general

[hollow imp]

Because it's exceeding the thinking limit set by arena

[echo aurora]

We're collecting these possible false positives in this thread: #1447983134426660894 could you also share the prompt used?

[verbal nimbus]

👀 I wonder how it does

[uneven peak]

I hope we get Opus 4.6 thinking 32k 🥀😔

[honest verge]

Is it just me or opus 4.6 thinking makes less detailed results than opus 4.6 no thinking?

[uneven peak]

Nah 4.6 thinking make more detailed results

[gilded shell]

HAY

[uneven peak]

Ka

[icy yew]

Ehh idk

Can you show like examples

.

[crystal rapids]

Why have a model up if it always exceeds the limit, also on the 2nd attempt it crashed after it's thinking so rather a token limit perhaps. I would be guessing just instability, everyone trying to use it all at the same time, ect.

[summer hound]

Yup that limit is screwing almost every prompt I try. I guess this model thinks so much more

[keen beacon]

I can’t even accept the terms 🙁 I think it was because the captcha was bugged 🙁

[patent oracle]

Q: Has anyone ever managed to convert a painting into a photoreal image using Nana Banana Pro? I can't seem to get to work tried all kind of prompts
Thank you!

[summer hound]

yea dude this has to get fixed. unusable

[loud verge]

Guys

Does grok 4 search not work anymore?

[toxic verge]

This is an image edit no prompt

[echo aurora]

We are looking into this.

[thick pawn]

It's been a problem for ages and it still hasn't been fixed

[echo aurora]

Can you try a different browser? With all of the issues you've seen today something is for sure off.

[echo aurora]

It is worth noting this something went wrong error message is the generic error message that'll trigger for many different reasons (including rate limit).

[thick pawn]

I get it a lot, sometimes I get that message, but I can just refresh the page and the image is actually generating perfectly fine. Other times I get the message for seemingly no reason multiple times in a row, and I have to wait a few minutes before it works again, and that's not due to me hitting my rate limit

[burnt sinew]

anyone here know how to get free ide usage for claude opus 4.6?

[toxic verge]

Idk how you guys will be able to fix captcha it does it’s job good if you loosen it it will be easier to bot. It’s a double edge sword

[frosty shuttle]

Same thing here. The model thinks for a long time and is interrupted by the response limit when it's time to give the answer. I think only removing the response limit will make it work correctly.

[stone cape]

kiro ide you can sub for $0 for first month i think if lucky

[toxic verge]

I don’t really follow social media. I don’t have a Twitter or anything like that and I kind of stopped using Reddit but you know how I knew that anthropic was about to release a new model?

[stone cape]

This seems VERY interesting https://discord.com/channels/1340554757349179412/1469144854264021174

[toxic verge]

https://youtu.be/tjW_gms7CME?si=9z9_yVZfgLZTV39j

Cause he always goes on the news before they release a new model and always starts preaching

[primal orbit]

opus 4.6 easily gonna be top 1 model on lmarena by a large margin

[toxic verge]

Ofc it’s new

His tripping if he think it’s gunna replace software 🤣

It’s like saying ai video model will replace directors

[north obsidian]

In code, but in text? I don't think so

[shrewd citrus]

nah ai will definitely replace software engineers tho

like out of all the things any ai learnt

what was the thing it mastered first

it’s coding

apparently coding is the easiest thing for an ai to do

im gonna say that by the end of this year ai wld be better than 60-70% of engineers currently working

[delicate fable]

It already is I feel like with opus 4.6

[toxic verge]

Ya’all crazy

[proud bobcat]

My favorite moment is when OpenAI released GPT 5.3 codex but didn’t set up any api for it

This model was so rushed

Really makes me sad cause I thought they’d take their time and not just rush out a release

[frosty lava]

is the gpt 5.3 codex is available on lmarena ?

very interested to do a comparison between this one and the opus 4.6 with same prompt

[proud bobcat]

Not yet since OpenAI hasn’t released it on their api

[frosty lava]

oh alright

thank you

[nimble snow]

😢

i cant keep doin dis every prompt

[cloud zinc]

you are robot

[frosty lava]

if you have vpn it might be cause of it

[nimble snow]

i dont

im js cooked

[frosty lava]

then idk

the opus 4.6 non thinking was already better than the 4.5 thinking in my test so i wonder what the thinking version can do

[nimble snow]

what u play on rovblox mud

[frosty lava]

Nothing cause i can't find any good games for now

i already played every good existing one and im tired

[toxic verge]

Why
. 5 series all cursed

[nimble snow]

u play demon hunter?

[toxic verge]

Maybe be your browser. Try login in ans out your account

[maiden fulcrum]

What is the rate limit for battle mode in the arena and when does it reset?

[frosty lava]

no sorry

[broken elm]

is lm down ?

[past trail]

Is there any ai or website that can build full stack apps for free ...without any paywall

[sturdy mica]

i literally only play armored patrol on robiox

[robust sonnet]

Yo guys

What is the rate limit on opus 4.6 in code mode?

[maiden fulcrum]

These reCAPTCHAS are so annoying, It doesn't go away at all

I selected all fire hydrants, yet it is telling me to try again

[toxic verge]

Login in log out clear history and try different browsers

[frosty lava]

i think gpt 5.3 is a big improvement actually and is better than opus 4.6

can't wait to see even stronger model but the improvement is already cool

[stoic solar]

Heyy! Anyone have opus 4.5 free trick? I really need it like we can share files too

[frosty lava]

opus 4.6 seems to do exactly what your asking but its almost every time less aesthetically pleasing so idk

[stoic solar]

Where can i access opus 4.6?

[frosty lava]

both claude website and lmarena

[stoic solar]

Okay, thanks buddy. Does it have file upload and vision?

[frosty lava]

Sorry i don't know actually

[stoic solar]

Oki

I checked it, it doesn't have 😞

[frosty lava]

its here, try on lmarena

oh my bad

i just understand what you said

[stoic solar]

No I mean the upload and vision capabilities

[frosty lava]

yeah sorry

[stoic solar]

Yeah not an issue

Do you guys use antigravity?

[frosty lava]

Yes but im mad at it cause even with the option "always proceed" it will ask me to confirm every prompt, ive seen other people with the same issue and can't find a fix

[stoic solar]

Yes and the rate limits sucks

[frosty lava]

yh

[stoic solar]

Any other good ide?

Other than vs code with Co pilot

Cursor

?

[frosty lava]

I use cursor so i don't know if there's a better one

[stoic solar]

Depends on model

[thick pawn]

Videos are currently having an infinite generation issue it seems

[maiden fulcrum]

It didn't work

[frosty lava]

lol i tried opus 4.6 its been thinking sooooooo much ive never saw that before

brainstorming on my prompt

[thorn lantern]

thoughts on opus 4.6 vs. opus 4.5 for software engineering tasks for actual software engineers? In terms of real world usage? I noticed opus 4.5 edged out 4.6 in 2-3 metrics on the official release document, but opus 4.6 did better on most.

But in terms of real world use, what are people's thoughts comparing the two thus far? I've read that opus 4.6 can be "too agentic", but not sure if that's a universal opinion or just a one-off

[frosty lava]

i don't know why but its been 6 minute thinking non stop on my single prompt

no lag its just really thinking

too much

[thorn lantern]

I read that's expected, but I bet that's eating a ton of tokens..

and for simpler stuff, that's probably not needed

For example for devs who like to take control of the process and implement step-by-step with less "do it all for me"

[frosty lava]

if you want i can send you the whole thinking for a single prompt (its not done yet)

its impressive

how much it actually think

[thorn lantern]

That would be valuable to see, if you don't mind. you can dm me

[frosty lava]

yeah i send you that

[thorn lantern]

I've used 4.5 a lot, but haven't experimented with 4.6 yet

[frosty lava]

i sent you the thinking

[thorn lantern]

Got it, thanks! I'll respond in dm

[burnt sinew]

literally nevermind i forgot how good good chat + bad cli is

[toxic verge]

I’ve been testing this out for a little

Try making a new account not Gmail

[maiden fulcrum]

hmm

why

[toxic verge]

I’m not sure exactly but it helps sometimes

Unless your sending requests to fast to model

Then your going to get them

Also try resetting ur modem if all else fails

[frosty lava]

Okay i figured out something, opus 4.6 for some reason, when given a complex work to do, it will think without actually writing code, you can see a huge thinking during 5 / 6 minute then error from lmarena, but no code written, so you have to tell it to actually write code

its weird but yeah it is how it is

[burnt sinew]

When's opus 4.6 going to be preliminary on leaderboard

[echo aurora]

We'll be sure to post an announcement when the leaderboards get an update with opus 4.6

[slim spire]

Does max use opus 4.6?

[frosty lava]

Hey, is it an issue from lmarena that opus 4.6 actually think so much it end up saying an error happened please try again ? it's actually very very often, the model will think 4 / 5 minute then just "an error happened"

its when given a complex work

code

[left lodge]

New feature in testing. 👀

Take a screenshot
is just not working

Why is it even there? To take screenshots of the current session?

[slim spire]

it's still in testing

[left lodge]

And the transparent error popup is nice but why is it at the bottom?
It should be somewhere the bg is clear

[slim spire]

it's not even out it's still in testing so it might not work that correctly yet

don't expect anything that is still being tested to work instantly

[left lodge]

They shipped that means it should work

Atleast somewhat

Models in arena are executing commands! Theya re installating packages?!

[echo aurora]

I believe it has the potential to use all models in Battle (excluding the codenamed models).

[left lodge]

I hope these tools come to text modality or a completely new modality where it have all these tools available to use :>

Without the system prompt of code modality

[echo aurora]

There is going to be a some instability with the model today as it's getting a lot of use as you can imagine.

[frosty lava]

Thank you, and do you know when gpt 5.3 will actually be available on lmarena ?

[echo aurora]

Nex experiment! https://help.arena.ai/articles/4680044864-arena-experiments-image-to-code

[icy yew]

When the API is out

There is no API for arena to use them right now

[echo aurora]

For the most part I won't be sharing details about if/when new features/models/etc. are landing.

[frosty lava]

Oh i see

[echo aurora]

Thanks for sharing this, will flag.

[echo aurora]

Any particular steps to repro this? I wasn't able to get this error.

[rigid holly]

So whats the opinion about opus 4.6?

Cuz i just found out about it like this second

[icy yew]

Good

Dif better then 4.5

But benchmarks say the new gpt 5.3 codex is better

But until the API is fully out we don't know

[rigid holly]

I mean... its out on openrouter so who knows

[frosty lava]

opus 4.6 did a decent 3d world much better than opus 4.5

but gpt 5.3 might be even better

at coding and overall i guess

[left lodge]

Just click the button?
Maybe cause of brave browser, try to replicate it with brave browser on android.

Hey pineapple can you create a seprate coding channels?

Guess the models behind these.

https://019c31d4-e917-7510-a7a6-f593cc0fcf35.arena.site/

https://019c31d4-e917-7b02-b61c-dff4a28a6d74.arena.site/

They are made by sota models of the same lab

[icy yew]

We don't know but we do know it's better in terminal bench which is like a important benchmark so the rest I think Claude wins on it

[left lodge]

But the difference in one version upgrade is so much like bruh what

It could be this single instance but we will see

[icy yew]

Made this with Claude 4.6 thinking
https://019c31d8-ccde-7e39-afa3-188c58cd1868.arena.site/
What do y'all think

[toxic verge]

Who here is a nano pro?

[toxic verge]

I tried

[atomic lagoon]

How do I get this? Is it slowly rolling out?

[royal crater]

Hi

I want to work with my GitHub repo. But it doesn't support github connector like perplexity. So wht to do now ?

Like I want the ai to create pr make changes put commit etc

[icy yew]

Do you know what type of thinking the thinking mode of Claude 4.6 on arena is?

Low medium high or max

[undone saffron]

Me using Opus 4.6 after the announcement:

[sleek phoenix]

why did you edit it

it was correct

[royal crater]

I want to work with my GitHub repo. But it doesn't support github connector like perplexity. So wht to do now ?

[undone saffron]

Vencord
Good

[hollow imp]

I see that daily with opus 4.5 32k. Finally found someone that struggles with me

[golden ocean]

lmfao

real

[austere sundial]

Omg lmarena became Light...
You can't for explosions, you can't for soldiers falling off the horse....
Drastic even chatgpt does that

[fresh urchin]

Guys how good is opus 4.6?

[sleek crow]

@echo aurora

the thinking model dosent work

[gilded kiln]

Please add copilot too!!

[golden ocean]

I'll let my work speak for itself http://localhost:8000/index.html

[undone saffron]

Check if you get [this error](#general message) too

[spare rune]

That’s..

Interesting

[fickle venture]

Copy the prompt you sent him and refresh the website and send the prompt you copied

[fickle venture]

https://tenor.com/view/how-do-we-tell-him-mr-krabs-spongebob-meme-spongebob-meme-gif-24063814

[spare rune]

Thinking is only for code right now

Right

[fickle venture]

No it's on text

[spare rune]

Ohh

[fickle venture]

Code and text

[spare rune]

I might have to test it for creativiry

[fickle venture]

Take advantage NOW because there is no limit

[spare rune]

I lost my laptop so I lost all the motivation to do anything else 😭

[spare rune]

No

I’m gonna use it 1 time

[compact flame]

Guys what glasses mean on ai models

[fickle venture]

I was using it many times yesterday 😭 still didn't hit limits

[spare rune]

Smart or something

[icy yew]

Copilot is just chatgpt

[fickle venture]

Copilot doesn't exist

[icy yew]

Vision

Like it can see images or something

[compact flame]

I think vision is the file icon

[icy yew]

The file icon is pdf

[compact flame]

Oh alright

[icy yew]

Wait Claude 4.6 doesn't have rate limit?

[fickle venture]

Sunglasses is vision
File is pdf upload

[compact flame]

Alright thanks

[fickle venture]

I mean I used it many times YESTERDAY still didn't hit limit

[icy yew]

Like how much tho

[compact flame]

Maybe thinking version only has limits?

[icy yew]

[fickle venture]

I was using thinking

[compact flame]

Oh alr

[fickle venture]

Uhh I think 10-20 messages

[compact flame]

So how good is opus 4.6

[icy yew]

I mean I spammed it yesterday day night and didn't hit any limit

[compact flame]

Or is it just same as 4.5

[icy yew]

Smarter but 5.3 codex wins in one benchmark

[fickle venture]

I love it, I fixed a project that been working for while no other ai fixed it. Thanks opus 4.6

[icy yew]

But every other one Claude wins

[compact flame]

I mean one benchmark is no big difference but still good

[fickle venture]

Probably rn it's free but today or tomorrow will add limits

[compact flame]

Well damn

[icy yew]

Also I think terminal bench is like a important one for agentic coding

🫀

[fickle venture]

Yeah it is

I think like the way it run commands on terminal

[compact flame]

First time seeing chatgpt cook

[icy yew]

They released the codex 5.3 the nano second claude released 4.6

[compact flame]

I wonder if they'll add search to text arena for better results maybe

Like not cutting off ai from internet

[icy yew]

Grok is probably the best for up to date information

[fickle venture]

The heck are these models on arena

[icy yew]

The hell

[fickle venture]

Search already exist

[icy yew]

What ai is named beluga😭

[compact flame]

I mean like search for the text arena not the separate thing

[rigid holly]

So does anyone know the data training cutoff for this model?

Opus 4.5 i think it was early 2025

[icy yew]

Ye it's kinda annoying

[fickle venture]

If I remember I told a model to search online and it actually did while thinking on text arena

https://arena.ai/?chat-modality=search&mode=direct

[fickle venture]

Probably 2024

[rigid holly]

Thats less than before tho

[fickle venture]

Idk you can ask it and it will answer

[compact flame]

I wonder why are the trainings are even cutoff

[shrewd citrus]

because they can’t learn like that much reliable data or

[rigid holly]

Yeah well 4.5 says its early 2025. And 4.6 isnt out yet

[shrewd citrus]

like Claude says oh it has reliable data up to April 2025 but can still get info up to July or something

[icy yew]

Glm 5 is trying to fix this so it has access and gets info without even searching because it searched before

[fickle venture]

To not know what happened today so they won't take control of the world

[fickle venture]

4.6 is out...

[icy yew]

Your late

It's out

🫀 🫀 🫀

[rigid holly]

Not on the model list it aint

[shrewd citrus]

It is just search it

[fickle venture]

It's the last model just scroll all the way down or search opus

[icy yew]

It literally is though

Scroll down

[icy yew]

Like the last model in the selection

[rigid holly]

Ah it was hidden at the bottom

Expected it at the top

My bad

[fickle venture]

They did that on purpose I just can't prove it

[left lodge]

Broo

💀

[fickle venture]

Anthropic being themselves

[icy yew]

https://tenor.com/view/james-doakes-doakes-curious-suspicious-questioning-gif-14390669560130901665

[golden ocean]

https://cdn.discordapp.com/attachments/559403230438883351/1468572366748778734/attachment.gif?ex=69867c58&is=69852ad8&hm=b13740d4b33b98af98027a4ee7c86990f84f61b5d5a0ce1d05516d4faa980145&

[compact flame]

I guess nobody is safe from greed

[bright spade]

the exacution of code dont work ?

[left lodge]

Its on the last because its not on the leaderboard rn.
Sorthing is same as leaderboard and models not on the leaderboard are at the last

[rigid holly]

Well its still early 2025

Slightly dissapointing

Was hoping for more recent stuff to be more available

Like knowing who the new pope is

[fickle venture]

I think they did that because they know ai can see the future before us

https://tenor.com/view/ش-gif-9093367425263329574

[fickle venture]

Well atleast I have a domain

http://localhost:8000/index.html

[rigid holly]

Ok but still. Not even a middle 2025

[fickle venture]

They might do it on the next model because Opus 4.6 was in training in 2025

Probably

[light sleet]

Will gpt 5.3 better than Opus 4.6?

[sterile tartan]

Training for these Minor Improvements?

💀

[sterile tartan]

Probably Not

[rigid holly]

That is a good point and yeah

The 4.1 and sonnet 4 had 2024 data before the upgrade

[icy yew]

Only in terminal bench

[fickle venture]

Nothing confirmed if GPT 5.3 is coming this month just don't believe anything you see. But my guess is No

[rare fractal]

Where does the arena get the money to pay for all these expensive models?

[sterile tartan]

But Codex 5.3 is already Out

[light sleet]

Kk

[fickle venture]

See their blog

[icy yew]

But not the api

[sterile tartan]

SECRET

[sterile tartan]

True

[rigid holly]

I remember that in decdmber it was thoughg sonnet 4.6 or 4.7 was coming out that month

[fickle venture]

Well idk why but I think they trying to improve something on Gpt 5.3

[rigid holly]

Same deal?

[sterile tartan]

I C

Makes Sense

[golden ocean]

SO TRUE

[fickle venture]

I mean internet is just full of lies

[left lodge]

Now i think if this is true , what will they release in place of actual next sucessor of sonnet 4.5??

[sterile tartan]

Like Our Profiles

💀

[left lodge]

Haiku???

[rigid holly]

What about haiku?

[fickle venture]

Haiku is boring no one cares

[fickle venture]

Its okay They hide our identity so yeah

[icy yew]

Never even tried that

Probably dog water

[fickle venture]

Same lol

[sterile tartan]

Truth

[toxic verge]

They get funding

[sterile tartan]

Yeah from A Crypto Venture

[toxic verge]

https://www.prnewswire.com/news-releases/lmarena-secures-100m-in-seed-funding-to-bring-scientific-rigor-to-ai-reliability-302462025.html

They’re burning some money though

Just like most companies

Except they’re not technically a business somewhat

Lightspeed, Laude Ventures

[next ivy]

ngl i thought it was sonnet 5 instead of opus 4.6, did not expect that

[toxic verge]

https://deepmind.us.org/blog/lmarena-funding-hits-150m-at-1-7b-valuation

[rigid holly]

So

In terms of writing long texts do we still need to wait for it to polished or what

Cuz it keeps crashing

Does it still need those funky numbers at tge end of the model?

[left lodge]

Haiku is good combination of speed and intelligence

[icy yew]

Wasn't that for sonnet

Or is it way faster

[high dirge]

is there a way to find out the exact model the max routed to not just the organization

since it could be helpful to see what models are best at what prompts

[spare rune]

Hahahaha omg this is so real pls sen me the link

😂😂😂

[covert iris]

gimme money

[uneven lance]

In lmarena leaderboard Gemini 3 pro ranks 1st while in Artificial Analysis leaderboard Gpt 5.2 pro tanks first

Is Artificial Analysis biased?

They only rank using trust me bro benchmarks...

[golden ocean]

crack bench

[echo dome]

idk if this one works

[echo dome]

who invited beluga into arena bro

[frigid tusk]

why is opus 4.6 the lowest here

[north obsidian]

Yes, this AI just take ur prompt and put the best AI for it to answer u or speak

[icy yew]

To be honest it's actually decent

I wou still go for sonnet for balance tho

[north obsidian]

It can be since google model until xAI

[golden ocean]

so that their money burned less fast because less attention

[frigid tusk]

but its the best of all?

[echo dome]

mr krabs

[golden ocean]

🦀 money 🦀 money 🦀 money

[echo dome]

(that's the answer)

[frigid tusk]

https://tenor.com/view/ahoklollmao-of-course-ohhhh-oh-that-makes-sense-gif-11896286840118452668

[golden ocean]

REAL

[left lodge]

Yeah sonnet is for everyday tasks and you can haiku for fast and quick answer without being too conscious about it being hallucinated or wrong, it is fastest model by claude and is seemed better compared to other similiar size models

[left lodge]

It might be true lmao

It shows improvements from sonnet 4.5 but not from opus 4.5

[echo dome]

wait what just happened to sonnet creator

[echo dome]

wait gemini 3 flash thinking have thinking? i forgor

[fickle venture]

Dam Gemini 3 Flash is the easiest one to jailbreak lol

[left lodge]

Gemini 3 flash literally doesn't have a non thinking variant

[glacial dock]

I’m getting non stop time outs with opus 4.6 thinking, meanwhile with the same question 4.5 has never timed out …is this a bug or is it because it’s brand new and needs more time ?

[left lodge]

Cause of its hallucinations and behaviour

Make minecraft with touch controls

Opus 4.6 thinking
https://019c32cb-de47-77c5-aaf3-30b23c910a2f.arena.site/

Gpt 5.1 codex max ultra premium pro high figh wifi bluetooth
https://019c32cb-de47-78f8-8f93-6c88f6728a2d.arena.site/

Wth is 5.1 even doing wth is that

[north obsidian]

What model made this?

It's good but the floor

[hollow ivy]

Anthropic's Claude (Opus)

[left lodge]

Specifically opus 4.6 thinking

I just said Make minecraft with touch controls

One prompt

[north obsidian]

I liked it

[left lodge]

Gpt 5.2 codex
https://019c32dc-b394-79d2-a3f1-08e74bf4e7ae.arena.site/

Gpt models are so weird rn

[hollow ivy]

opus 4.6 takes forever to answer a simple prompt in arena

[hollow ivy]

OAI is history

[left lodge]

Cause its not made for simple prompts ¯_(ツ)_/¯

[north obsidian]

Yeah

[wind ember]

now even direct chat has captcha?

[hollow ivy]

Anthropic (and Deepmind) have defeated OpenAI

[wind ember]

like come on ...#

[left lodge]

Its literally made to think and reason for the hardest problems

[hollow ivy]

-# later, xAI might join the victors

[left lodge]

It already had from the starting

[wind ember]

this is annoying

[north obsidian]

I never had like it, I prefer Gemini and Claude

[wind ember]

it was only on voting session

[left lodge]

Yeah gpt lost it with gpt 5 launch

[ocean ferry]

can anyone try this prompt for Gemini 3 Pro GA?

Create a nice looking and rich SaaS about Gemini 3 Pro GA by Google Deepmind, it must has a mock about the Gemini 3 Pro Preview which is so lazy and it's fixed on Gemini 3 Pro GA, output in single html, must use tailwind css(cdn) and i don't want shiity website, should be really cool and good and should never use emoji in the html.

https://x.com/i/status/2019458279856877818

Chetaslua (@chetaslua)

🚨 How to Access new Gemini 3GA

go to @arena
battle mode - write your prompt - chances are set at 25%
how to difference
gemini-3-pro with google logo is normal
gemini-3-pro without google logo is GA

show me your demo now

X · Feb 5, 2026, 5:09 PM

i never get it bro

i only get gemini with google logo bruh

so please any1 try it

[wind ember]

not working anymore

[hollow ivy]

does this guy favor Lua language?

[left lodge]

No, google captchas were already on direct chats

[toxic verge]

I don’t know how they’re gonna replace them. They need it for anti bot

Because it’s really effective

[wind ember]

they need to do smth about it

[toxic verge]

That’s the thing I don’t know what alternatives there are I can’t think of any

[left lodge]

Wait i just noticed chat titles are not first prompt

Hmm

[golden ocean]

10 gallons of water per custom chat title

[toxic verge]

Fr

[wind ember]

a lot

[left lodge]

Thats not correct bro 😭

[wind ember]

not sending any prompt anymore

[toxic verge]

Oh yeah, it’s not correct but there’s a bitter drop of truth in there in a general sense

Exaggerated

[left lodge]

Claude opus 4.5

non thinking

https://019c32dd-76b3-7fd8-8986-8e3397f6fb27.arena.site/

This one is nice one too ↑

Gemini 3 pro
https://019c32dd-76b2-7edd-b2a2-750076c5c961.arena.site/

It isnt even rerendering anything 😭

[ocean ferry]

the ga deleted?

[zealous sparrow]

actually it does

It's just a gemini 3 pro model

One is direct with logo
another is stealth with logo

[left lodge]

I literally have zero interest in gemini models cause of their hallucinations and attitude issues

[wind ember]

its only available in battle mode?

[zealous sparrow]

probably so

[left lodge]

Yes

[zealous sparrow]

i have a gen with 2 gemini 3 pros

[wind ember]

mm i see

[ocean ferry]

can u try it with my prompt plz

i have tried it for 2+ hours and i only get the gemini 3 pro with logo

[zealous sparrow]

the GA is also with logo, you can only tell by quality

[shrewd citrus]

does anyone else have the problem where opus 4.6 thinking just thinks for too long

and then stops working

[left lodge]

Just three looks wierd they should have extensions showing

[shrewd citrus]

like 4.5 would think for a minute max before it starts outputting something

[north obsidian]

Better than Claude lol

[left lodge]

It is by claude

👀

[north obsidian]

Hmmm

[ocean ferry]

bruh😭

[north obsidian]

I see it now

Claude 4.5 was better than Claude 4.6 thinking 🤔

[north obsidian]

It made a floor

[toxic verge]

Attitude?👀
.

[uneven lance]

Gpt 5.3 codex is so bad at frontend

But backend it shines

[left lodge]

Yeah i dont like its character, its lazy, doesn't accept its own mistakes, doesn't follow instructions, a literal karen

When i say dont be lazy, instead of doing work it literally says i am not lazy 😭

[uneven lance]

I had to beg with Gemini flash so it follows my lead

[left lodge]

Yeah flash is even worse

[uneven lance]

Within thinking it's the worst

[left lodge]

They dont have any reliability

[uneven lance]

It removes features

When I ask it to add a feature it removes the previous version of the code too 😡

So now I just use GLM for coding

[left lodge]

Glm is good

Its tool use capability i like it but it outputs literal articles even for simple questions

[uneven lance]

A system prompt makes it give one line answer...

I have to instruct it how to use its tools

The model is so agreeable too

Isn't gpt 5.3 codex assisted by 5.2 codex in it's creation?

No wonder the front-end capability is meh

[left lodge]

Good doesn't mean sota

Good is good not perfect

[wind ember]

what is it good at?

frontend?

[left lodge]

Try yourself

[wind ember]

i did

im asking you what is it good a t

maybe im missing something

[spare rune]

Same

[wind ember]

although i still think glm 4.7 is the best chinese model yet

[left lodge]

I haven't used it for front-end

[wind ember]

alongside deepseek v3.2

[left lodge]

What about Kimi 2.5?

[wind ember]

meh

starting to look more like gemini 3 clone

its heavily trained on gemini outputs

[spare mango]

Gemini, the "best" chatbot.

"in 2024".

[zealous sparrow]

what model did you use

Fast or thinking

[spare mango]

Pro.

[zealous sparrow]

it knows what year it is, just does silly messups..

[spare mango]

If it knows what year it is, then why does it do the silly messup.

That's like me saying I sometimes forget my name.

That would reduce my credibility and reliability by a lot.

[north obsidian]

2 y ago I saw a brutal AI error in math, was about meta AI it made all the equation but in the final it was like 255 + 1 it said 257

[shrewd citrus]

i hate it when models say “you are absolutely right”

like why can’t they be right when i first asked the question

[spare mango]

Honestly same.

I have to argue with, and debunk the AI's false claims, after which I realize I just wasted my time arguing with an AI.

[icy yew]

Gork the best in time lines tbh

[golden ocean]

You're absolutely right

[icy yew]

It doesn't really get years wrong or dates

[icy yew]

Gemini 3 when it doesn't hallucinate absolutely cooks especially in language

Sadly it hallucinates like crazy

[icy yew]

Claude ofc

[hollow ivy]

I agree, but i still want to see, if Claude Opus manages to get 100% in this poll, and which model lands second place.

[keen beacon]

how to fix it?

[icy yew]

It just happens I think

Unstable

[keen beacon]

🙁

[hollow ivy]

(..the 2nd poll)

[keen beacon]

but does it go back to normal? because I already tried closing the browser and everything, and it didn’t work , it’s still giving this same problem

[eternal saffron]

website is down?

[icy yew]

Uh no

[keen beacon]

🙁

[icy yew]

Could be a problem on your side

It works for me

[eternal saffron]

alr mate

[frozen osprey]

Does nano banana pro work

It keeps giving errors

[somber sky]

did you use multiple accounts like me?

[frozen osprey]

Nah

[somber sky]

or generate any feminine related?!?!!?

huh????!?!?!?

[sterile tartan]

What are Opus Rate Limits on Arena?

[plucky sparrow]

has anyone tried this on opus?

[modest prism]

Opus 4.6 thinking gives a timeout when thinking longer than a certain time. Any plan to fix it

[vast fern]

@echo aurora are there any plans for adding gpt 5.3

[icy yew]

They can't

The API for 5.3 codex isn't out

[golden ocean]

is there an api for talking to @icy yew

[icy yew]

After the API comes out

[icy yew]

https://cdn.discordapp.com/attachments/1283423642579107905/1302654122008252446/image0.gif

[vast fern]

login

[icy yew]

https://tenor.com/view/respect-moment-gif-15169320514465779908

[keen beacon]

dont work

[frosty shuttle]

The new Claude program is still unusable for me; it thinks for a long time before giving an answer, but is interrupted by the site's limit. Has anyone managed to use it yet?

[shrewd citrus]

exact same thing is happening to me too

[fleet lintel]

do we have gemini pro ga candidate on LMArena?

What is the ranking according to this group?

claude 4.6 > gpt 5.3 > gemini pro ga
OR
claude 4.6 > gemini pro ga > gpt 5.3

i believe these are the only two possibilities 🙂

[timber iris]

https://discord.com/channels/1340554757349179412/1469259617661091964

[glass perch]

Why is opus4.6 thinking forever man

I swear it never stops thinking

[icy yew]

At most for me it takes the maximum 30 seconds

For hard stuff

[glass perch]

Im tryna get it to make a story game

That lasts like 10 mins

Which might explain it

[north obsidian]

My ranking Claude > Gemini > Chatgpt

[north obsidian]

Code?

[plucky basalt]

dude why cant i test any model on this website

it reasons for 5 mins and it breaks

[burnt sinew]

What's ga

[icy yew]

General availability or something

[sturdy mica]

wow you guys havent touched gpt 5.3 codex

[obsidian cargo]

I'd definitely rank claude 4.6 over gemini 3 at this point, esp with gemini 3 being so terse

[celest orchid]

lol

[mystic olive]

Gemini 3 is still better overall tho

[icy yew]

The main problem is it hallucinates like a lot

[obsidian cargo]

idk my big problem with gemini 3 is it doesn't do long outputs. also it always wants to name characters Elara Vance

[cosmic falcon]

Can u guys release a premium version of the arena , so theres dedicated support , most of the time server crashes on my experience , just an opinion btw

[icy yew]

Wdym server crashes

Does the website not work or the ais

[acoustic garden]

What's causing this error? I haven't used it at all today, I don't have a limit.

[icy yew]

With Claude?

[acoustic garden]

yes

[icy yew]

Oh no that's normal

Claude code with opus 4.6 is unstable

[acoustic garden]

I use 4.5

[icy yew]

Well
Idk

🫀

[acoustic garden]

(

[sturdy mica]

Eli Vance half life

[toxic verge]

Gemini sucks. All though it’s all I use right now only because of nano banana

Well not sucks but has issues

[frosty shuttle]

It's always the same, Google nerfs models after 5 days of release, the same thing happened with Gemini 2.5.

[junior spoke]

Nano banana lagging rn right

[burnt sinew]

Ah so would ai studio one count

[burnt sinew]

For chatting sure maybe. For coding? Not in the slightest compared to opus 4.6 only thing going for it now is that its free on cli

[pulsar crystal]

i love max
why does max not tell me what model responded?
it would be useful to know

i guess max does does not know it?
because athropic is internally routing?

[echo aurora]

Going to respond in #ask-here in a bit

[echo aurora]

Are you able to create a post in #1343291835845578853 and explain a bit more what's going on? What modalities are you using, is this mobile or desktop issue, is there an error message, etc. Anything you think is relevant would be helpful to know.

[echo aurora]

It is going to be a bit unstable, but it is working. I'd recommend trying the steps in this article when you run into this error message.

[zealous sparrow]

Python in codearena?

[dense pumice]

why lmarena isn't opening

[burnt pulsar]

Opus-4.6-Thinking is too unstable for me with longer tasks, 4.5-Thinking-32K is way better.

[toxic verge]

Ya gem 3 is sad to see what it became

[icy yew]

Google says It makes these insane codes and when you actually try it it ain't the same level as they showed

[burnt sinew]

Yes for me it times out. I have been copying its thinking context and pasting it back to it so it can have the very long thinking times. Otherwise it will just keep trying and failing

[proud bobcat]

babe wake up

gpt 5.3

[echo aurora]

This has been flagged to the team btw cc @burnt sinew

[burnt sinew]

Tbh I thought it was an issue with every model... even gemini 3 officially errors out at around 10 minutes

[burnt pulsar]

Thanks, I usually analyze/optimize Mesa/Linux Kernel files of around 2000 lines of code. Opus 4.6-Thinking really struggles there.

[burnt sinew]

Wouldn't you use 4.6 non thinking then?

Or... what i said earlier with copying thinking context manually before it errors

[burnt pulsar]

It usually errors out within the thinking process already. But sometimes it finishes thinking but then only gets not that far with the answer.

[wind ember]

thats glm 5

[proud bobcat]

....

every model that has alpha in its name

was a cloaked

openai model

oh my god the joke flew over my head

im such an idiot

[burnt sinew]

If you copy the thinking it'll pick up from there

[burnt sinew]

Does the server in your tag really have free ai usage like lmarena?

[burnt pulsar]

Yeah, that was a workaround that I've used so far.

[burnt pulsar]

Well, yupp gives out credits and each prompt costs credits, but you get credity by reviewieng the output (with some gamification, so sometimes you lose some, sometimes you get much more credits per review).

Payout to EU has been suspended though, hence I didn't make any money there.

[burnt sinew]

Like they don't give credits to eu?

[burnt pulsar]

You still get credits, but you cannot cash out via Paypal to EU at the moment. But I am more in there for science and the access to the latest models.

[burnt sinew]

Cash out what??

You can make money from there?

[burnt pulsar]

Credits = Money -> Cashing out your earned credits.

Yeah, 1000 Credits are 0,90 EUR at the moment.

[burnt sinew]

How much free credits do they give you

Did leaderboards just update?

No announcement yet

[burnt pulsar]

There is a base of free credits at the start. But you need to earn more credits to pay for each chat. The model costs vary. The earnings vary a lot, too.

[burnt sinew]

Crazy

[burnt pulsar]

It somehow works, though.

[burnt sinew]

1502 THINKING MODEL to 1576 NON thinking

Aye there's the announcement

[burnt pulsar]

But as I wasn't able to cash out at the moment (and it might take many more months to resolve it), it is more interesting for people outside of the EU.

[burnt sinew]

@echo aurora What's the difference between code and text->coding

[mighty surge]

wich is better rn? Opus 4.6 or Codex 5.3?

[quartz pike]

opus absolutelly murdured the leaderboards lol

even tho it failed in my benchmark

[burnt sinew]

From what

[quartz pike]

to gemini 3 pro

[burnt sinew]

What is

I mean what is that from polymarket?

Doesn't look like it

Ah

[quartz pike]

making a portfolio website

[burnt sinew]

@echo aurora Also what was the point of not including thinking 4.6 and have just normal 4.6?

[echo aurora]

Code is going to be Code Arena leaderboard & Text->Coding is coding tasks done in Text modality, leaderboard here.

[echo aurora]

We're still gathering votes for thinking version.

[burnt sinew]

Shouldnt there be more votes on the thinking one? I would assume more people would use that one

[echo aurora]

Votes are generated via Battle, not Side by Side (where you can manually select a model)

[burnt sinew]

Are there currently plans to make code arena eventually include non pure front-end tasks?

[zealous sparrow]

@echo aurora I think 4.6 thinking has higher error rates

[burnt sinew]

It does he already forwarded it

[zealous sparrow]

ah gotcha

[echo aurora]

I wouldn't want to share more info about future plans until we're ready to, but overall our team is wanting to bring a lot more features to Code Arena

[quartz pike]

oooo

[burnt sinew]

Yeah thats fine

[limber panther]

yo

4.6 opus is really good, the only issue it has is no access to external textures and libraries

when coding

im really excited for claude 5 sonnet tho

its supposed to be huge and even better at coding tasks than opus

[stray aspen]

whats the rate limit of claude 4.6 think

[limber panther]

i think like 10 prompts

or 15

[stray aspen]

great

[limber panther]

someone made a farm game using 4.6 opus

its good

[stray aspen]

send it

[burnt sinew]

Thats your job

[limber panther]

u cant do that in lmarena

[burnt sinew]

Yeah you can provide it links to assets

[limber panther]

i meant if it had access to search and browser websites that would be really great

[limber panther]

does it work on the current models in arena?

if u provide links, it cannot open them or extract anything

[burnt sinew]

It works on every single model anywhere..

[stray aspen]

how do i send opus 4.6 think images

[burnt sinew]

Yeah but it can include them as assets??

[burnt sinew]

Nope

[limber panther]

so you send the links and say to put them in the assets?

@zealous sparrow

https://019c344b-e8de-7dd0-ae14-f03fa0c3273e.arena.site/

[burnt sinew]

You don't need to open a link to use assets in code

[burnt sinew]

Yeah direct link to the images

And say the dimensions of the image

[icy yew]

What's the rate limit for Claude 4.6 thinking

[topaz epoch]

i neeed help which is best opus 4.6 or 4.6 thinking for python coding?

[icy yew]

Thinking ofc

[burnt sinew]

Like I made flappy bird 1:1 clone using that it just took all asset links

[limber panther]

yea i realized, im dum lol

i just need to search for the links and ask it to put them as assets

[topaz epoch]

thank dude

[limber panther]

althought gemini 3 pro could put the assets itself, like planet textures without even asking for it

i think gemini 3 pro training data is pretty good

[burnt sinew]

Yep

It did that for flappy bird

But it used external asset links

[limber panther]

yeah, my point is that gemini 3 pro has alot of assets in its training data and can use them without the user noticing

[burnt sinew]

Yeah sure or just try asking opus to use its own asset links

[iron laurel]

What is the level of thinking for 4.6 Thinking?

< or > than 32K?

[honest verge]

16k I think

[iron laurel]

So it's better to use Thinking 32K

[honest verge]

They don't have it for now

[rancid turtle]

Hello

[honest verge]

It's only for opus 4.5

Not for 4.6

[iron laurel]

Yes that's what I meant

[echo aurora]

We don't have these rate limits publicly listed. Although this is something we're considering.

[rancid turtle]

is arena ai downloaded for ios or no?

[inner relic]

Did they fix the response bug

[icy yew]

It's a website

[echo aurora]

Nope, there isn't an app, but this may change one day.

[echo aurora]

Which bug are you referring to?

[inner relic]

"Something went wrong"

[rancid turtle]

Okay, thank you

[inner relic]

i am talking about the bug that cuts the response off.

[burnt sinew]

You mean after it thinks for a while?

[echo aurora]

I'd encourage you to scan #1343291835845578853 for a similar post and share/tag me there. Or if there isn't one that directly lines up with the problems you're having create a new post.

[icy yew]

Btw is Claude 4.6 dynamic thinking is the same on arena

[rigid holly]

Alright i tried the 4.6 model in writing stories

Dont know what to think honestly

Like the writing is not BAD

But it feels drier in dialogue for one than previous models

Can't speak about code or other such things. I Dont use ai models for such things as code or image generation

[inner relic]

I am using claude opus non thinking and this happens..

[rigid holly]

Oh yeah that happened to me to in writing. I was using thinking tho. 4k words work i think cuz i also had it work on shorter chapters, but 8k words get crashed

[honest verge]

I think opus 4.6 has some limits

It says every time "I have to make it in my limit of 20 steps"

When it thinks

[rigid holly]

Tf does 20 steps even mean

[icy yew]

What's 20 steps

[rigid holly]

I will say this tho. The restrictions are more loose in what it rejects from writing than the previous model

[latent merlin]

Hello I am new here

[honest verge]

Idk

But it says 20 steps

[icy yew]

Soo the thinking is set to high could be the cause of the thinking limit making it get a error

[topaz epoch]

Bro i was using 4.6 thinking and it keep getting stuck in the middle because of thinking

[icy yew]

Idk

[topaz epoch]

Exactly

[prisma cipher]

The lmarena team needs to increase the context limit in all responses for models of this type; this way, that frequent "error" is avoided. It even avoids giving the model instructions for response limitations, while also saving time and writing.

The typical limit is approximately 8350 words per answer in Claude, but lmarena has to increase the limit until everything is completely finished and not limited.

[honest verge]

You can do it without this but it will require splitting prompts

[prisma cipher]

Mine requires a detailed and complete answer; therefore, it is necessary to apply these restrictions, instead of putting everything in a single response.

[honest verge]

Finally opus 4.6 thinking actually think for some time not just for 1 second

[echo aurora]

This has been flagged to the team btw.

[icy yew]

This is probably the core issue regarding the errors when using Claude 4.6

Hope it gets fix soon

[gleaming roost]

[honest verge]

Idk it doesn't happen to me now

[gleaming roost]

Perhaps it's just a matter of luck

[icy yew]

It's fixed for me

[gleaming roost]

[granite tide]

opus opus let me use opus

[honest verge]

Please arena I need opus 4.6 thinking 32k

My opus 4.6 thinking is kinda homeless

I live with my opus 4.5 32k

[honest verge]

Maybe it can't make too much code or text

Because of the limit

[prisma cipher]

Thank you for the notification.

[uneven peak]

@echo aurora how old are you?

[icy yew]

Asking a manger his age

https://cdn.discordapp.com/attachments/1276881295447691330/1354443134377005128/attachment-7.gif

[uneven peak]

Ayo chill fam 😭

[limber panther]

@echo aurora why do i get an error that corrupts the whole project, when I use the coding arena?

i tried this in many chats, and 50% of the chats get corrupted at the end of coding when publishing the app

[prisma cipher]

Read the following:

#general message

[limber panther]

so its the context limit destroying the whole project when its done coding?

[prisma cipher]

It is the limit of the context that prevents the work from being finished, that is why the error appears.

[prisma cipher]

You can limit it to 8350 words per response and see if it actually finishes the job. You can try it.

[limber panther]

but my specific prompt was to produce long and complete code...

seems like arena.ai wont allow that

[prisma cipher]

Use the model in text mode and generate the code to copy it later. Use reasoning mode.

[limber panther]

code arena is the only affected part of the context limit?

[golden ocean]

i'm a new soul, i came to this strange world, hoping i could learn a bit about how to give and take but since i came here felt the joy and the fear finding myself making every possible mistake

[prisma cipher]

No. Text mode is included.

[limber panther]

yeah it does cut mid coding

but it doesnt corrupt the whole chat

like code arena...

[prisma cipher]

It includes instructions for dividing the answer into parts, with each part having a maximum of 8350 words. It's useful.

The other thing is to write the code directly, clean, without comments, without artificial simplification, and completely unified. It's a very powerful instruction.

[red meadow]

why am i always getting an error when claude 4.6 gets done with its task in code mode?

[echo aurora]

We are looking into these reported problems, but it's worth trying these steps in the meantime as they may help: https://help.arena.ai/articles/1645798556-lmarena-how-to-something-went-wrong-with-this-response-error-message

[icy yew]

Now this is impressive
W Claude
https://vt.tiktok.com/ZSaw8kyst/

[hollow snow]

where is this opus 4.6 think in the leaderboard

[prisma cipher]

Include these instructions at the end of your prompt:

Each response must be consistent with all of the above and without deviations, proactively correcting anything without waiting for explicit instructions from the user.```

This instruction is very powerful, especially when it is something serious and in production mode, but useful for testing the model's capabilities.

[toxic verge]

Time it out space out requests

[prisma cipher]

Good luck.

[limber panther]

thanks for helping man

[stray aspen]

claude4.6 gave me a pretty nice roblocks camera system

[limber panther]

can it code an nice GUI

for roadblocks

💀

[stray aspen]

yes

its actually cooking lol

and its not even the thinking version

[limber panther]

show me the image, cuz i wanna see

[limber panther]

i used 4.6 opus in html games today, it cooks good

[viral cedar]

how is claude 4.6 opus like 5x better than gemini 3 pro

[limber panther]

i made a good solar system simulator with assets

[limber panther]

dw, sonnet 5 is gonna be even better

sonnet 5 is the master of coding

[prisma cipher]

Do you have online samples so I can see how it actually works?

[limber panther]

its not complete yet, but i can give u the arena.ai test link

cuz it has a minor bug in loading Earth's texture

https://019c34b5-2776-7bb5-b2a6-4d6ba40b36d1.arena.site/

[viral cedar]

google lowkey kinda disappoting me rn

they gotta catch up

[limber panther]

google is cooking, we just wait

[viral cedar]

like idk how benchmarks show that claude 4.6 opus is like 3% better or sum

but its even evident w/ webdev and documentation making

i gave it requirements saying make me documentation for so and so programming lang, and it completely half asses it and ignores half of my instructions.

[limber panther]

4.6 opus is way better than 4.5 opus, if you test them both you gonna see a big difference

[viral cedar]

meanwhile claude 4.6 opus basically turns into slave and acts like its being held at gunpoint

[limber panther]

lol

[viral cedar]

or even a simple half-assed prompt saying make me UI like palantir

gemini and 4.5 opus will just half-ass it as usual

4.6 opus will immiedately cook up and make u whole UI lib that actually looks decent and is bug free for most part

[prisma cipher]

It's a start. What I like is space warfare and first-person perspective.

I have in mind that Opus 4.6 will help me create a unique and realistic universe to integrate my character into, but I will do that at some point if possible.

Console, PC, and mobile games are linear and feature repetitive stories. My universe will be more than that.

It's just for playing around for a while, not for getting addicted.

[toxic verge]

All ai does this . This is the way they’re optimized.

[atomic lagoon]

Because you need to go more in depth for it. Also tell it to not be basic or be lazy whatsoever or it will be by default. AI does what’s fastest not what’s best quality

[limber panther]

4.6 is significantly better than 4.5 once you test it yourself

[stray aspen]

we need image uploads for opus 4.6

[north obsidian]

Up

[echo aurora]

Big +1

This is being worked on

[steep jewel]

this is terrible

[steep jewel]

i could pitch you like 100 much better ways to simulate a plent's atmosphere. i made procedural textures in blender in like 5 minutes that look 100x better than this. not to mention the shadows just dont work

[limber panther]

its non thinking, and my prompt was pretty simple

[steep jewel]

yeah fair enough

[limber panther]

yeah but, you're not my AI assistant who responds less than 5 seconds..

lol

[steep jewel]

yes, and what i make also isnt a steaming pile of dogshit

[limber panther]

yeah ik devs are better than AI at coding, but this is impressive for someone who doesn't know a bit about coding languages

[steep jewel]

nobody would choose this over something made by a human. its only value comes from being extremely easy to have an ai make for you and free, which is not negligible but necessary to understand you're pitching quantity > quality

[limber panther]

sonnet 5 is also reported to be much better than this 4.6 opus

[steep jewel]

at coding? no lol

[stray aspen]

how do i use sonnet 5

[limber panther]

which is said to be released late feb early march

[steep jewel]

you cant yet, its not released

[limber panther]

at 83% SWE bench, its way better

[steep jewel]

oh ok i didnt see

did they put out the numbers

weird they'd make a sonnet model super good at coding when coding is quality > quantity

[limber panther]

from leaks, it seems to be close to that number. no one knows anythin yet

[steep jewel]

and opus is supposed to be good at complex, structured tasks

[steep jewel]

leaks are usually pretty terrible sources of information. this goes for anything

[limber panther]

yup, there are no actual sources on it yet

[steep jewel]

literally any time one of the big ai companies does anything now you have 60 wojaks on twitter saying its agi superintelligence from the preview builds they've sent out

[limber panther]

when 4.6 opus launched, it was pretty decent not too impressive

[steep jewel]

yeah i tried it. its pretty good

[limber panther]

people expected sonnet 5 with better coding and stuff, but it got delayed

also sonnet is pretty cheap at $3 per input / $15 per output compared to opus

[steep jewel]

i've been thinking about a system of fine tuning over the top of the base model where you have a few elo based examples the ai is trained to respond like to specific criteria. essentially what is already done with safety but for code

i also believe you can create a "perceived prompt" that the ai sees and the stupid half-thought-out prompt given by the human. you have an intermediary ai that goes in and edits the prompt so its good and leaves little to the stochastic imagination

nano banana already does this, as well as hunyuan, qwen, and most other ai companies

[proud bobcat]

K2.5 instant is really strong

Damn

Thinking mogs it but it’s nice to see

[prisma cipher]

What is known about the context? Is it 1 million or more?

[modest prism]

Please help how do I fix opus 4.6 thinking timeout error

[verbal nimbus]

What's surprising is how well it scores on long context

[toxic verge]

I think that’s the key here

It could also be argued and a case can be made that perhaps is actually occurring isn’t necessarily an improved model as much as it could be improved memory and hardware on their end

[verbal nimbus]

It's the only open source model that's competitive on long context reading comprehension:

Gemini 3 Flash's score is insane, but Opus 4.6 scores higher in MRCR needle-in-a-haystack. Opus is still not on the above benchmark yet though.

[thorny drum]

was opus 4.6 an anon model first

[toxic verge]

Gemini is fraud

Probably has the worst memory issues of all the models

That’s how I feel when I use Gemini

[verbal nimbus]

On the app, yes, I think they do prune parts of the context. You can actually test this by asking the fast model to output a transcript of the entire chat so far. If you closely, it actually leaves parts out (not entirely sure if it's just the model, but probably not).

[toxic verge]

I don’t even bother for one reason only I don’t code and I don’t see the reason for long text because you’re still dealt with the problem of the models all hedging hard

[verbal nimbus]

I guess I can test it more but it's a bit time consuming to recreate a long convo.

[toxic verge]

It’s like musical chairs

They alter the words and meanings of the semantics and hedging is one of the most messed up things about AI in my opinion

Grant more authority to model than it does to users intent

Here’s an example

[proud bobcat]

Another day another Chinese banger

[toxic verge]

You see how it alters the words now imagine with a long context

It completely stripped away the emotion, the individuality, the uniqueness of expression from my statement into

Look kimi instant

Va thinking

[stray tusk]

Hi

[molten robin]

i hate this endless generating bug so much.

[main nexus]

ong 😭

[molten robin]

it slows down my LUA experiments SO MUCH

[verbal nimbus]

Wow, what a prompt 🤣

[rugged abyss]

What model is beluga? Is this an alias or do I just not know that model?

[hazy forge]

show the output we may be able to tell

[rugged abyss]

I got it again, its from Amazon
https://codepen.io/Emilio-the-encoder/pen/raLZKRo

[hazy forge]

seem to be actually pretty good

[shrewd citrus]

wait when did bezoz step down from ceo 💀

[green yacht]

5 years ago ish

[echo aurora]

July 2021 😏

[shrewd citrus]

so for the past 5 years I’ve been thinking that Jeff was still the ceo 😭

[balmy mist]

which company is pony alpha??

[frosty lava]

where do you found it ?

[stray aspen]

<@&1349916362595635286>

[glacial dock]

How the quack is everyone doing

[hardy lion]

pretty good, how about you Ducky?

[copper cape]

is here anybody looking for the developer?

[glacial dock]

Just working to pay that duck support

[spare rune]

Lua or luau

[molten robin]

glua

[spare rune]

Oh

[molten robin]

I do GMod lua ai experiments

[spare rune]

Ok

Omg broo

Why is ro * lox a banned world

I’m gonna die

[strange sluice]

because life isnt always roadblocks

[toxic verge]

Cuz of scammers

[spare rune]

Elon Musk

See

It works

free

Money

Free money

bitcoin

Btw

[old garden]

https://tenor.com/view/grilled-chicken-grilled-chicken-eat-grilled-chicken-eat-gif-9823209876167517923

[strange sluice]

hi

[toxic verge]

Who is that

[fiery gull]

Let's say a guy who liked children a lot

[toxic verge]

Like cheese pizza?

[austere sundial]

OMG Lmarena stopped Someone is work with some function that I probably won't use

[sturdy mica]

diddy baszucki

hell no

[old garden]

david baszuki?

[old garden]

what about it

[old garden]

https://tenor.com/view/i'm-new-here-say-hi-discord-im-new-here-say-hi-im-new-here-say-hi-discord-new-member-discord-new-gif-8065649195492145845

[sturdy mica]

websim is cancer

dude you have so much stuff

[old garden]

ik websim is lowk going bankrupt or somethingf

i have 1.2k folowers on ther i think

[sturdy mica]

i joined this server like months ago i just rejoined after leaving for a while you can search my name

[old garden]

ok

[sturdy mica]

holy

your websim page is full of slop

nobody is playing this bro 🙏

[old garden]

ik

iwas just testing

if the ai knew how to make btools system

i never released that game to the public

many of my projects are unrelased

like 99% of them

[sturdy mica]

yes you did

i was playing it

all your private games are public

theres a lot

[old garden]

https://websim.com/@Trey6383/sonic-2d/50 i mean ive been working on this

lately

and this https://websim.com/@Trey6383/earthbound/83

[sturdy mica]

its too fast

[old garden]

ik

[sturdy mica]

if you stay still you glide lol

wow this game is awesome

[old garden]

ive just been so caught up with my more important projects that

i havent had time to work on

the quality ones

[old garden]

you can see though that i was working oj changing the speed, the sliders of speed, and accel rates

[sturdy mica]

cool

what game is that

[old garden]

earthbound

https://en.wikipedia.org/wiki/EarthBound

[sturdy mica]

yeah i mean on websim

[old garden]

o

[old garden]

this ^

the intro is bad rn

i havent had time to fix it

but recently a lot of websim staff have been fired
free credits have been removed
some of the other popular users are just quitting

[sturdy mica]

websim has always been slop

lol

[old garden]

i personally dont agree with that statement
it was so good
when free users got 50 free gens a day
and the team gave out free max subscriptions (i was one of the first to get one)

[sturdy mica]

no im talking about the cheap AI generated games on the front page

nothing was ever fun on that platform

[old garden]

thats not really websims fault