Grok 4 for its price is like selling a hamster for the price of a car

[tidal ginkgo]

true

[stray aspen]

Hes actually smart AF tho

[tidal ginkgo]

elon chill

[solid brook]

Yes in logic

[stray aspen]

It's great for math

[solid brook]

But it is not worth the price at all

[stray aspen]

Yeah it's a robbery

[tidal ginkgo]

not in a million years

[stray aspen]

Just use lm arena lmao

[hoary elbow]

If Grok 4 was actually the best AI model at its time then I would understand all the hype

[solid brook]

I love lmarena.

[tidal ginkgo]

hey

are there any benches on every model of the gpt-5 family?

like 5, 5 chat, mini and nano

[tidal ginkgo]

which gpt5 is

[solid brook]

Forget 5 chat it is pointless

[hoary elbow]

Yeah, and it’s way less expensive

[solid brook]

5 mini i think is close to o3

[tidal ginkgo]

yeah

[hoary elbow]

But you can use it for free on lmarena just like Grok four and you can also use it for free in ChatGPT

[tidal ginkgo]

yeah but limits

on gpt

hey

[solid brook]

I didn't see any limits on the gpt 5 models

[hoary elbow]

Yeah, but there’s no limit in lmarena though

[tidal ginkgo]

how is gpt-5 doing on writing?

creative writing?

[hoary elbow]

GPT five is doing quite great on writing

[tidal ginkgo]

for now

[solid brook]

I think i heard a few say that it is worse than 4o in creative writing

[tidal ginkgo]

maybe because chat is more for public

[hoary elbow]

To anyone who has tested GPT five pro how did it feel?

[tidal ginkgo]

its better in that

[hoary elbow]

How was it?

[solid brook]

I have not used it but i think it is just a minimal upgrade

[hoary elbow]

All right

[tidal ginkgo]

yeah

is there more context window on pro?

[solid brook]

No

The chatgpt website context window is sht

[tidal ginkgo]

yeah

[solid brook]

Proof please

[hoary elbow]

I can’t wait for Gemini 3 pro and Grok five

I wonder how it is

[solid brook]

I heard gemini 3 is going to release december

[tidal ginkgo]

gemini 3 closer

grok 5 farther

[hoary elbow]

All right

[tidal ginkgo]

but much MUCH better

[hoary elbow]

I understand it might be much better because it’ll come after

[tidal ginkgo]

but gonna cost more than a house in 2025

[hoary elbow]

But what were y’all experience with Grok four because there was too much hype for it when it came out

When it coded, bad games was I just using bad prompts

Is it really the best model?

[solid brook]

Let's hope all of them create good models with good prices to keep the competetion alive

[tidal ginkgo]

is that proof?

[solid brook]

No the only thing it is good at is reason and logic

Uhm i don't see any gpt pro

[hoary elbow]

Their input prices and output prices are very cheap

[solid brook]

Actually on chatgpt website context window is limited 32k. The contex window in pic are for api

[tidal ginkgo]

?

[tidal ginkgo]

yeah

[solid brook]

And gpt 5 pro doesn't have any api

[tidal ginkgo]

sadly

[hoary elbow]

No one talks about copilot of around here so I just wanna clear this out. GPT five is now on Copilot.

[tidal ginkgo]

yeah but nobody cares about gemini today

[solid brook]

Gemini 2.5 pro becomes dumb after 200k context

[tidal ginkgo]

already sorted it out like 2 hours ago lol

[solid brook]

The 1 mil is useless

[hoary elbow]

Ok

[tidal ginkgo]

nobody uses copilot anyways lol

[solid brook]

Gpt 5 is nerfed on copilot

Lmarena gpt 5 much better

[tidal ginkgo]

idk

[hoary elbow]

I feel like GPT five on Copilot might be a faster version

[tidal ginkgo]

haven´t tried copilot

[hoary elbow]

Let me test it out with GPT chat

[tidal ginkgo]

yeah maybe they using gpt 5 chat

[solid brook]

I feel it is gpt 5 mini not gpt 5

[hoary elbow]

They’re using GPT five chat

[tidal ginkgo]

EWWWWW I JUST GOT 5 CHAT IN BATTLE

[hoary elbow]

I told them the exact same thing and they gave the exact same answers

[solid brook]

Lol

[tidal ginkgo]

no temp lol

[obsidian shell]

wait which gpt5 they have in copilot?

[tidal ginkgo]

chat says @hoary elbow

[solid brook]

Nerfed

[obsidian shell]

chat?

[solid brook]

It does not reason much

[obsidian shell]

i thought it was pro normal nano and mini

[tidal ginkgo]

here

[tidal ginkgo]

pro normal chat mini nano

[solid brook]

On lmarena you can get the model to think up to 3 or 4 minutes. But not on copilot

[obsidian shell]

am talking about the pro plan of copilot

[tidal ginkgo]

yeah

[tidal ginkgo]

nobody uses copilot lol

[obsidian shell]

i do

[hoary elbow]

I wanna clear something up

[tidal ginkgo]

well lmarena better

[hoary elbow]

It says it’s GPT four

So Copilot just straight up, lied to our faces

[tidal ginkgo]

yeah

everyone does that

even chatgpt website

oh

[solid brook]

They have to give it system promt to say it is gpt t

[tidal ginkgo]

nvm

[solid brook]

Gpt 5

[hoary elbow]

Maybe they didn’t change the system prompt yet

[tidal ginkgo]

maybe

[hoary elbow]

But when they change the system prompt, I might know, or it might straight up lie to my face and say it’s the original version of ChatGPT five

[tidal ginkgo]

he doesn´t lie lol

[clear spear]

so umm... why is there a gpt-5-chat? what's that about?

[hoary elbow]

It’s a small version of ChatGPT five

[solid brook]

Hmmm guys when i go use gpt 3.5 on legacy lmarena i get nostalgia

[hoary elbow]

And it’s bad

[tidal ginkgo]

gpt-5-chat is used in the chatgpt website

[leaden palm]

Same deal as chatgpt-4o-latest I believe

[hoary elbow]

Well, not that bad but like it’s worse than GPT five

[tidal ginkgo]

gpt-5 is the better, more reasoning model

[tidal ginkgo]

lol

[clear spear]

okay

[tidal ginkgo]

lol the pfp

oH ITS A DOG

[solid brook]

The said they want to simplify the models. But i feel it got more complicated

[tidal ginkgo]

yeah

wwasn´t on the livestream at all had to go out

just saw the first 5 min

[hoary elbow]

Is there anything better than GPT image right now?

[tidal ginkgo]

imagen 4 ultra

much better in text

[sturdy mica]

gpt-5 high is op

oh my god

[tidal ginkgo]

? what is it?

[sturdy mica]

i just said

gpt-5 high

[tidal ginkgo]

yeah ig

[sturdy mica]

yeah you guess?

[tidal ginkgo]

closest we got to AGI

right now

[tidal ginkgo]

yes

[hoary elbow]

I feel like whenever AI gets like a very big update and they suddenly become like the biggest model I feel like when it updates it might not be like as better so I feel like in like 2029 the first AGI will come

[tidal ginkgo]

u just said 2027-28 5 min ago lol

[unborn lantern]

In pro version?

[hoary elbow]

Well, I didn’t really think about it before

I said or 2029

[tidal ginkgo]

i keep my ground

closer to 28 than 27

but just hypothesis lol

[solid brook]

Guys gpt 4 was 2 years ago. The diffrence between gpt 5 and gpt 4 is a whole lot. 2027 will be AGI

Or 2028

[tidal ginkgo]

yeah

oh wow

2 years ago

time flies

[tidal ginkgo]

closer to that

[hoary elbow]

What video generators do they have in video Arena?

[tidal ginkgo]

3 fast

and kling master

and wan

i think so

[hoary elbow]

All right

It also has sea dance too. I saw one of the pros.

[solid brook]

Google has crazy advancments in I

AI

Damn no image

Compare this to all other AI companies

[solar hollow]

yes google is king

[patent aspen]

They didn't even mention publishing 80% of the seminal AI papers of the last decade

ChatGPT couldn't exist without Google

[rocky mauve]

ai is surprisingly growing and evolving so fast, last year ai was garbage and couldn’t do anything good

[solid brook]

I feel google has AGI internal

I mean classified

The might have it

Dude i am saying classified

How would you know

Hmmm

Overall every tech before releasing gets used by the goverment first

[astral prawn]

Rip btw I feel for u

Anyone else here using local system prompts with gpt5? How do you feel it’s performing?

Kind of sucks we can’t tweak temperature anymore

[prime mulch]

@echo aurora all image generators expect flux kontext and gpt are dead

[autumn cloud]

yo

what's best AI for coding atm? gpt 5?

[prime mulch]

Gemini 2.5 pro and gpt 5

[solid brook]

Gpt 5

Use the api

Or lmarena

[autumn cloud]

thanks

[lament bone]

how do i delete all data in lmarena?

[gentle elk]

[quiet pollen]

What happened to Claude family?

[solar hollow]

no

[random canyon]

What is the difference between the regular gpt-5 and gpt-5-chat?

[solar hollow]

sota is nowhere near agi yet

and if you want to believe unlikely, unverifiable things, go ahead believe in "agi internally" or some god

[solid brook]

See i think the goverment has a model far more better than the public sota

[solar hollow]

what good does this for you without any evidence in that direction?

did you ever find any indication of it at all?

[keen beacon]

Maybe the gpt-5 have high reasoning.

[solid brook]

I mean it is pretty clear

If you look in history

[lone relic]

claude opus 4.1 is acc pretty good

[solid brook]

Every big tech before being public was used secretly

[warm fulcrum]

pricing tho 🫦

[lone relic]

so is opus 4 itself

for coding

[lone relic]

assume that is not considered 😢

[keen beacon]

Where is this

[warm fulcrum]

assume we billionaires 🫦

[warm fulcrum]

gatekeeping

[lone relic]

lmarena has claude opus 4,1

[solid brook]

Limited

[keen beacon]

Wdym

[lone relic]

isnt that openrouter

[warm fulcrum]

noooooo

[keen beacon]

Thanks

[solid brook]

You have to pay btw

[lone relic]

yeah

[solid brook]

The api price

[lone relic]

nothing is free in this world

[solid brook]

Oxygen?

[keen beacon]

?

[lone relic]

we have to pay with co2

[solid brook]

You have to pay for api price in openrouter not free

[lone relic]

there acc are free apis tho there

[keen beacon]

I never used OpenRouter, but since LMA has Opus 4.1 now why would I want to use it

[acoustic cliff]

we have to pay with our data

[solid brook]

I know but top models you have to pay

[lone relic]

yup

[lone relic]

ye thats true fs

[lone relic]

amazing at writing, great for coding

esp if agentic

for that warp gives limited free opus 4.1

wait are we allowed to share referral links here??

[solid brook]

Gpt 5 is way more efficent than opus 4.1

[lone relic]

tru

[keen beacon]

How many messages

Did you try it

[lone relic]

250 creds

per month

50ish at least

i managed to hit 70-80 though

[solid brook]

Even if opus 4.1 is better it is very minimal not worth the price. Claude must change or it is cooked

[lone relic]

valid lowk

with gpt 5 out

[keen beacon]

Artificial analysis shows GPT-5 is bad at coding.

[solid brook]

Wait for gemini 3 that will blow everything

[lone relic]

it got dumbed down today cuz of some bug 🙁

[lone relic]

i have gemini 2.5 pro lol

[solid brook]

It i s good

[lone relic]

ye its hella good

[solid brook]

Just say think hard in promt

[keen beacon]

lol

[solid brook]

And btw 2.5 pro got nerfed. Original 2.5 pro was at 4 sonnet or opus level

[lone relic]

are we allowed to share referral links here?:

[solid brook]

Idk

Ask the mods

[lone relic]

oh yeah ik that 🙁

<@&1349916362595635286> am i allowed to share referral links here?

[zealous panther]

check rules

To ensure we have an inclusive and welcoming community we have some rules everyone should review and adhere to. The moderation team has final say over if violations of these rules have or have not occurred, along with the actions we take in response.

✅ Act in accordance with Discord’s Terms of Service and Community Guidelines.Violations of these terms and guidelines should be reported directly to Discord. It’s also recommended to be familiar with Discord’s Safety Center for more information on how to remain safe while using Discord.

✅ No NSFW, Harmful Content, or Spam. This includes, but isn’t limited to, hate speech, harassment, racisms, sexism, homophobia, illegal content, inappropriate profile pictures, sharing of inappropriate content, and so on.

✅ Treat others with Respect. Be kind, assume good intent from others, and keep disagreements respectful. It’s encouraged to share your disagreements, but only if it’s done in a respectful and productive way.

✅ Do not promote or advertise. This includes sharing of: social media, other Discord servers, or involved projects in an promoting manner.

✅ Avoid political and religious content. As a space that’s inclusive to many different worldviews we ask to avoid topics related to politics and religion in order to maintain an inclusive space. It is okay to have discussion related to new policy or laws as long as it’s related to AI.

✅ Do not impersonate staff, moderators, or others. Efforts to impersonate LMArena staff, server moderators, or other community members is not allowed, even in a joking manner.

✅ Message in English only. Please keep discussions in English.

Most importantly, remember why we are here, to advance the understanding and application of AI!

[keen beacon]

GPT-5

Wait this the photo from yesterday.

This is from now.

98% at AIME but somehow can’t solve a equation

[solid brook]

Guys

It is not funny anymore

[keen beacon]

Imagine the thousands of simple equations GPT can’t solve.

[solid brook]

Every single ai is garbage at math without reason

[lone relic]

gpt 5 has been dumbed out

[keen beacon]

So benchmark is fake?

[solid brook]

I mean without thinking

You must enable thinking

[lone relic]

today gpt 5 has been dumbed out, sam altman posted in twitter abt it

[lone relic]

ye

[keen beacon]

Let me see

[solid brook]

Why? It is good for me

Lemme see x

[lone relic]

without thinking it was bad

with thinking tho unbeatable tbh

[celest zenith]

I fine-tuned OpenAI’s OSS 20B reasoning model using the most popular medical reasoning dataset and published the results on Hugging Face. The model can break down complex medical cases step-by-step, identify possible diagnoses in clinical scenarios, and answer board-exam-style questions with logical reasoning.

During training, I used 4-bit optimization and enhanced the model’s performance in medical contexts while preserving its Chain-of-Thought reasoning capabilities. The training format includes “question,” “Complex_CoT,” and “Response” fields—allowing the model to first reason in detail, then provide the final answer.

You can check it out here:
🔗 https://huggingface.co/dousery/medical-reasoning-gpt-oss-20b

I’d love to hear feedback from anyone working on or interested in medical AI.

[keen beacon]

They said they will double the rate limits for Plus

[lone relic]

mhm

[keen beacon]

Don't know about that. Deepseek V3 calculated it just fine without reasoning.

And the model is from march

[solid brook]

Okay so what is the point of this argument?

[keen beacon]

Without reasoning, models can do math just fine too