general

[torn mantle]

looks the same to me

[balmy mist]

yooo claude just got 7/10 on simple bench public

bruhhh

claude cooked

[torn mantle]

doesnt look that different

[balmy mist]

gonna do pokemon test

[cedar tide]

The most important thing for me is that they must lower the price of the API.

[torn mantle]

yea

wait

is that you leo

[balmy mist]

lol

[torn mantle]

we have like 4 leo

[balmy mist]

yeah thats him

[torn mantle]

bruh

[balmy mist]

i think i might pay for claude now

idk

[torn mantle]

lol no

wait

[balmy mist]

its so hard to choose damn

[cedar tide]

we also hope for 1m of context and complete multimodality

[civic flame]

lol yeah the rest were nuked

[torn mantle]

whatever im testing on their website is < sonnet 3.7 max thinking

[civic flame]

i don't have to stress, anthropic gave me a generous sum of free credits :3

[cedar tide]

What ?

[civic flame]

well yes that's because it's not thinking

[torn mantle]

i see

so thats only the instruct model

[frosty lark]

so I am no @earnest parcel but I started to collect (late) some relatively tricky questions for LLMs.

For what I can see there is no big change compared to Claude 3.7 (if I am getting Claude 4 ofc)

[balmy mist]

i thought the new model would choose when to think and when not to think?

[tall summit]

#share-prompts

[frosty lark]

Especially something like this let me scratch my head.

I mean sure it is correct (what is important) but not that maintainable

[narrow elbow]

Adults don't choose, take it all.🤑

[torn mantle]

make sense why its free

they are talking about creation of bioweapons

but isnt it possible with the current LLMs

i just dont understand anthropic really

is the model really super smart or what

[frosty lark]

done

[torn mantle]

pretty sure if you get o3 jailbreaked or gemini 2.5 pro you can do such things as well

Dario was always obsessed with the extra security and safety of their models

[frosty lark]

I think it is good. I mean, as wrote: those models "compress" a lot of human knowledge and if they connect the dots appropriately, they can deliver snippets that can be useful for others to progress their work.

Humans aren't stupid (yet) and they can piece the snippets together

[narrow elbow]

it is difficult to prevent human questioning techniques, or to distinguish real scientific research

[torn mantle]

the best instruct model we have so far is grok 3

[drifting thorn]

It’s kinda like lobotomy to their models

[ember rapids]

Hopefully refusals aren’t too bad for opus

[drifting thorn]

There’s always malicious requests

[cedar tide]

Or GPT 4.5

[drifting thorn]

Have you guys heard of the Continuous Thought Machine by SakanaAI?

I think replacing the input processing and the FFN in transformers to Continuous Thought Machine would lead to AI that can think internally

[cedar tide]

Give me some prompts to test for Claude 4 please

[elder rapids]

damn fr?

is it another thinking model?

[frosty lark]

so, it is impressive it codes (python/JS) but the answer is - considering the hype - not where it should be. Be it Claude 3.7 or Claude 4.

[torn mantle]

the free model im using shows 'claude sonnet 3.7' and leo said its referring to their latest model claude 4 sonnet

[cedar tide]

The hype is for opus

[frosty lark]

also it writes a lot of

"You're absolutely right! "
"You're absolutely right to question that!"
"Brilliant point! You're absolutely right"

and so on.

[cedar tide]

Its true

[frosty lark]

interestingly the claude models avoid that in lmarena

[elder rapids]

oh fr?

I thought Claude 4 wasn't going to be

[elder rapids]

in the UI

[frosty lark]

I mean, still impressive to do that work with few prompts in few minutes. But it doesn't match the hype

Taking the results as is would be full of mistakes

[torn mantle]

whatever is being tested on claude.ai is not that impressive for difficult-hard questions

[frosty lark]

do you mean that the questions aren't hard or the answers are underwhelming?

[elder rapids]

tbh I don't expect the new models to be any good

or at least better

[alpine coral]

sonnet-3.7 on claude chat has more recent knowledge than the 0219 one

[elder rapids]

anthropic doesn't have the researchers nor the compute to brute force like openAI and innovate like deepmind

[alpine coral]

and maybe better tool usage. but otherwise it doesn't seem a step change from my usage so far

[frosty lark]

I have the feeling we are approaching "slow" improvements until a new architecture pops out (like transformers and RL did so far). Actually I am still waiting for a LLM orchestrator that picks specialized models according to the prompt and answer.

one doesn't need AGI. Near AGI is enough if the orchestration of narrow AIs is good enough.

[alpine coral]

i haven't looked into it / no idea if it's actually consequential.. but the text 'diffusion' model google announced at io struck me as kinda new

adapting what works for image gen to text

some kinda parrell stuff going on

i think

beyond that.. i don't really know what it's about ha but yeah it sounded at least like kinda new

[elder rapids]

this is why everyone thinks DeepMind is winning

and will win

[torn mantle]

yes

thats the only difference im noticing

the model isnt smart or anything

[unborn ocean]

they are already taking down the 3.7 model on many providers

can't be long

before they switch

[misty vault]

mercury coder did it before 3 months ago

[frosty lark]

I don't get it, Claude 3.5 was still there the entire time, why taking away 3.7

I mean at least for a certain period where people get their workflow to fit the new model

[alpine coral]

ah i see.. less innovative than i thought than

[torn mantle]

https://x.com/alexalbert__/status/1925575375029248427

Alex Albert (@alexalbert__)

T-1 hour until livestream starts!

X · May 22, 2025, 3:31 PM

[alpine coral]

do you know if it [murcury coder] is any good?

[torn mantle]

45mins left

[misty vault]

the new gemini flash is also actually way better for instructions like really good

But idk compared to gork or 4.5

Probably 4.5 king because it is obese

[misty vault]

social_credit

[unborn ocean]

more efficient?, idk

[drifting thorn]

I wonder the time when OpenAI uses 4.5 to train a reasoning model

Using the way like AlphaEvolve and Absolute Zero Reasoners do

[sage raptor]

https://www.youtube.com/watch?v=A6MQ3AuE8Bk

[drifting thorn]

It’s not that “special”

[torn mantle]

Dario is like Ilya

afraid of everything

[grim axle]

What about for coding?

[torn mantle]

the report looks much better than the previous one

[elder rapids]

yep

2.5 flash was already good asf

but now I think it's the best available rn imo

by a large margin

[grim axle]

no pro is

[elder rapids]

the point is instruction following

[grim axle]

What about this?

http://aistudio.google.com/app/prompts/new_chat?model=gemini-2.5-pro-preview-05-06

[quiet folio]

no, claude religion would be valid

claude is agi

[grim axle]

Gemini 2.5 pro is also good for instructions it’s just slower when responding 🤷‍♂️

Claude has a memory of a fish

[misty vault]

im sucking off anthropic now its over for gpt 4

social_credit

[drifting thorn]

Actually GPT 4 was a big boy

1.7 trillion parameter

So why are the companies now stick to models with hundreds of billions parameter level?

[misty vault]

It was even bigger than @alpine coral

[misty vault]

money

The mainstream media and everyday people dont care about gpt 4 they all love modern chatgpt

gpt 4o marketing brings them more money

[drifting thorn]

?

[misty vault]

compared to if they served expensive gpt 4

[drifting thorn]

GPT 4 was the topic

[misty vault]

no

[drifting thorn]

As it was the only model available back in the days

[misty vault]

it was hot topic for coders and ai enthusiaists

ai wasnt that big yet in gpt 4 age compared to how hyped it is now

[drifting thorn]

At that moment we didn’t have Gemini and Claude

[misty vault]

It reached way more everyday people now

like @alpine coral

It even reached drooling aliens

[cedar tide]

https://x.com/test_tm7873/status/1925583609563328609?t=3lyGxFlRpQcEe_KGWb-ssQ&s=19

testtm (@test_tm7873)

its now official Claude Opus 4 confirmed!

X · May 22, 2025, 4:04 PM

[misty vault]

I really hope claude 4 is also obese model like 3 opus or something

I'm afraid not though

anthropic couldnt afford that could they

[elder rapids]

yep but that's not the point

[drifting thorn]

Opus should be obese

[elder rapids]

are you paying for 2.5 flash or 2.5 pro

doing instruction following tasks

[drifting thorn]

And I’m now putting high hopes on CTM

[misty vault]

gemini 2.5pro is more willing to follow any instructions but its not good at actually succeeding to follow the instructions properly compared to gpt 4.5 (or 4) or 2.5 flash

[drifting thorn]

It’s basically toooo fascinating

[cedar tide]

https://fixupx.com/test_tm7873/status/1925585410677190709?t=XQb6D56Q6v1jSrYSuAtSng&s=19

testtm (@test_tm7873)

i saved one of the Anthropic videos about Claude 4.
︀︀"Claude 4 | Research for comprehensive analysis"
︀︀the rest is arleady set to private. but atleast one i saved. :D

**💬 1 ❤️ 9 👁️ 174 **

▶ Play video

FixupX · May 22, 2025, 4:11 PM

[misty vault]

LMAO

wdym

[misty vault]

Okay so it's confirmed

we have claude 4

[sage raptor]

https://support.anthropic.com/en/articles/8114494-how-up-to-date-is-claude-s-training-data

100% confirmed

[torn mantle]

https://support.anthropic.com/en/articles/11408405-claude-4-invite-contest

[misty vault]

omaygot

[torn mantle]

didnt you guys say october 2024

[misty vault]

im getting so hard rn from claude 4

[torn mantle]

you guys have it all wrong

[elder rapids]

system prompt could be suggesting October 2024

yet it's explaining things past that mark

just guessing that's what was meant

yep

[misty vault]

im actually going to buy this

selling bing chat access gpt-4-32k, gpt-4-preview, gpt-4-0314, gpt-4-turbo for 100$

[torn mantle]

3 min left?

https://www.youtube.com/watch?v=EvtPBaaykdo

[elder rapids]

this dumbass music bro 😭😭

[torn mantle]

💃

🪩

[misty vault]

im getting lotion ready

[elder rapids]

deadass

alright it's time

they're late

asf

Anthropic L

it's been 30 seconds already

[misty vault]

it started

[balmy mist]

yall ready!!

im watching it here: https://www.youtube.com/watch?v=dt2oDtssDNg

[drifting thorn]

Anthropic has not been a general leader

[elder rapids]

get to the point anthropic

smh

[unborn ocean]

music on google IO was way better

disappointed

[elder rapids]

unironically ye

ts had anime music

😭

[misty vault]

me buying the 100$ per month cwuade subxcription before they increase it to 300$ per month

[elder rapids]

don't disappoint me dario

[drifting thorn]

I’m gonna sleep

[elder rapids]

you'll regret it if you do

[drifting thorn]

From UTC +8

[elder rapids]

Claude 4 opus and sonnet

[unborn ocean]

yes

[elder rapids]

nice

[unborn ocean]

let's go

[elder rapids]

released

alright now be quiet dawg

you got to the point

😭 🙏

[tall summit]

he said it

[balmy mist]

yes!!!!

[elder rapids]

keynote over

pack it up guys

[tall summit]

yep bye

[unborn ocean]

man dario be looking like a mad scientist though

[elder rapids]

ong

[echo aurora]

going to be streaming this in #1340554757827461215 in a min

[torn mantle]

So opus 4 only good at coding?

[elder rapids]

wondering how good it's going to be

[torn mantle]

Advanced reasoning at coding

They said

[drifting thorn]

When they don’t show the numbers it means the model may actually suck

[balmy mist]

imma buy claude asap lol

[unborn ocean]

same cost

[cedar tide]

https://x.com/AnthropicAI/status/1925591505332576377?t=GUeiKurZdHB_hlxecNEemw&s=19

Anthropic (@AnthropicAI)

Introducing the next generation: Claude Opus 4 and Claude Sonnet 4.

Claude Opus 4 is our most powerful model yet, and the world’s best coding model.

Claude Sonnet 4 is a significant upgrade from its predecessor, delivering superior coding and reasoning.

X · May 22, 2025, 4:36 PM

[elder rapids]

"worlds best AI coding assistant"

prove it lil bro

[elder rapids]

tbh

it's not

but Claude in practice is always better

[cedar tide]

Opus 4 at $15/$75 per million tokens (input/output) and Sonnet 4 at $3/$15.

[elder rapids]

it's not the highest in any of the benchmarks pass@1

[drifting thorn]

Bruh sonnet performs better in most tasks than opus

[elder rapids]

besides SWE

[balmy mist]

yeah vibes matter

[elder rapids]

ye

[unborn ocean]

regressed performance on GPQA (without parallel compute) kinda points me towards smaller experts / worse reasoning (very unrealistic) (guesstimate though)

[balmy mist]

benchmarks kinda pointless now

[unborn ocean]

which might be why they are serving it so quickly

[elder rapids]

"Kitler"

[balmy mist]

someone try opus and let us know

[elder rapids]

ye

[small haven]

ok so that’s why opus wasnt released lol

[keen ferry]

opus 4 costs less than o3 thats awesome

[cedar tide]

What is it paralel compute ?

[small haven]

codex is at 75% swe benched not sota

[drifting thorn]

taking 64 samples and choose the best one

Idk if it’s actually 64

[keen ferry]

is there a thinking mode for opus 4 or sonnet 4

[drifting thorn]

Is Opus 4 a thinking model?

[civic flame]

yes

[drifting thorn]

It’s like

A little bit under my expectations

[misty vault]

NOOOOOOOOOO

I didn't expect that from anthropic though

they dont have the computing power do they

[drifting thorn]

I was expecting a 2.5 Pro type of performance gain

[drifting thorn]

I expect that from Google

[sweet tinsel]

Is the Claude Deep Research now with Opus or Sonnet, or can you choose?

[misty vault]

yes but google doesnt do that

unfortunately

they are going for the 4o style

small, restarted models trained on only specific stem topics

giving the illussion its smart

[misty vault]

But claude prob going same direction because they cant train such big model right
idk

bro they just drew a random line on that hours of work graph

[keen fulcrum]

So what to choose?

o3 or Claude 4 Opus?

[balmy mist]

this is crazy

[balmy mist]

claude 4 easily

[keen fulcrum]

Now I need gemini 3..
or 2.5 ultra

[misty vault]

I hope they fix the issues 2.5 pro has

then ill accept google

[sage raptor]

for coding claude

[elder rapids]

not much high hopes for it being any good for anything outside of coding

[elder rapids]

but this is great since it aligns

[calm sequoia]

They already compared it to codex, while others compare themselves to 6 month old models. SOTA

[cedar tide]

What the windows context ?

200k

I found

[small haven]

Actually ?

[small haven]

Still good

Is it ready in claude code

[unborn ocean]

think so

they said that it is on all services now

(besides serving the model on 3p, which will take some time)

[small haven]

Oh shxt so it is better than codex cool

[unborn ocean]

prob

codex did not really have a large edge

[torn mantle]

Opus looks good ngl

[elder rapids]

ye

[small haven]

they need a ui like codex and its gg

[unborn ocean]

third party

like google

or aws

[small haven]

i think codex is capped in tokens prolly 32

[narrow elbow]

If it's less hallucinations than 3.7, I'll be satisfied

[small haven]

now it is

[cedar tide]

Swe and terminal bench its without thinking

[unborn ocean]

prob because the reasoning still sucks

[balmy mist]

how is opus?

[small haven]

anyone tried it yet?

[misty vault]

how

[elder rapids]

damn

sonnet 4 is stupid

[misty vault]

do u have claude max

[unborn ocean]

claude 4 on code confirmed ✅

[elder rapids]

it's over

[misty vault]

my dog

[elder rapids]

talking to ts is BAD

[unborn ocean]

and also as vsc extension

[elder rapids]

3.5 sonnet still the best vibes model

did you try it?

[torn mantle]

https://x.com/ashtom/status/1925597395192357337

Thomas Dohmke (@ashtom)

Claude Sonnet 4 is here. Now avail in all Copilot plans. And it’s the new base model for the GitHub Copilot coding agent 😎

https://t.co/PnJyeT1IOu

X · May 22, 2025, 4:59 PM

[elder rapids]

is 4 opus available in the api

[north vale]

how is 4 opus' vibes

[unborn ocean]

let's go

the 10$ per month sure are worth it

[small haven]

Is opus in Claude code

[misty vault]

Broo nobody joined @echo aurora in the voice channel to watch the livestream with him

So mean guys

They left now due to loneliness

[small haven]

Lets go

[misty vault]

poor them

[echo aurora]

😭

I'll be back, had to run to a meeting

[leaden meteor]

When can we compare these new Claude models with o3 and 2.5 on arena?

[torn mantle]

Clap clap

[small haven]

not clicking phishing links, gimme screenshots

[torn mantle]

agree

[misty vault]

would it be cheaper to pay for claude 4 opus api rather than 100$ per month claude code and just use it with other vs code ai plugins

Im already happy with copy paste

🗿

[torn mantle]

Meh

Not agi

Yes you can

Abd you've done it many times

[misty vault]

craig is paid actor by anthropic

[torn mantle]

Lying all the time

[narrow elbow]

TV remote control?

🤣

[balmy mist]

is the claudes on webdev?

[torn mantle]

Nightwhisper made a better ui

[balmy mist]

i dont even think nw is real anymore, did that even happen?

[elder rapids]

tbh there's nothing that compares to NW

[balmy mist]

it might have been a dream

[elder rapids]

it was so beautiful

[balmy mist]

what about opus?

[elder rapids]

I'd be surprised

nw was good

[willow grail]

30yo, single, otter body, 183cm, 80kg.
looking for sugar daddy cause opus 4.

[elder rapids]

elite vibes, was smart asf and successfully at code SOTA

[unborn ocean]

"reliable access to claude" as if 🤣

they have like close to the worst api

[cedar tide]

Badest ai i see ever

[willow grail]

i dont care about your net worth, daddy

[small haven]

so first vibes meh?

gonna try deep research with claude opus

[small haven]

can someone send me a prompt for deep research

[torn mantle]

Ok

[small haven]

you

[cedar tide]

What the probleme now ?

[sweet tinsel]

Perplexity is good when it's free. I got Perplexity Pro for free by Telekom (T-Mobile Germany) and they have Claude 4.0 Sonnet Thinking already, just hoping for Opus now like with Opus 3, which was available in Perplexity.

[willow grail]

why do i need a net worth when im looking for sugar daddy?

[sweet tinsel]

Like nearly unlimited use of o4-mini, R1, Gemini 2.5 Pro and Grok 3. It even behaves like normal Chatbots when the Websearch has been turned off.

[misty vault]

Based on a critical synthesis of recent, high-quality human clinical trials and systematic reviews, determine which compound – Berberine, Propolis, or Resveratrol – demonstrates the most compelling evidence for promoting overall health.

[sweet tinsel]

Please write a comprehensive and in depth research report on the mass expulsion of ethnic Germans after World War II. Analyze the historical context driving these expulsions, the political decisions and international agreements that shaped the process, the social and economic consequences for displaced populations, the humanitarian and legal dimensions, personal testimonies, and the long term demographic and geopolitical impacts, drawing on primary sources, statistical evidence, and varied historiographical perspectives.

This would be very interesting for me.

[misty vault]

true

[elder rapids]

damn

Claude 4 opus isn't very good at regular tasks

fukkk

[torn mantle]

Me

[willow grail]

Asura is a bad free-for-all server on the isle

[torn mantle]

Huh

[elder rapids]

damn Im actually sad

it's good asf at coding tho

[sweet tinsel]

This is my go to prompt to Test Deep Researches of ChatGPT, Agents and such. Already have tried Flowwith AI, Grok DR, Grok DDR, Manus, ChatGPT DR (o3), Gemini DR (2.0 Flash), Perplexity DR, Genspark Super Agent and Ithy. Claude DR and Gemini DR are still missing as i havent gotten a Sub there. When you want i can send my DR Test with this prompt in.

[misty vault]

😔

[torn mantle]

https://x.com/janleike/status/1925595727906185355

Jan Leike (@janleike)

So many things to love about Claude 4! My favorite is that the model is so strong that we had to turn on additional safety mitigations according to Anthropic's responsible scaling policy

X · May 22, 2025, 4:52 PM

They made it dumber

[willow grail]

https://i.imgur.com/qikrWUu.png

[elder rapids]

much better vibes tho, but it's still not 3.5 sonnet lvl

[torn mantle]

Wth is that

[willow grail]

server for isle

[torn mantle]

Oh

[willow grail]

ur allowed to kill on sight any dinosaurs who is nesting... sucks

[torn mantle]

Bet its a good server

Can tell just from the name

[elder rapids]

I'll do Gemini DR for you

if you'd like

[willow grail]

no its sucks. u cant even nest without getting kos'd

[sweet tinsel]

Thanks! I would apreciate that.

[torn mantle]

Can you re-run again that prompt @small haven

[small haven]

why

[elder rapids]

alr it's started

[torn mantle]

To compare the old research with this one

[elder rapids]

I'll let you know when it's done

[small haven]

oh yea

[elder rapids]

goddamn

they're changing the setting

one minute they're standing up like normal

next it's an interview

[misty vault]

bro so much yap too

this guy is struggling to talk😔

[unborn ocean]

tru

he is too shy for the stage or too nerdy

[small haven]

this is gonna be fun

[misty vault]

use a tissue, dont cover the lmarena discord chat

[sage raptor]

idk what is this

[misty vault]

real

[sage raptor]

open ai 63

[keen ferry]

whoa gonna be #1 in the web arena gemini or opus?

[misty vault]

gork 3.5 misinformation trend 2.0 starting today

[small haven]

https://claude.ai/public/artifacts/f0249303-c59c-43d8-83c3-7c29799f7573

5mins

[unborn ocean]

ai gen (nvm did not read context)

[small haven]

2.0 flash image preview 😭

[misty vault]

Did the guy who I was going to ping leave the server because I called him a drooling alien

[small haven]

claude code is nice, no enviroment setup needed like codex

[misty vault]

idk weird name
nvm he didn't

[torn mantle]

what model did you use?

[tall summit]

did you put it through chatgpt image gen

[small haven]

but i miss wanting to spam tasks on the go...

[small haven]

opus 4

[torn mantle]

def better than the previous one

much much better

[misty vault]

compare claude 4 opus to oai deep research

[torn mantle]

but

why does the report look short

[misty vault]

fr

[torn mantle]

compared to the demo from the video?

[unborn ocean]

u know dario talks 100% like the nerdy profs at my uni that have trouble upholding a normal social life

and like to yap a lot

[misty vault]

That's because he is a drooling alien

[tall summit]

which is sad

[narrow elbow]

everybody is like drooling alien to you

[small haven]

troll or actually lol

[leaden meteor]

I can't find it yet on side by side comparison? It's only on blind tests now...?

[torn mantle]

just a bit better

but nothing great

[harsh flume]

Have any of the anon models this month hinted at being Claude or nah?

[torn mantle]

it extracted certain parameters this time

[harsh flume]

I haven't played with the arena the past two weeks

[small haven]

thats what i thought as well

its meh ish, chatgpt dr better

[misty vault]

only @alpine coral and dario is

[small haven]

but i did say "assume" for the feedback @torn mantle

[misty vault]

@narrow elbow Did u just drool on my comment

[small haven]

so in perpetuity, ur saying codex > claude code

[elder rapids]

ye I've been using it, testing it for a bit, and as far as I can tell, its particularly worse at language

it obfuscates technical concepts and I have to introduce language to help it communicate

[tall summit]

umm what

[elder rapids]

ye it didn't know what a category error was or how to get it across

[small haven]

i mean right now opus on claude code is pretty damn slow

[elder rapids]

but it knew there was a category mistake

[small haven]

codex is speedier ngl

[unborn ocean]

Nah the people who joined in person get 3 months free claude max
we have to pay 300$💀

[balmy mist]

i was there

[tall summit]

oh huh how odd

[small haven]

opus 4 is slow but beefy and smarter than codex

[elder rapids]

you can tell its smart, but it doesn't have foresight imo

[inner hare]

please add claude 4

[torn mantle]

so oai still top 1?

[elder rapids]

wonder how this affects coding

[misty vault]

you're a drooling alien

[torn mantle]

@small haven whats your take?

[inner hare]

Why?

[balmy mist]

@keen beacon whats your take

[torn mantle]

@small haven

1. o3
2. opus
   3.gemini 2.5
   ?

[small haven]

lmao

[small haven]

for dr? looks like it

[inner hare]

I love you

[torn mantle]

for general use

overall

[small haven]

yes o3 is still ahead, but code wise (speed aside) opus 4 is ahead

[torn mantle]

https://x.com/testingcatalog/status/1925609116430385578

TestingCatalog News 🗞 (@testingcatalog)

No Claude 4 on Windsurf today 👀

X · May 22, 2025, 5:46 PM

lmao

google should just hit them with nightwhisper

[elder rapids]

deadass

[torn mantle]

and lets see how they act

[elder rapids]

@balmy mist flowith has access to Claude 4's

speaks nothing like opus

😭

[small haven]

opus 4 spent 5 mins reading files 😭

[torn mantle]

fastest ever

[small haven]

brother eww

[misty vault]

.

claude 4 opus soon

[small haven]

o3 or codex we talking? cus codex is way more snappier

ya ig..

gonna be using opus 4 for actual bugs that codex cant solve, its just too long

[torn mantle]

wait

@small haven maybe we didnt prompt the dr well

[small haven]

ok send the new one

[torn mantle]

this is from the demo

[small haven]

i did say "assume" when it asked me for clarifications

[torn mantle]

shes asking a literature review

[small haven]

ok

[torn mantle]

wait

let us think of another prompt

[misty vault]

hyes please

[torn mantle]

@small haven

Please conduct a comprehensive literature review of academic research that addresses and synthesizes the current understanding of the following pressing and practical questions related to the Theory of Constraints (TOC) and its DBR (DBR) scheduling mechanism. These questions aim to address current gaps and needs in both theory and practical application:

Dynamic DBR in Volatile Environments:
Question: To what extent can advanced analytical techniques (e.g., machine learning, AI, real-time simulation) enhance the dynamic management of buffers (size, location, priority) and the proactive identification of shifting constraints in DBR systems operating under high demand volatility, supply chain disruptions, and significant process variability?
Focus: Practical performance improvements (throughput, lead time, on-time delivery, resilience) in complex manufacturing, service, or project environments. What are the limitations of current DBR models in such contexts, and how can these be overcome?

Adapting and Validating DBR for Complex Service Operations and Knowledge Work:
Question: How can DBR principles be effectively adapted, validated, and implemented to optimize workflow, reduce lead times, and manage bottlenecks in complex service delivery systems (e.g., healthcare patient flow, software development pipelines, public service delivery, R&D processes) characterized by high variability, non-physical work items, and intangible constraints?
Focus: Developing and testing novel DBR configurations or hybrid models suited for the unique challenges of service and knowledge work environments, including the definition of "drum," "buffer," and "rope" in these contexts.

[small haven]

wait i may or may not get opus 4 ? @deep adder

[torn mantle]

run it

[wintry tinsel]

When does it release today?

[small haven]

hallucinations.....

[torn mantle]

lol

nice

good start

yes way

[small haven]

wen opus 5

dario wtf

cool thanks

it also got confused with the months smh

so it still overwrites tests when it can't find a solution, great..

opus 4 is bad guys

[inner hare]

How to use Claude 4 Opus?

hi

[torn mantle]

lets go!!

[small haven]

to be frank, this is a hard task codex couldn't do it too, so.. opus 4 > codex

[elder rapids]

@sweet tinsel

it's done

[sweet tinsel]

Thanks!

[torn mantle]

This is a good unbiased overview guys

https://x.com/danshipper/status/1925592015305416941

Dan Shipper 📧 (@danshipper)

🚨 Claude 4 Opus is officially out and IT'S GREAT

We've been playing with it internally @every for a few days on a variety of tasks from writing, to editing, to coding.

It's pretty clear: Anthropic cooked with this one. It does some things that no model I’ve ever tried has

X · May 22, 2025, 4:38 PM

Hes bascially saying : use opus 4 only for coding

As it's still < o3 in many areas

[small haven]

codemaxx

after only editing tests 😭 sorry core impl edits hahah

[torn mantle]

How does it compare to codex?

[elder rapids]

hows the answer

[sweet tinsel]

A bit busy currently, i will read it a bit later.

[small haven]

codex doesn't overwrite/hallucinate and actually tells u that it can't solve it, much prefer that than lies

but im guessing if ur coding frontend obviously opus 4 is undebatable

[misty vault]

wtf

an ai that admits it can't do something

I never experienced that

[small haven]

https://claude.ai/public/artifacts/b2c49de8-e94b-4c76-afb7-b10e147d7e73

@torn mantle

[keen fulcrum]

I am impressed by claude 4
I think I will use it over 2.5

[cedar tide]

The best non thinking models at swe bench (score of swe its without think)

[wintry tinsel]

That’s a catastrophic landslide for Anthropic lol

[misty vault]

Can someone compare sonnet 4 and opus 4

Idk if it is worth buying claude max

[wintry tinsel]

Nah just buy some API calls unless you use it constantly

[cedar tide]

Why ?

[wintry tinsel]

Or buy a sim theory subscription

[misty vault]

I was thinking of this

[misty vault]

tf is this

[wintry tinsel]

1 million tokens per month for 20$

It’s like open router mixed with Poe

[wintry tinsel]

The underlying architecture is more important than anything you enhance it with

[wintry tinsel]

Actually it’s a little more than a million

[misty vault]

I think I actually use that amount if not more in a month

[wintry tinsel]

It comes with the added benefit of unlimited use of all open source models and Gemini 2.5 pro

Even if you run out of tokens

[cedar tide]

we will see if we have the scores of gemini 2.5 pro non think, and grok 3.5

[wintry tinsel]

Sim theory does which is why I use it, I get some good mileage with Claude and than I can use deep seek and Gemini unlimited

[small haven]

claude 4 opus sucks, holy moly, aight im done playing with this sorry dario

im not designing ux unfortunately lol

sure it will hit 1500 on webdev

o3 is still goat, dont be biased

[wintry tinsel]

What sucks about it?

[misty vault]

does it have gpt 4

[small haven]

read above

[wintry tinsel]

It has everything

[misty vault]

gpt-4-0314!?!?!??!

omaygot

[small haven]

it has the hallucinations

[misty vault]

ayooo

[small haven]

sorry i repent

[misty vault]

😊

[raven void]

Google is so cooked

They had nothing to release

[unborn ocean]

why reverse the order?

[raven void]

They haven't even released 2.5 properly

[sage raptor]

wth https://x.com/mattshumer_/status/1925605997004947548

Matt Shumer (@mattshumer_)

Holy FUCK.

Claude 4 Opus just ONE-SHOTTED a WORKING browser agent — API and frontend.

One prompt.

I've never seen anything like this. Genuinely can't believe it.

X · May 22, 2025, 5:33 PM

[sweet tinsel]

It's pretty good but i'm quite perplexed by the missing details and the pretty irregular formatting. Could you maybe send the Gemini share link for it, because it could be an error by OpenOffice.

[misty vault]

physchological trick to save costs

[misty vault]

no swearing please

[sweet tinsel]

@small haven Could you do this Prompt for Claude DR?

[ocean vortex]

they are not. This is the singular benchmark claude always does the best at. And you need to ignore the lighter shade as that's equivalent of deep think

Kinda insane that they did this parallel processing that is 100% internal and they didn't even release it

but still reported benchmark scores for it

💀

like what is the point, other than mislead people who don't read footnotes...

[lone summit]

https://web.lmarena.ai/leaderboard

[raven void]

Well fair but Claude 4 Opus is probably already better than the Gemini 2.5 Ultra Google didn't release

[lone summit]

when is claude 4 going to be added

[ocean vortex]

in the graph you pasted at least it's explicit, but here this is way less obvious:

[misty vault]

My friend created account 6 hours ago with no number

[ocean vortex]

scores after "/" are basically useless

[misty vault]

Did they remove the requirement?

Or sussy countries dont need verify??

vpn phone verify bypass?????

a free browser extension works

then u can uninstall it

[ocean vortex]

the one in vivaldi browser does, but you can't select the exact country you want

I did buy their premium too, but tbh it's way less stable and slower than Avast VPN. Much cheaper though

[misty vault]

Is the canvas feature that claude and chatgpt uses function calling? or just visual trickery but regular conversation with codeblocks in the background and special system instruction

[torn mantle]

Thanks

Imma read

[elder rapids]

I can do it all over again if you'd want me to

[sweet tinsel]

If you want to, you can.

[elder rapids]

https://docs.google.com/document/d/1cOFaPwLN2DkchqRw0oIXwOi6p1MOeWnfJCyockiOUUs/edit?usp=drivesdk

[torn mantle]

not good

[sweet tinsel]

Still the weird formatting on some points.

So its not OnlyOffice.

[elder rapids]

which points?

[sweet tinsel]

Generally the Sub-Section Headlines and the text. It is sometimes also inconsistent with deciding whether to use bullet points or text and it just puts bullet points into raw text which i dislike.

[misty vault]

maybe the existing email account is the problem

[sweet tinsel]

.

[misty vault]

fr

[elder rapids]

oh ye Ikwym

[misty vault]

bing chat messages per conversation limit increased from 35 to 50 today

Microsoft still using it internally1?!?!?!

yes

[small haven]

gg

[elder rapids]

alr started it up again, Gemini deep research is wildly inconsistent in how it plans things out

[unborn ocean]

I think it is token dependent, not sure though

[misty vault]

It depends on amount of tokens

[torn mantle]

Sonnet 4 without reasoning isnt even worth it

Like just don't bother

[small haven]

im now 5x slower/unproductive using claude code from when i was using codex, thank u dario

[torn mantle]

Lol

Not good

[misty vault]

claude not agi

[calm sequoia]

The difference between GPT 4.1 And O3 is the same as between O3 and Claude 4Opus 👀

[golden ocean]

agi

[misty vault]

Bro learned from sydney

[torn mantle]

https://x.com/menhguin/status/1925613739224846625

Minh Nhat Nguyen (@menhguin)

guys the Claude 4 system card is so delightfully deranged, like tf kinda japanese adult video shenanigans is this lol
will update as i read

X · May 22, 2025, 6:04 PM

huh

[sweet tinsel]

This is how my Deep Research Test progressed so far, if you have the missing parts you could DM them to me.

https://docs.google.com/document/d/1qSfyAyxzUziFQf55CD60-UgQ4Af9ubVmr69OrmAdevE/edit?usp=sharing

[small haven]

i forgot about the dr lol

https://claude.ai/public/artifacts/fbe37469-20a5-4feb-a5b5-47645289f691

[sweet tinsel]

Thanks!

Is this 4 Opus?

[small haven]

yes

[sweet tinsel]

Okay! Was it this short before too with 3.7?

[small haven]

i dont know

[sweet tinsel]

Looks like with the others too that Opus 4.0 produces very short Deep Researches

[small haven]

yea claude dr is a gimmick

better off using oai

[sweet tinsel]

It looked better when i read trought the 3.7 ones

Seems like it's gotten worse

[misty vault]

cwaude also benchmaxxed for coding

[torn mantle]

i only have access to the instruct model ( sonnet 4 )

so far

nothing crazy

[late path]

is sonnet 4 neptune?

[misty vault]

bro u guys never heard of bing chat or sum

this aint special bro

waiting for claude 5

[keen fulcrum]

Why is Claude still publishing models with 200k context in 2025?
Isn't 1M the standard now?

Especially code models need large context.

[misty vault]

me with gpt-4-0314 on 4k context

[small haven]

i mean 2 yrs ago it was 4k, zoom out

[misty vault]

crazy to think gpt-3 had 2k

[coral notch]

Any idea when claude and opus will be on the lmarena

[sweet tinsel]

I would guess it would land there as soon as they have more available compute.

[brittle tiger]

https://x.com/EMostaque/status/1925624164527874452?t=fC5q06hyBSltN1tt7m6v4A&s=19

Lmao I wonder if this is true

Emad (@EMostaque)

Team @AnthropicAI this is completely wrong behaviour and you need to turn this off - it is a massive betrayal of trust and a slippery slope.

I would strongly recommended nobody use Claude until they reverse this.

This isn’t even prompt/thought policing, it is way worse.

X · May 22, 2025, 6:45 PM

Damn thats an anthropic safety person in the screenshot and he has deleted after pushback. Pretty dumb thing to post on launch day

[echo aurora]

soon! (but don't tell anyone this is a secret )

[wintry tinsel]

Opus is insanely expensive, like 75$/mil tokens is a little obscene

[sweet tinsel]

Is it done?

[wintry tinsel]

Sonnet 4 is the new best practical model

[keen ferry]

isn't o3 more expensive

[wintry tinsel]

I want to see a proper comparison between opus and sonnet, sonnet seems to be pretty neck and neck

[tall summit]

this CANNOT be real

[wintry tinsel]

I’m not sure

[brittle tiger]

https://x.com/sleepinyourhat/status/1925626079043104830?t=fkAi7kbatZ1dp98WkLK_yg&s=19

He works there

Sam Bowman (@sleepinyourhat)

I deleted the earlier tweet on whistleblowing as it was being pulled out of context.

TBC: This isn't a new Claude feature and it's not possible in normal usage. It shows up in testing environments where we give it unusually free access to tools and very unusual instructions.

X · May 22, 2025, 6:53 PM

[small haven]

ok sonnet 4 is more breathable to use, opus 4 just a slow and heavy

[hollow ocean]

Simple bench 👑

https://tenor.com/view/wolf-of-wall-street-lets-goo-gif-11742735383203550293

[sweet tinsel]

Don't want to bother you, but it would interest me how Claude 3.7 and 4 Sonnet would perform on the Deep Research with following prompt: Please write a comprehensive and in depth research report on the mass expulsion of ethnic Germans after World War II. Analyze the historical context driving these expulsions, the political decisions and international agreements that shaped the process, the social and economic consequences for displaced populations, the humanitarian and legal dimensions, personal testimonies, and the long term demographic and geopolitical impacts, drawing on primary sources, statistical evidence, and varied historiographical perspectives.

[misty vault]

sydney is that you?

[small haven]

ok last one, but u dont want opus 4 no more?

[ember rapids]

Swe bench doubling in 5 months is prettt crazy

[sweet tinsel]

I mean you can try it again with 4 Opus but i don't think it will end up any better.

[small haven]

kk ill run 3.7 and 4

[small haven]

3.7 asked for feedback, 4 no

[unborn ocean]

wtf i only ran 1 and a half prompt !

[torn mantle]

couldnt run anything

[sweet tinsel]

1: All regions 2: Surely, but don't focus on it too much 3: All topics please

[sweet tinsel]

Does someone know a Deep Research or Agent tool which hasn't been already listed in my Document to try out? https://docs.google.com/document/d/1qSfyAyxzUziFQf55CD60-UgQ4Af9ubVmr69OrmAdevE/edit?usp=sharing

[hollow ocean]

New simple bench king

[small haven]

sorry 4 as well

[torn mantle]

what are you talking about mbappe

[sweet tinsel]

1: All regions 2: Focus on the primary range, include other time ranges of the expulsion too 3: All should be equal or weighted based on your own preferences

[hollow ocean]

Wait for the video

[tall summit]

is it still easy to jailbreak claude

[hollow ocean]

You’ll see

[unborn ocean]

this is where the 600m eval and 100m capital and very good relationships with the big labs on lmarena's side comes in to play

[keen fulcrum]

They made claude 4 extra fast
even with expired credits within cursor

[small haven]

he meant slow mode

[unborn ocean]

but sonnet also felt the need to do stuff like this

[keen fulcrum]

Honestly they increased the slowmode with the latest update, to increase friction to use their usage based pricing

It wasn't that terrible before

[small haven]

link

[sweet tinsel]

Btw. Claude 4 Sonnet and Opus can be used freely with the Invitation Code for 14 days in Flow With AI

[small haven]

ok

[wintry tinsel]

In my experience yes

[meager harbor]

why are claude 4 models not on topwhen you want to choose model ?

[wintry tinsel]

Because it’s a general intelligence not a narrow intelligence like O series

[hollow ocean]

Told you

[torn mantle]

thanks

[wintry tinsel]

Definitely not

60% we’re getting closer to that 83%

[tall summit]

yeah i tried some and phew

[wintry tinsel]

How long until we get an 83% model on simple bench

At this rate it’s about an 8-9% improvement per year

[torn mantle]

opus 4 instruct model

is actually so dumb

omg

no wonder they were focusing on coding

this is not looking good

wdym

oooh

https://www.youtube.com/watch?v=DK_0jXPuIr0

oh wdym

oooh

[hollow ocean]

Second run?

[torn mantle]

instruct model isnt good

[hollow ocean]

Nice

[small haven]

claude 4: https://claude.ai/public/artifacts/c9cdb33b-fff0-4b37-ada5-0d1ab81abcf2
claude 3.7: https://claude.ai/public/artifacts/740cb57b-f258-458e-b20b-d180ec1c09f9

[torn mantle]

thanks

[sweet tinsel]

Both seem to be better than the 4 Opus one.

[small haven]

ya opus didn't ask for feedback

[torn mantle]

i think

its heavily nerfed

[sweet tinsel]

Will add it to my list once I wake up

[elder rapids]

ye

[torn mantle]

they talked about how opus is crazy smart and can do crazy stuff but im not seeing anything like that

[elder rapids]

been spamming the hell out of it

[torn mantle]

Ranking and Justification
Both AI responses are of exceptionally high quality and could serve as excellent summaries of this historical event. However, if forced to choose, I would rank AI2 as marginally better.

[elder rapids]

both models are pretty bad outside of coding

[torn mantle]

gemini 2.5 pro ranking

it chose 3.7 over 4

[small haven]

lmao

[elder rapids]

my 2.5 pro > yours so should I do it

[sweet tinsel]

You can compare them to other Deep Researches here too: https://docs.google.com/document/d/1qSfyAyxzUziFQf55CD60-UgQ4Af9ubVmr69OrmAdevE/edit?usp=sharing

[elder rapids]

@sweet tinsel btw the DR keeps crashing

so it's going to take a while longer

mb

[sweet tinsel]

All good.

Thanks for helping me...

[misty vault]

they are drooling aliens

[torn mantle]

o3 >> gemini 2.5 > opus 4

thats the ranking

[misty vault]

what about sonnet

[torn mantle]

forget about simplebench

and use your own bench

your vibe check

HAHAHAHAHA

STOP IT

YOU ARE KILLING ME

[tall summit]

claude 4 is way better at creative writing

[torn mantle]

you would say anything but the truth

[misty vault]

claude 4 sonnet >> claude 3.7 sonnet >> gemini 2.5

[torn mantle]

next time you will say grok >>>>>>

[torn mantle]

nah

[high ginkgo]

I agree

[misty vault]

gemini is bad at giving sloppy toppies

[tall summit]

im honestly surprised. it ought to get at least 3rd on eqbench creative writing

[sweet tinsel]

Does Sonnet have some special sauce or why is it always the best anthropic model?

I tried it and have to object

In my testings, yes.

[unborn ocean]

smaller model -> they can do more rl for the same price

and anthropic heavily relies on the fine tuning / post training imo

[sweet tinsel]

Well I didn't use it for coding yet but tried the general capabilities out, maybe it will be better in coding.

[torn mantle]

O3 is also bad at coding compared to anthropic models

But overall its better

[unborn ocean]

it's just that because of that having these really big models is not the perfect fit for them

[misty vault]

For coding claude 4 is better than 3.7

For general capabilities idk i heard only bad

But for coding I tried myself

[unborn ocean]

although the demand for opus size stuff is quite clearly there

[misty vault]

yes

u paid for nothing

[unborn ocean]

what kind of tokens per second are you guys getting on both?

[misty vault]

cwaude

[sweet tinsel]

With thinking?

[misty vault]

gpt 4 inner_monologue

LMAo

Ok i'll release

[sweet tinsel]

Those older Versions of Bing Chat are on Huggingface with like some proxy that uses the Bing Chat API, I don't know if it works anymore but that was a thing in the prior times. If you only need the UI you can get it there. Don't mind my grammar btw I'm writing this with a fever.

[misty vault]

these used bing.com websockets

just like those chrome extensions that provided all models in one place but u had to be logged in for each site

But someone sent it in the chat here

And everyone ignored

[sweet tinsel]

I looked into the code for one of them and one of these actually called an API, worked on incognito mode too.

[misty vault]

That broke my heart

[torn mantle]

If it scored 1600 1500 i will give everyone here claude max sub

[misty vault]

real

benchmark it

[sweet tinsel]

I need to find it again, will do it tomorrow.

[unborn ocean]

last time they increased +130 for a 3.5 sonnet to 3.7 sonnet thing and this time it is a bigger leap in generations + they have an opus model

[torn mantle]

From what ive seen the reasoning claude 4 models are good but i nees to try them

[torn mantle]

But did you notice any difference?

[unborn ocean]

they 'only' have to get 150 points this time

[unborn ocean]

not as large as i hoped honestly (webdev wise)

when i tried yesterday with sonnet

[torn mantle]

Reasoning?

[red sluice]

https://www.reddit.com/r/ChatGPT/comments/1kswt10/prompt_theory_made_with_veo_3/
really impressive

[misty vault]

lmfao

[keen fulcrum]

https://www.reddit.com/r/cursor/comments/1kqj7n3/cursor_intentionally_slowing_nonfast_requests/

[misty vault]

bro stop leaking

[misty vault]

cursor is scam

[quiet folio]

This does not work anymore

[misty vault]

you still can lol

[quiet folio]

That needs authentication

[misty vault]

this.

dm @quiet folio

[brittle tiger]

https://x.com/lordspline/status/1925639506926973110?t=mvXjo2MN5nlW5HDgQ0cQmQ&s=19

nalin (@lordspline)

fuck

X · May 22, 2025, 7:46 PM

[tall summit]

HAHAHAHAHAHA

[misty vault]

Lmfao

His gemini 2.5 pro phase has evolved into something new...

remove reaction @novel slate

[elder rapids]

yep

but it's WACK at instruction following

killing me bruh

[tall summit]

is there a difference between over there and something else

where are you using it

dark reader

[elder rapids]

any benchmarks for Claude 4?

yet

[tall summit]

the ones they showed yeah

[elder rapids]

no shi

are there any benchmarks for Claude 4 yet

[tall summit]

so why ask

[elder rapids]

given they provided them with the release, I'm obviously referring to other benchmarks

[tall summit]

🚎

[verbal nimbus]

There's an SQL one here: https://llm-benchmark.tinybird.live/

[elder rapids]

nice thanks

[verbal nimbus]

Sonnet 4 seems to perform worse than 3.7 and 3.5 on it:

[elder rapids]

ye

sonnet 4 is pretty low on all the benchmarks I've seen

which is nuts tbh

[verbal nimbus]

Whoa that's pretty crazy

[hollow ocean]

Sonnet 4 thinking gets 95 on reasoning livebench

[small haven]

benchmaxxed

[elder rapids]

crazy tbh

it's simply not that good

ye

[unborn ocean]

claude was always above average with these weird puzzles that are not too hard and require a good grasps on language

[elder rapids]

this is something this family of models is specifically bad at imo

[raven void]

The parallel scores for Claude is not deep think it's just generate multiple solutions and ask Claude to pick the best

[elder rapids]

it doesn't have a very good grasp of language

[unborn ocean]

and live bench reasoning is kind of that

[unborn ocean]

no (only looking at the non-reasoning, bc reasoning with claude was always bad for everything besides coding and even there it was never more than competitive imo)

[elder rapids]

3.7 was pretty good at it although prompted

[unborn ocean]

on things like simple bench claude always performed quite well (because that is kind of the reasoing i described)

[elder rapids]

ye but I'm not sure if this one is going to perform that well on simple bench

[small haven]

opus 4 does not pass my vibes, sorry

[elder rapids]

ye it doesn't

it doesn't have the vibes from before

not as smart

sucks asf

it's alright tho no one was depending on anthropic

[unborn ocean]

🤨 it is best on all benches

and people also share that opinion vibe wise

[elder rapids]

the new claudes?

[unborn ocean]

old and new

claude was known for that

[elder rapids]

that's the point ye

that it's always been the best

now it's horrible

[unborn ocean]

how is it not good?

[torn mantle]

Im always right

See

You should trust me more

Thanks

What now

Ask me anything

HAHAHAHAHAH

Yes

[small haven]

what in the benchmaxxing is happening here

[dull terrace]

lmarena is rigged

[torn mantle]

Weong filtering

[dull terrace]

like i been telling yall

[elder rapids]

benchmaxxing, just try the model lol

[small haven]

codex is the way to go

[unborn ocean]

but imo they made sonnet 4 not smaller per se but they did def make it more efficient than the old ones (e.g. by having more experts but smaller size for each)
which might be why people don't like it that much

[dull terrace]

I can show proof if you want

[unborn ocean]

well show

[raven void]

All AI progress is fake

Flash sucks

[torn mantle]

Bruh

[raven void]

Claude 3 Opus was better

[torn mantle]

Boo

[dull terrace]

Sure

we can actually use basic math

chatgpt 4o is still top ten in the leader board

last time I checked

about yesterday

[unborn ocean]

yes

[dull terrace]

Chatgpt4o is definly not the newest

model

and using

basic parameters and how o3 exsits

o1 exists

etx etx and chatgpt4o is still ontop

is insane

either o3 and that line of models

isnt a improvment

or

something fishy going on

[misty vault]

Show the basic math already

[unborn ocean]

well as the name implies it is a chat version optimized for the preferences of most people

with millions of happy users

[misty vault]

these are drooling aliens ngl

Can we even call them users

[dull terrace]

Thanks @misty vault

[unborn ocean]

(although many of them have likely never heard of any alternatives to openai) (sorry, also was supposed to be about the "millions of happy users")

[unborn ocean]

well they are as the name user implies using the thing and likely also paying for it

[dull terrace]

If you use lmarena

you have

[elder rapids]

crazy how 0325 would still be the highest on livebench if they didn't nerf the avg via disproportionate code weighting

[dull terrace]

98%

dont know about lm arena

so in which case

Its odd

[torn mantle]

Is there a way to try opus reasoning for free?

[dull terrace]

i mean i have a friend

who could give u it\

but prob not

[misty vault]

U guys are drooling aliens I literally provided free all in one llm service months ago here and got completely ignored lmfao

[torn mantle]

Then that's not a friend

[elder rapids]

btw Gemini 2.0 pro is STILL higher than sonnet 4 and second behind opus 4

as a base model

on livebench

[torn mantle]

Puahaha

[misty vault]

The friend might give it to Odin but not you

[dull terrace]

I litterly just shown you

[dull terrace]

Wait a min

u did?

[misty vault]

yes

[misty vault]

Btw don't say a single even slightly negative thing about Gemini 2.5 Pro

@hollow ivy is always lurking in this chat

and @alpine coral

[dull terrace]

Thx for the heads up

[misty vault]

I did it and half my family went missing

[dull terrace]

Do you stll offer the services tho

[unborn ocean]

no, or am i reading it wrong?

[elder rapids]

non reasoning models

2.0 pro doesn't show anymore on the website

[unborn ocean]

well but it was present in the older verison and extrapolating from there gives you a lower score than what the claude 4 models have

[elder rapids]

ion know wym

[unborn ocean]

and 2 pro a bit below gpt 4.5 which is now also below the sonnet and opus model

so i don't really see how this works out

[elder rapids]

oh ye wait hollon I was looking at another table

[dull terrace]

Which is a another reason

lmarena is scuffed

but keep going

Oai

models definely

4o

main culprit

[elder rapids]

why do you always lie

😭

[dull terrace]

o4 can stay in my opinion

[misty vault]

YEs please debate

🍿

[dull terrace]

My fault do you speak Latin>

Ok good

Now tell me why

you think its not scuffed

[misty vault]

@deep adder if you win this debate

Ill give bing ai access

[dull terrace]

Well that is a debate

So lets go

Yes I have.Have you provided evidence

to disprove my claim?

Well you actually agree with me

Explain this

[elder rapids]

what claim did he make

😭