general

[strong ferry]

They won't. At least not for a year. They're stubborn.

[outer flicker]

did opus really comfort you guy?

[hollow ivy]

i did some roleplay with it, but that was long ago)

comforting? not really, but entertaining and intelligent

[strong ferry]

It was the only model that could go into long depth and detail in summarizing and storytelling. Sonnet is garbage.

[hollow ivy]

Opus-4.6 was one of the very few models, which had an actual (long-term) understanding

[outer flicker]

just like blade runner 2049 k lost joi, we lost opus 😄

[hollow ivy]

at least, we still can use Sonnet-4.6 for free
Let's be thankful for that!

and also GPT-5.3, 5.2 and 5.1

and Grok-4.20

and Gemini-3-flash

and Gemini-3.1-flash-lite

[outer flicker]

yeah

[hollow ivy]

sonnet isn't bad

and GLM-5.1 and Qwen-3.6

[strong ferry]

Sonnet 4.6 is only available under web search, not text though

[outer flicker]

gork and gpt know comfort us is just officer K

[hollow ivy]

oh

but in battle, it is available

grok not gork lol

[outer flicker]

why they don"t add onther searching model?

[fast dew]

?

It's available there too

[limber dock]

so it wasn't just me then haha. Opus is honestly amazing though, and if it were just a tiny bit cheaper, I totally would've gotten it 🥹🥹

[strong ferry]

Oh, my bad. I added thinking and it only gave me 4.5. Let me test it.

[hollow ivy]

and there's also Minimax, Kimi and Mimo

[strong ferry]

For real. It was the only text model I used. Gemini Pro used to be good until they nerfed it sadly.

[hollow ivy]

what about Gemini-2.5-pro ? that is still available

i used it last year for RPGs, was quite good

(if prompted well)

[strong ferry]

That's what I mean. The Geminis feel nerfed compared to the initial launch. Back then, I could ask for an entire chapter be generated in 5000 words. Now, I'll be lucky if it gets up to 2500.

[hollow ivy]

even with advanced prompting technique?

using special syntax, etc

have you guys tried out new Qwen (3.6) and new GLM (5.1) for roleplaying?

[sand loom]

Anyone testet sonnet 4.6 vs Qwen 3.6 Plus in coding? Can it compete?

[rustic gale]

Tbf, gemini was bad with long outputs even when 2.5, at least when it came to creative writing. Which is honestly defendable from the devs' POV, but it is what it is

[strong ferry]

Just saying. If I asked Opus for an 8000 word chapter, it delivered.

[limber dock]

I feel exactly the same. Like, Gemini is definitely getting more advanced, but the 'vibe' just isn't the same anymore... Sorry, I don’t really know how to put it into words🤷🏻‍♀️

[rustic gale]

This. Maybe not 8k, 8k is kind of overboard, but 4k is achievable

[hollow ivy]

but what if you set the temperature above 1.0 for 2.5pro (in ai studio) ?

for roleplaying/creative writing

[rustic gale]

(also perhaps the reason why it costed astronomical values for arena. 😄 Imagine how much it costs in API; every single output is like plus 10k tokens)

[strong ferry]

Nah, I get you. Opus actually gave off that emotional feel rather than technical. You could feel like a human wrote it. While with the others, you can clearly tell it's just a machine talking.

[hollow ivy]

even with newest grok?

[rustic gale]

It will do what raising the temp will do. Output length is something hardcoded into the model; even JB people can't really overcome it

[strong ferry]

I don't bother with Grok. Again, limited in word count as well.

[hollow ivy]

what about Mistral?

(for RPGs/sandbox & creative writing)

[normal abyss]

grok 4.2 is pretty underrated

[hollow ivy]

and Command (by Cohere) ?

and Seed (by ByteDance) ?

[strong ferry]

Not sure how it delivers, but I don't usually do creative writing these days. I ask for long-form essays based on movies and stories well known. Any other model besides the latest Gemini pro and Opus always hallucinates the stories and end up being like 90% false. It gives me a headache.

[fast dew]

Lol is the 4o situation happening again

[limber dock]

Yes, exactly! That's exactly what I meant. And now I'm starting to miss Opus..😦

[rustic gale]

Command is a coding model last time I checked. And both are kind of third echelon. Because tiny. Which isn't inherently bad, there's just the league of models you can run on a 4090 and the league of those you can't

[hollow ivy]

with poe.com you have 1 message per day (better than nothing, i guess)

[strong ferry]

Guess the old saying remains true. "Nothing good lasts forever."

[rustic gale]

4o was a machine which was unhingedly trying to do stuff in the 'sound human' department. It was so-bad-it's-good category. It won over so many hearts because at the time it was the only one who could do that. Opus does this without the excesses

[fast dew]

Difference is opus is actually a good model

I don't get all the emotional stuff but i agree with that

[hollow ivy]

except life itself, which lasts forever

"Life finds a way." - Dr. Ian Malcolm (Jurassic Park)

[rustic gale]

So was -4o, again, at the time of its release

[fast dew]

Mehhh

For its time probably

Yh

I remembered the first time i got 4o

[rustic gale]

There was a time people were developing stuff with -13b hosted locally. And don't you all forget this.

[fast dew]

I cant lie though I didn't like it

[strong ferry]

Opus also had a way of deeply connecting plot threads too. You could go on for a while and it will try to avoid being repetitive too. Man, this sucks.

[fast dew]

Remember when GPT-4 only had a 4k token window

And we thought 32k was too much

[hollow ivy]

so, which models comes closest?

Sonnet-4.6-thinking?

does it have a thinking-variant?

[light siren]

honestly they dont have a choice

[strong ferry]

Hold on, testing Sonnet now. Wish we had thinking but will have to test the normal one first

[mild spade]

Are you using opus for creative writing too?

[strong ferry]

Yep!

[fast dew]

Lol from what i've seen like 95% of the community r using opus for creative writing

[limber dock]

That was actually super helpful, man!! Thanks a lot!

[mild spade]

cool bruhh I was able to produce like different fanfics 100k+ words now

[strong ferry]

Oh snap! So Sonnet 4.6 gave me a 15K word response 😲 The ONLY problem is the slight lack of emotion. So the best way to compare the two is

Opus - The Author
Sonnet - The Assistant

[mild spade]

coz it's good at that

[fast dew]

Cant u tell its like AI tho

[mild spade]

Nopeeee my readers didn't notice

[fast dew]

Lol

Where u posting them

[mild spade]

Ao3

[fast dew]

Wattpad?

oh

[mild spade]

people are amazed

opus is just really natural

[mild spade]

Noticed the lack in emotion too TT

[hollow ivy]

there's also anakin.ai

(but no opus there, but other models, which might be interesting)

[mild spade]

I miss opus

[fast dew]

Lowkey i dont think its coming back

U probs gonna have to wait for a new model

[mild spade]

I just trust

[strong ferry]

Guess it's better than nothing for now I suppose. I'll use Sonnet until they hopefully give us Opus back.

[limber dock]

honestly, I saw this coming too, but now that it's actually here, we're all just kinda bummed about it, right?

[mild spade]

Atleast we get something

[fast dew]

Sonnet?

[mild spade]

The new model, if ever

[fast dew]

O

[mild spade]

😁😁😁

Trust

[hollow ivy]

and gab.ai, which gives you ¢150 when registering, but after they are used up, it costs you money

(so maybe 1 message to opus-4.6 there)

[fast dew]

U know they might be uhh

[strong ferry]

Pretty much. Happens with everything honestly. It's fun while it lasts but eventually it will run its course or simply be taken from us.

[fast dew]

Releasing GPT-5.5 next week

[small swan]

Is there a way I can create videos using Seedance for free for testing purpose ?

[hollow ivy]

it seems Google-Deepmind has a model-release soon®©

[outer flicker]

why you edited so much lol

[hollow ivy]

lol :)

because my fingers suck ^^

(and the keyboard)

[outer flicker]

still im use lm for chatting help task, make code, and roleplay

[hollow ivy]

which model do you now use mostly, in arena.ai?

[outer flicker]

my finger too

[hollow ivy]

maybe we need to eat more micro-nutrients?

(broccoli ^^, actually yumm if spiced correctly)

[outer flicker]

just say im like all cause im in neutral side 😄

[hollow ivy]

have you tried out sonnet-4.6 in coding?

is it decent?

[outer flicker]

everytime im try is pop up 'something went wrong'

[hollow ivy]

in Python it should still rock

as that is the easiest language for AI

[outer flicker]

but im make driffent ai work together to make a appliction all time zone of all country

in coding mode

[hollow ivy]

second-easiest might by Java or javaScript/TypeScript

java is good, if python is too slow

but C# is faster (and of course, Rust/C/++/Go also are faster)

also Nim and Zig (both are faster than java)

of all really fast languages, C might be the best for AI

(Go is slightly slower than C, but is good for multi-threaded apps)

[outer flicker]

are you try coding in arena is good?

[hollow ivy]

i only used text-chat mode

and compiled/tested it myself

with Opus-4.6-thinking, that's a breeze, of course

[outer flicker]

yo im like coding but each model do driffent task

[hollow ivy]

have you ever coded with GPT-5.4-high?

is it decent?

[outer flicker]

is qwen good at chatting? they say the model is 1m token

[outer flicker]

is only good sometime

[hollow ivy]

Qwen is not as good as Gemini-3.1-pro, but probably better than g2.5-flash

i wonder, which chinese model is the best for

• coding & debugging
• creative writing & world-building
• GMing (intelligent gamemaster-ing)

[outer flicker]

oppen ai the first ai oppen AI era

[hollow ivy]

lol?

closedAI now

[outer flicker]

yeah before

cause more good onther model now push gpt to down

im found grok imagine not free anymore 🙁

[hollow ivy]

grok-imager is gone?

i wonder, what "spud" can do..

[outer flicker]

yeah go check it

[hollow ivy]

the new OpenAI model

[outer flicker]

can make some not strict too much like onther ai

[hollow ivy]

i have the feeling as if this month another model-update will come out…

yeah, Grok is not as censored in imaging

and also in roleplaying

[outer flicker]

even grok have ani

[hollow ivy]

btw there's also Grok.com and openrouter.ai (and perplexity.ai)

[outer flicker]

grok roleplay good?

[hollow ivy]

it is not as strict, so could be fun for certain people

it allows more freedom in your RPGs

but not as intelligent as Gemini-3.1-pro

GLM might also allow more

(And Mistral)

[outer flicker]

im will try

[radiant pewter]

My chat of me direct message with claude kept loading infinitely does not give an answer, and it does not help to refresh the page or log out, someone else with the same problem?

[modern wedge]

So when they will bring back the Opus? (I writing goddam peak fiction but now it has too pause)

[hollow ivy]

Opus-4.6? maybe in 2027

but then, there will be better models

[modern wedge]

Fr???

Hell nah

[outer flicker]

anthropic leak their new model name capybara

yeahhahah

[molten cargo]

how did you all even use opus in arena? I used to hit the rate limit after 4-5 responses, surely that's not enough for writing stories with it or whatever the use case was

[hollow ivy]

i wonder if Capybara/Mythos-5 really were april fool's jokes (according to some people, they were)

[radiant pewter]

I'm just using sonnet, does anyone else have the same problem?? Or recommend a solution

[hollow ivy]

maybe they were patient

[molten cargo]

maybe, but it was like a 1 hour timeout after 4-5 responses. I could never be that patient

[rustic gale]

If you revisit them over the course of weeks, you can hit the context limit. I know this because I did it several times, though I suppose it was opus-4.5 back then, end of last year 😄

[hollow ivy]

that's 1 message every ~12 minutes, which could work, if those messages were really long and the player was really patient (lol)

[rustic gale]

And yeah, we're not talking about chatting about life. Creative writing can easily mean something like 'I send a prompt of 2k tokens and would like to have it expanded by you'

[sour spear]

Opus was a pain to use because of that. It's understandable they removed it, because it's insanely expensive. If you pay for API, they can easily charge you 1-2$ for a single reply. Fantastic quality, but their pricing is just insane.

I paid 10$ extra on top of my subscription, and got 6 replies out of the model

[hollow ivy]

yeah, we need much cheaper inference

there's a new chip for that

an ASIC-based one, iirc

[sour spear]

They're starting to use Google chips, that could help, too

[whole knot]

when is opus coming back ?

[hollow ivy]

but Cerebras also had a good approach

[hollow ivy]

maybe next year (2027), maybe never

depends on inference chips available worldwide

https://www.heise.de/en/news/AI-inference-cast-in-silicon-Taalas-announces-HC1-chip-11185112.html

https://blog.themenonlab.com/blog/ai-inference-accelerators-compared

[native flame]

Hi guys , since opus is no longer available, do you know another site where one can use it??

I used to use it for japanese language learning:'v
It was the better of the three

[rustic gale]

Tbf, for these kind of purposes you probably don't need that much advanced models, so whatever other two were in your shortlist, just switch to them

[heavy blaze]

so quick question, if i have an already made website, how do i upload it to arena so i can use the models there to edit it.

[native flame]

mmm I tested with the other two, and they gave wrong answers:'v, that's why Opus was the more trustful and useful

[heavy blaze]

where did you get the arena champion role?

[rustic gale]

Well, without big names there's... dozens more on the arena? Gotta catch them all. Then again, maybe you just need an actual course; it's not like the task was insormountable, or particularly hard, before LLM

[native flame]

???
Sorry, what do u mean?

[heavy blaze]

the yellow role, where did you get it

[vivid coral]

for those asking why opus is gone, just look at twitter and the controversy around Anthropic right now. They are heavily limiting the $200 users now. They are PISSED. They screwed everybody that paid big for it. It's the classic bait-switch we see from some of these bad actors. Promise to be the only "safe AI", "moral AI", "pro-Black", "pro-LGBTQ", "anti-capitalism", blah blah blah.....Then once they bait you in, they switch up on you, and now only very rich and elite "nepo-babys" can afford and use it, who are 99.99% white and straight. The same kind of "bait and switch" many dictators have used to gain power. They ran the playbook perfectly, step by step. And now, you are seeing the result.

[heavy blaze]

so basacly Anthropic fooled us all

[native flame]

Ahh tbh I joined the server way longer before, sadly I don't remember

[heavy blaze]

damn

[vivid coral]

let me clarify, Arena didn't or never would remove something over differences with a company, it's simply that they can't afford it. The switch up Claude just pulled has made it impossible for normal casual users. Even people "well off" are having to switch to something else. That's what it comes down to. A sad, yet simple story. And a lesson for us all.

[robust jay]

Hi

If anyone can test it, please rate it and tag me with @robust jay

[heady sky]

is it virus??

[civic plaza]

I really hope they add opus back. The other AI’s are okay, but they are getting on my nerves.

[robust jay]

No

It's not a virus.

[signal pelican]

oh sure, if it was, why would the person ever admit it anyway?

[inner relic]

e

[robust jay]

It's Google AI Studio, but I made the app for mobile myself.

[rapid merlin]

boy what does this have to do with inclusion 💔

[signal pelican]

right... you can't just paste the file here

[rapid merlin]

the usage limits on Claude are absolutely ridiculous tho, yeah

I only got the 20 dollar plan but bro

[heady sky]

Claude is bs, it got leaked and within 1-2 weeks all chinese app will be same level

[rapid merlin]

I got usage capped within TWO MESSAGES

and the second one didnt even finish

[robust jay]

If you want, I can send you a video of me installing the same file and using the app to prove it's not a virus.

[signal pelican]

no thanks

[robust jay]

Use an antivirus program so you can see.

[signal pelican]

no

[heady sky]

@signal pelican stop being a menace wtf

[signal pelican]

you're free to test that for him then

[heady sky]

so u made an app for google ai studio? why?

[signal pelican]

and there it is...

[vivid coral]

Google AI Studio's limits are pretty bad now too, better off just using it here

[robust jay]

No, it's basically Google AI Studio but as a mobile app, and simply because I had nothing to do and also because I don't have a PC to use it on.

[icy frost]

spending money for ai

[heady sky]

wdym? i pay for ultra version

[rapid merlin]

if it's fun why not

tho the 20 dollar plan would be useless if it was really needed for like anything practical

[heady sky]

not about fun, i use it for work and make a lot less errors due to if way more efficiently, easily worth $150 a month

[rapid merlin]

With what usage?

[robust jay]

Someone brave enough to test the app I made, please, and for the love of God, give it a rating and say it's not a virus.

[heady sky]

its not about brave, its useless app

[rapid merlin]

The quarter daily limits are absolutely horrid

[heady sky]

ai is too expensive. chatgpt is losing money on their high-tier subscribers

i use it in my work a lot

[rapid merlin]

you use gpt?

[heady sky]

nah it sucks

[rapid merlin]

I was talking about claude

[heady sky]

i use gemini

[slow axle]

Is video arena working?

[robust jay]

😭

[heady sky]

bro i got warning cuz i asked same q here

[slow axle]

I think its broken

[robust jay]

Out of sheer spite, I'm going to make an LMArena app for mobile phones.

[slow axle]

Im not asking for that, but in the website I can't create videos

[fast dew]

It's only 20 bucks jit

[slow axle]

I think its a bug

[storm dust]

well i think mods are just fed up so they just added that 🤣

why did bro remove the screenshot

[proper quiver]

i finally vibe coded a roblox system since sonnet 4.6 worked

[storm dust]

"working hard"

i didnt even had to watch the video

i worked smarter

i understood everything from the name

https://tenor.com/view/thinking-think-think-of-it-gif-12729650

[rapid merlin]

is gpt image 2 still in battle mode btw?

[severe canopy]

realease gpt2 right now oks !

[dusty wagon]

Wait so why was opus removed again? I first saw it was about due to financial issues, and then i saw another reason

[pastel ember]

Looks like the guys from Grok paid to have other top-tier competitors like Opus removed. 🤣

[whole sundial]

haven't heard or seen anything about them being back so no

[crimson ivy]

When are the opus models going to come back?

Did they remove them because of the error issues?

[sharp marsh]

Dude

Opus 4.6 got removed

?????

[analog owl]

Hi @echo aurora , how are you? We talked in private a few months ago IIRC, about the captcha of death. Just so you know, Im getting the captcha of death again. Hope everything is alright on your side. Also and of course, happy easter to you, your family and loved ones as well.
Thank you very much for arena.ai and everything you do to make our lives better and easier.

[vapid holly]

Yes its extremely dumb

[normal abyss]

welcome to the grok community 👋 (grok is kinda underrated though, i kinda like it)

[kindred sorrel]

I saw on YouTube that videos can be generated in a Discord server, but I can't find it. Can someone help me?

[normal abyss]

its within the arena.ai website now

[kindred sorrel]

Bro, I saw on YouTube that you can create videos directly from a Discord server.

[normal abyss]

it used to be like that a few months back

[kindred sorrel]

But

[normal abyss]

yeah thats from way way back

arena.ai also works on mobile btw

if your on that

[sharp marsh]

How limited is it on their website

Gemini on aistudio was nerfed a LOT and I suspect grok to be the same

[normal abyss]

for grok rate limits?

[sharp marsh]

I was able to do like hundreds of requests to gemini without ratelimiting, now it's limited to 5-10 a day

[sharp marsh]

Yes

[normal abyss]

wait where lol? hundreds???

[sharp marsh]

Yup

[surreal zephyr]

yes google ran out of computers

[normal abyss]

i think its around 20, i never hit rate limits so i never really find out (could be much more or much less)

[kindred sorrel]

I want to create videos on a website, and I’ve made some, but there are a lot of options missing. On many websites, there are options like 10 seconds, 5 seconds, 4K, 8K, etc., but I don’t get those. When I click on the video option, it only shows an option to select an image and nothing else.

[surreal zephyr]

gemini had 100 requests they now nerfed to 20

[sharp marsh]

I was able to spam 7 hours straight in aistudio coding, gemini 2.5 and 3.0
When 3.1 released all models ratelimits were nerfed

[sharp marsh]

10*

Also not 100, but 1000

[kindred sorrel]

Now what can I do? Please tell me.

[sharp marsh]

It got nerfed to 500, then 250, then 100, then 10

[sharp marsh]

Oh that's bad, I will keep using gemini 3.1 pro with multiple gmails lol

[rustic gale]

And next step is zero because why would there be free cookies

[normal abyss]

arena.ai is very good for making videos if you hit the right model (you do not get to pick models here though)

additionaly higgsfield is pretty solid

[surreal zephyr]

the web int has 100

[sharp marsh]

What?

[kindred sorrel]

Bro, how can I select a model when no models are even showing up?

[normal abyss]

you dont on arena.ai, its all randomised

[kindred sorrel]

I think I’m bothering you a lot, please forgive me.

[sharp marsh]

Anthropic really killed free users huh

Jesus christ

[normal abyss]

no its kk lol

[kindred sorrel]

🙃

[normal abyss]

do check out higgsfield though, it may be what your looking for, with the options you said before (mind the rate limits dont reset daily like arena.ai, im pretty sure its monthly)

[normal abyss]

are you wanting any specific video model?

[kindred sorrel]

Alright bro, can you tell me any free websites where I can generate videos in really high quality? If you can, please tell me.

[normal abyss]

this might not be what your looking for exactly but this has a few free Veo 3.1 fast generations https://tryveo3.ai/features/v3

not sure if thats ultra high quality though

[fickle pumice]

image 2

[pastel ember]

Hitting limits is hard, even on the free plan which is generous, and in voice mode I’ve never reached them at all, even after talking for hours.

[kindred sorrel]

Bro

[normal abyss]

oh dang thats new

with free video generations its pretty impossible to find a fully free high quality option with settings and model selection icl

[kindred sorrel]

I’ve tried ten times, but it’s not working.

[normal abyss]

wana give higgsfield a shot

[sharp marsh]

Free account?

How good is coding there

[kindred sorrel]

the thing you just mentioned, is that a website?

[pastel ember]

Yeah, Grok free is pretty generous.

[sharp marsh]

Also what model were you using of grok?

Is it the thinking mode

[normal abyss]

yeah

[kindred sorrel]

Thanks bro

[pastel ember]

I mostly use voice mode and auto, and rarely use thinking mode

[light sleet]

48 Hours, I'm in uae and not scared but cautious.

[sharp marsh]

Ah that would be the flash model then

[normal abyss]

here is a little preview

[light sleet]

Seedance 2 is free in capcut

[sharp marsh]

In reddit there were people saying thinking models have a 16 requests every 13 hours limit

Which sucks

[normal abyss]

also grok imagine is on here

[pastel ember]

free plan?

[sharp marsh]

Yea

And anyways gemini 3.1 pro is better than grok 4 so I'd rather use that

1 million context too

[normal abyss]

il say this now though, you will get 1-2 videos, a month 😭

[pastel ember]

Gemini feels overly corporate and heavily censored

[quaint lily]

how to stop beign forced using battle mode on direct since battle mode is broken

[pastel ember]

Wait for the skip button to appear

[quaint lily]

never appears

it's broken

[sharp marsh]

Way better than chatgpt

In terms of censorship

[golden ocean]

fr

where is bro using gemini from

use google ai studio

[sharp marsh]

^^

[golden ocean]

it gladly helps u destroy the world even

[sharp marsh]

And if it doesn't cooperate, gaslight it

99% chance it will let you do whatever

[rustic gale]

You have a pretty tame definition of 'whatever' if you think standard gemini 3.1 is that easy. 2.5, maybe, and even then. 😉

[sharp marsh]

Gemini 2.5 can be jailbroken

Extremely easy too

[kindred sorrel]

😂😂😂It would have been fine if I could make one or two videos a month, but that’s not happening. Bro, when I try to make one video, it’s already asking me for money.

[rustic gale]

So is 3.1, but the question is whether simple galighting, as you put it, is enough, without resorting to specialized JB techniques, unless you consider gaslight as JB

[sharp marsh]

Oh nah 3.1 cannot be jailbroken

Neither 3.0

I tried

[rustic gale]

See? As I said, you have a pretty tame definition of 'whatever'. They most certainly can, but much fewer people have the know-how which they're also ready to share. Some exist

[sharp marsh]

You said 3.1 can be JB

[kindred sorrel]

@normal abyss 🤦🏻

[sharp marsh]

I said no

[kindred sorrel]

@normal abyss hello bro

[sharp marsh]

I didn't say whatever is perverted stuff either

If you're into that it's your problem

[rustic gale]

While getting NSFW (even of tame and vanilla variety, really) is probably the most often employed usecase of jailbreaking, it's not the only one. And yet again, with gemini 3.1, those can be done. Now that I think of it, perhaps there can be some additional problems with 3.1 pro, but definitely fewer with 3.1 flash.

[normal abyss]

depends what model you choose and the settings

you get 10 free credits

[regal girder]

I must be having some issues with tokens, because on Gemini 3.1 Pro after 5 prompts it displays an error and asks me to create a chat 😥

[quaint lily]

is there an option to remove battles from direct completly?

[quaint lily]

it can there are some jailbreaks

[rustic gale]

Just open side-by-side and tell the other model to shut up constantly after explaining the situation 😄

[quaint lily]

that means creating a new chat

[rustic gale]

Well, that limitation was not present in the initial query. Restart. 🤷‍♂️ Remember that direct chat is basically charity since the site needs those juicy comarisons in addition to all your prompts

[golden ocean]

transformer

[strong ferry]

OH NO! Now I'm getting stupid token usage limits!!! WTF

[golden ocean]

LOL

[regal girder]

the same problem 😥

[bright shard]

@echo aurora I have a problem; the second message I send in code mode I get this message. And then it won't let me continue with the project, it's quite annoying.

[light siren]

by chance are u using a model with a low token usage limit

[bright shard]

I don't know, I use Gemini 3.1 Pro

[light siren]

yup...

gemini is bad at coding anyway

why are u even using it

[frozen sparrow]

https://tenor.com/view/evening-good-evening-good-night-sweet-dreams-good-night-sweet-dreams-gif-9529461948075944960

https://tenor.com/view/доброе-утро-улыбка-теплые-слова-пожелания-доброта-gif-15202361525408956675

[dense fern]

Is Claude Opus missing for everyone? What happened to it?

[sour spear]

They removed all Opus variants because those are expensive af.

Anthropic do have the best models, but they're using way more compute than Gemini or others to achieve this, hence the obscene pricing. 😉

[sour spear]

Just use AI Studio.

[wicked talon]

Microsoft copilot is actually valid

It uses Claude and gpt

[undone saffron]

#general message

[prime karma]

@sullen quest I haven't seen the Gamma 4 and Quik 3.6 on my leaderboard yet. Have they arrived yet or have they arrived?

@oak python @quasi atlas @light siren I haven't seen the Gamma 4 and Quik 3.6 on my leaderboard yet. Have they arrived yet or have they arrived?

[light siren]

never heard of quik 3.6

u mean qwen

[prime karma]

Which website and AI is this made from?

[prime karma]

Yes qwen

[normal abyss]

higgsfield, 1-2 videos a month, its kind of a 1 hit wonder

[light siren]

qwen would probably be here once they're sure their wallets can handle the model

about gemma, it's already here

[prime karma]

😂

[light siren]

u mind telling pineapple to respond to my modmail once hes here

lowkey I need to get my extension approved

[storm dust]

oh hey liam

[light siren]

hi

can u do that too

[storm dust]

any news about the extension?

me?

[light siren]

yes you

[light siren]

pineapple man hasnt reponded to me in 2 days about it

[storm dust]

i cant dm him

i need to friend him

[light siren]

in here

his dms is closed

that wont work

[prime karma]

What

[storm dust]

@echo aurora im sorry for pinging you but liam wants to talk with you

[light siren]

@echo aurora hes a staff

[prime karma]

Oo sorry

[light siren]

all good

if u see him speaking here tell him liam asked if he can respond to the modmail

havent seen him myself all day

wonder what happened

[rustic gale]

I dunno, weekend?

As in, this... place is in 6 digits member-wise. And it represents a corporate entity. Something tells me being a community manager here is a job. Which is paid and which has at least some semblance of work hours. 😄

[trim condor]

https://info.calbar.ca.gov/knowledge/en/my-state-bar-profile-attorney-annual-renewal-guide#license-status-changes

[light siren]

that would explain why his dms are closed

[cedar citrus]

he prob also is getting alot of hate of claude opus 4.6 removal

which is bad

[cedar citrus]

how do i make it use claude??

[rustic gale]

I mean, someone has to. The only good solution to this was to never even start offering it. Or immediately provide clear and concise info on what's the limits and not this 'we'll do whatever the f--- we wish and will never tell you exact numbers because f--- you, you're illiterate peasant anyway'.

[stuck orchid]

Freedom for Claude Opus!

We won’t let Claude Opus models be mistreated on LmArena or pushed out by others!

Claude Opus models have every right to be on LmArena!

They are excellent models

They shouldn’t be so underestimated, as if they don’t belong here at all

[hollow ivy]

battle-mode: exists

you just have to be smart and patient, then you can use Opus-4.6 (sometimes)

[storm dust]

even though some or most people mistreat them 🤣

[hollow ivy]

People who regularly receive gifts for free, tend to see it as their right, not a gift anymore.

(i can understand, why pineapple was not here, when it happened)

we forgot, that they spend money for us

[sweet tinsel]

Honestly Opus is overkill for most chatting tasks and only really excels at coding.

[hollow ivy]

and they still do it, and we can use over hundred of models - still for free

[stuck orchid]

He thing is, users will think that since Claude Opus models were removed from DirectChat, LmArena is being unfair to Claude and trying to discourage people from using their models. This could cause a backlash

[rustic gale]

Mistreat, lol. Were there limits? No, or at least those barely understood and often circumvented. Was there comms about what is the correct way to use them? No. Maybe about how much? Also no. Gas occupied all volume available to it; so does human nature when seeing a freebie. And then omg, abuse, creative writing for 5k words per single output (source: me, doing that all the time, for months). What was expected, really? That a model known that it's very good at something will be used for something else?

[sweet tinsel]

Back then you couldn't even really use those things in direct chat, when it was called LMSys Arena, you either had stark global rate limits or they weren't available in chat. See this time window of usable Opus in direct chat as a temporary grace period, which now ended. Also not only Claude models have been removed.

[storm dust]

what are you talking about

[rustic gale]

Given that I still see both gpt-5.4 and gemini 3.1, I still doubt that.

[stuck orchid]

It’s strange why they did this. These are very good models. They definitely belong on LmArena. It's not clear at all why they are being removed. Could it be because of their involvement with the Ministry of Defense?

[sweet tinsel]

According to the announcement #announcements message

[rustic gale]

Because a big chat could cost a hundred bucks. Each.

[hollow ivy]

i dont think so, Opus just is too resource-intensive

[stuck orchid]

Hmm, It’s strange. Very strange. There must be a specific reason why the Opus models were removed. It’s unlikely that it’s just LmArena’s own attitude toward these models

[hollow ivy]

Arena was close to going bankrupt.

[sweet tinsel]

Literally just costs. It is really expensive, even in comparison to the top-models from the other labs.

https://platform.claude.com/docs/en/about-claude/pricing

[stuck orchid]

I think if LmArena is one of the main benchmarks for all models, it's unlikely they’ll run out of money

[hollow ivy]

The Opuses are the most expensive models.

and arena.ai has limited funds

[stuck orchid]

It says something there about caching, and with it, the price is much lower. I wonder if caching was used for requests from LmArena to Claude

[hollow ivy]

(how many were there? opus-4.1, opus-4.5, opus-4.6; was there opus-4.0 ?)
(also their thinking-variants, which are even more resource-hungry)

[sweet tinsel]

Most likely, but it's short-lived and only works for already used prompts / contexts.

[rustic gale]

Well, if arena devs have found a money tree which covers most of their expenses, then the anthropic devs haven't found the GPU tree which covers all their capacity. For the use of which they're generally paid for.

(spoiler alert; there are no trees. In case it wasn't clear)

[sweet tinsel]

Also there is not really an incentive to operate a direct chat other than to lure users in to vote in battle someday. Direct chats don't have an effect on ELO rating, which is the product that they essentially provide to the labs.

[hollow ivy]

and if 100k users used an Opus each day, we get $100-200k expenses just for Opus-usage per day

or even $500k per day
that is $15 million, just for Opus, each month

[stuck orchid]

The Claude models already seem to be the most popular ones right now. Their GPU costs are likely fully covered, given how expensive their models are

[rustic gale]

Battles in direct, tho. Also all the juicy prompting

[sweet tinsel]

They simply don't have the GPUs, Anthropic invested late into GPUs. That's the reason why they also introduced lower limits during typical times of higher demand for their own subscribers.

This video explains it fairly well: https://youtu.be/j_kJNYLI6Tw?is=arvbc6RYyeKW9zQY

[static heart]

Is the Opus 4.6 trough Arena and the official claude.ai website the same? i bought the Claude Pro Subscription today and Opus 4.6 on Claude.ai directly seems different to the one on arena. i tested prompts i prompted before in arena and it responded verry differently

[hollow ivy]

Interesting.
Can you test with different temperatures?

[loud verge]

I remember some announcement about models being removed from direct chat.

[opaque maple]

hey can i ask you guys smth

[hollow ivy]

and/or different P-top values

[opaque maple]

i know its ai slop but

[loud verge]

Is that actually happening? If so, when?

[cedar citrus]

bro what is this delete this bradar this is completely unrelated bro

and its kinda mid idk

also the ai companys give acess to lm arena the ai models for free because they will get more visibility and free datasets

free datasets is the most valuable thing

[static heart]

where can i play with that?

[hollow ivy]

in google AIstudio

i thought, Anthropic had something similar

[cedar citrus]

if they had this they would go bankrupt

[sweet tinsel]

Not really. He could use the API although.

[hollow ivy]

(i cant use it, because they demand a phone number)

[cedar citrus]

use openrouter

[sour spear]

Have you ever actually paid for Claude Opus? Try it, their API costs will blow your socks off, it's completely ridiculous. 💸💸💸

[hollow ivy]

that needs money, to recharge your credits, which i dont have

[cedar citrus]

the ai companys give acess to their models for free in exchange of free datasets

simple

[hollow ivy]

(and poe.com only offers 1 message per day; opus-4.6 included)

[cedar citrus]

free data is 100x more valuable than the costs

[cedar citrus]

try genspark

[unreal hatch]

its say something inappropriate in that bottle

[static heart]

nah im using opus 4.6 from the official claude.ai site. dont see any settings to play around

[cedar citrus]

genspark gives unlimited usage to paid users

[hollow ivy]

that one only offers a one-time credit infusion

[sweet tinsel]

You missed the point of direct chat. See #general message.

[hollow ivy]

no daily credits

[sweet tinsel]

They don't get the Data for RLHF from the direct chat.

[cedar citrus]

makes sense

still

they should only remove from direct

[sweet tinsel]

And it's not worth it if people never go over to the battle mode and just use Opus directly.

[hollow ivy]

users who paid, that is the point

[sweet tinsel]

Opus is still in battle mode.

[unreal hatch]

btw the all models have some limits on genspark like Opus 4.6 cant write 3000code line

[winter cypress]

i want create video

[cedar citrus]

dawg its unlimited

i coud literally use 100t tokens

and they arent charging a single cent

i could recreate 100 node js and they arent charging me a single cent more

[sweet tinsel]

Fair use limits are going to quickly show you how there is never a real unlimited product.

[hollow ivy]

but it still costs you $240/year, which i dont have

[cedar citrus]

im brazilian

so for me it would be 1200 reais

[slate hare]

Hi everyone I just want to inquire about Claud Opus Removed

[winter cypress]

i want create video

[cedar citrus]

but honestly i understand your pain

[sweet tinsel]

Use the website.

[slate hare]

Does anyone know the answer?

[sweet tinsel]

You haven't even asked yet really.

[light siren]

wdym

[hollow ivy]

lol, i never said, i have pain
because, there is a chance to get opus-4.6 in battlemode

(but low, so one has to be very patient)

[cedar citrus]

uh hate bro

[cedar citrus]

pain from not being able to afford stuff

like me

[slate hare]

Why was Claud Opus 4.6 Thinking removed and will they bring it back?

[light siren]

ofc it's bad

[cedar citrus]

then why did you say wdym

[compact comet]

poe.com is actually cheaper than regular claude? their rates seem reasonable

[light siren]

I thought u meant smthn else

[sweet tinsel]

Just look a bit up in the chat. I somehow covered the topic already not long ago.

[cedar citrus]

is it unlimited chat usage?

[compact comet]

no of course not

[cedar citrus]

ah okay

[cedar citrus]

i dont think its cheaper

[compact comet]

it's 75k points per 100k output tokens

[cedar citrus]

api costs are higher than subscription costs honestly

[rustic gale]

'Who would explain that to you, you pleb' and 'no' are the answers you should count on, regardless of what is actually said.

[cedar citrus]

actually not bad

[sweet tinsel]

Depends on your actual usage.

[hollow ivy]

all Opus* models had been removed from direct chat and side-by-side chat (not from battle mode), due to cost

[compact comet]

and they give you 1M points per month

[cedar citrus]

not bad

[sweet tinsel]

But it's true that most subscriptions are subsidized.

[cedar citrus]

they do more for subscriptions because you are giving them money monthly

[cedar citrus]

api is one time only

[sweet tinsel]

I know. The API is directly usage based.

[cedar citrus]

yes

[slate hare]

Will they bring it back or not?

[hollow ivy]

not this year

[sweet tinsel]

See here. #general message #general message #general message

[hollow ivy]

(…at least, that is very unlikely)

[slate hare]

Why

[sweet tinsel]

But note that it's still available in battle mode. The actual product.

[cedar citrus]

prob not until anthorpic lower costs

i wish anthropic lower costs of their products

[compact comet]

or leechers stop abusing their services

[sweet tinsel]

#general message

[cedar citrus]

yes

[slate hare]

How long will this take?

[cedar citrus]

with anthropic blocking openclaw with the subscription usage the usage limit will rise prob

[compact comet]

somewhere between never and never

[cedar citrus]

they gotta do optimizations

[cedar citrus]

like they did with opus 3 that was 75 dollars now opus 4.6 is 25

[sweet tinsel]

They actually don't have to do it. But how the space is moving you can't predict it that accurately but Anthropic has a tendency to create high cost models.

[cedar citrus]

i think they will now that they became the new "chatgpt"

[compact comet]

massive ai models are expensive to run and reasoning takes 10x the cost, more news at 11

[hollow ivy]

i wonder, when arena.ai will add Spud..

[slate hare]

What About Gemini 3Pro

[hollow ivy]

(aka GPT-5.5)

[cedar citrus]

because everyone uses it now their cost will be high so they will optimize

[hollow ivy]

3.1-pro still exists in arena

[sweet tinsel]

They simply don't offer Opus to free users and let Enterprises pay for overpriced API prices.

[compact comet]

nobody offers their strongest models to free users, that would be utterly insane

[cedar citrus]

yes but the thing is they became the new chatgpt everyone is using them so it will be higher costs for their servers so they will run some optimizations idk if they will lower api costs but sure they will raise usage limits

[slate hare]

He doesn't exist.

[slate hare]

I'm asking for version 3 pro and not 3.1

[hollow ivy]

yep
replaced by 3.1-pro

[cedar citrus]

opus 4: $15/M input tokens
$75/M output tokens opus 4.6:
Feb 4, 2026
1M context
$5/M input tokens
$25/M output tokens

3x less

[slate hare]

How much does it cost to subscribe to anthropic using Opus

[cedar citrus]

what?

[compact comet]

claude limits for opus are ridiculous

[west lodge]

alot

[compact comet]

the API is cheaper

[cedar citrus]

i dont think so

[compact comet]

unless you go for the max subscription, in which case the limits are reasonable

[cedar citrus]

oh okay yeah

[sweet tinsel]

For the broad masses they aren't really the new thing and Sonnet is already fairly cost optimised, so that's what you can see as an optimisation with its trade-offs.

[cedar citrus]

that i agree

[cedar citrus]

sonnet honestly has great pricing

[sweet tinsel]

Distillation isn't lossless.

[west lodge]

shill

[sweet tinsel]

(Hint: it's still not great in pricing)

[cedar citrus]

the only thing they have to optimize now is opus and then they raise usage limits and keep api costs

[sweet tinsel]

You're better off using GPT and Gemini models.

[cedar citrus]

because entrepise gang will pay alot money

and people will subscribe to claude

monthly money

[sweet tinsel]

They can't simply put the optimising lever on it. There are quality drop-offs always expected with a lower parameter count.

[cedar citrus]

fair enough

this will be some sort of industrial revolution

im sure

its like with pc parts

a part from 2006 at the same price as a part from 2026 the 2026 version will perform way more

will happen with ai

[compact comet]

the inevitable crash of the insanely inflated AI race will lower prices but don't count on it

[cedar citrus]

mainly because of robots

[west lodge]

im hoping for corporate/closed AI To crash and burn

[cedar citrus]

dawg i dont think the ai race is stopping

[west lodge]

and for open weights models to reign supreme

[cedar citrus]

the ai race is the third industrial revolution

[hollow ivy]

Gemma 4 is decent (for an open model)

[cedar citrus]

count my word

[west lodge]

yeah ive noticed

[compact comet]

I don't need your word, economists, scientists and politicians agree with me

[west lodge]

im just looking at how i can get a cheap gpu with actually good amounts of vram to run the decent gemma modes

[compact comet]

stop consuming corporate slop

[west lodge]

because 12gb isnt enough

[cedar citrus]

gang

do you prefer

writing code for a software in three months

getting 1 second of sleep

[west lodge]

i CAN run the MoE model with some offload but its kinda slow compared to the cloud options and i cant give it much context so

[cedar citrus]

or do you prefer doing it instatly getting the same money

[west lodge]

not really a good replacement for gemini at

m

[compact comet]

what is blud talking about

[west lodge]

https://cdn.discordapp.com/attachments/1077690385855418510/1474328003637940296/attachment.gif

[sweet tinsel]

Good luck. You won't get it. It eats 101 GB of RAM with Q8 Quantisation with GGUF.

[compact comet]

AI is very useful yes, it is still crashing and burning in its current state

[west lodge]

woah not that big

im talking about 31B

its decent enough

[compact comet]

it's not that hard to understand that it's not sustainable

[cedar citrus]

best option 3090 or just wait for intel arc b70 with 32gb of ram

[sweet tinsel]

Still a lot of CPU offloading necessary.

[cedar citrus]

intel arc is 1000 dollars

[sweet tinsel]

You're better off renting an H100 from runpod or such.

[cedar citrus]

for a 30b paramater model nope

[west lodge]

if i can get a good GPU taht can keep gemma 4 31b in full vram with a good context window im set

[sweet tinsel]

This thing eats up more than you think.

[west lodge]

ik benches arent anything to trust but

[cedar citrus]

if a 70b model then yes

[west lodge]

this is kindof a win

[sweet tinsel]

Have you actually tried running it?

[west lodge]

especially if i use heretic

[cedar citrus]

depends if its completely quantizated

[sweet tinsel]

Did you never use any local AI?

[cedar citrus]

my gpu aint that good

[west lodge]

if 31b actually does better than g3 flash in real life then im set

[sweet tinsel]

It eats WHOLE 101GBs with Q8.

Already quantitised.

[cedar citrus]

i have an i7 9700 rx 550 and 64 gb of ram

[cedar citrus]

im using claude to prove you wrong

[hollow ivy]

with that machine, you are not poor

[west lodge]

wtf ishtis bro

[west lodge]

god damn

[cedar citrus]

bro the 101GB figure is for BF16 unquantized at FULL 256K context, nobody runs it like that locally 💀

Q8 31B fits on a 32GB GPU. Q4 31B fits on a 24GB GPU (RTX 3090). that's documented on unsloth's page and confirmed by people actually running it in llama.cpp

also the 26B-A4B MoE is probably the better pick anyway — only 4B active params per forward pass, way faster, and a 3090 runs the full 256K context on it with room to spare

so yeah, 3090 is the move

[sweet tinsel]

I tried to run it myself, so I know it for a fact.

[sweet tinsel]

I used Q8.

[rich sable]

I'm new here, does anyone know where the Claude Opus model has gone?

[cedar citrus]

what GPU were you running it on and what tool did you use?

because there's a known LM Studio bug that shows inflated numbers for gemma 4 31B — someone on the unsloth HF page literally posted about it and closed the issue after confirming it was wrong. in llama.cpp the actual VRAM usage at Q8 is way lower than 101GB

also if you were counting system RAM + VRAM together during CPU offload that'd explain it. that's not the model "eating" 101GB, that's just your tool reporting total memory across RAM + VRAM

[hollow ivy]

It's gone for good.

(was too expensive for arena.ai to sustain)
but it still exists in battlemode

[sweet tinsel]

Also with vLLM on Runpod and Huggingface Inference services it tells me I need 2 A100s to fit this thing.

[west lodge]

damn this would be a great offerr (if it wasnt like mined on or smth) if it wasnt like a whole channel away from me

[cedar citrus]

bro vLLM and HF Inference run BF16 by default, no quantization

2x A100 = 160GB total — that lines up perfectly with full precision weights + KV cache overhead, not Q8

you're literally proving my point. the moment you quantize to Q8 with llama.cpp it fits on a 32GB GPU. those cloud services just don't do local GGUF quants

[west lodge]

how fast would a 31b model even run on a 3090

[compact comet]

it probably wouldn't even run ngl

[west lodge]

wym

it would run fine if you just quantize it

[rich sable]

life hack

[compact comet]

not for any reasonable definition of "running"

[cedar citrus]

im answering you

give me a sec

[sweet tinsel]

That's true, that's a small error of myself in logic. Part of being responsible is to know when to signal defeat, even if you're drafting your messages with Claude, I looked at the official page from Google and it doesn't eat up this much.

[compact comet]

yea sure but it's like 1 token per 5 billion years 😭

[west lodge]

great

[cedar citrus]

31B Q4 on a 3090 sits around 30–34 tok/s which is totally usable

but honestly the 26B-A4B MoE is the smarter pick — it hits 64–119 tok/s on a 3090 and runs the full 256K context with room to spare, quality is only slightly below 31B

so if you want speed + big context, go MoE. if you want max quality and don’t mind slower, go 31B

[west lodge]

huh alright

cuz atm i only hit like 15tok/s on my 4070 😭

with the MoE

[rich sable]

soo real life hack?

[cedar citrus]

prob smt wrong i think so let me check

[cedar citrus]

that's because the 4070 only has 12GB so it's offloading layers to CPU RAM, which kills speed

on a 3090 the whole model sits in VRAM and you get the actual 30-34 tok/s — no offloading bottleneck

[west lodge]

uhm yeah

[cedar citrus]

guys i have a question

[west lodge]

why are you getting claude to generate your responses btw

that did NOT warrant a prompt

[sweet tinsel]

You know what, I'll spin up a VPS real quick equipped with an A100 to see rq if it works with Q8 quants tomorrow. I'll message you @cedar citrus.

[cedar citrus]

who is better runpod or vast.ai

[cedar citrus]

okay artupia

[cedar citrus]

your the goat

[sweet tinsel]

Actually I'll do it tomorrow because it's Late here in Germany.

[sweet tinsel]

And I want some sleep on those festive days.

[west lodge]

https://tenor.com/view/johnny-joestar-king-von-king-von-johnny-joestar-anti-piracy-screen-johnny-joestar-tiktok-sticker-netflix-gif-15496977903175842722

[hollow ivy]

how good is the Gemma 4 E4B SFP8(8bit) in roleplaying?
(That is the biggest of gemma, which i might barely be able to run here, lol)

[west lodge]

truly who gaf

[rich sable]

fake

[cedar citrus]

great you live in europe you are lucky anyways good luck

[sweet tinsel]

Use whatever you please, in my experience runpod has more stable prices though that don't jump around and their UI just feels smooth.

[rich sable]

If you want to reuse Claude Opus 4.6, go to a chat where you've already spoken with him, and there you go, you can reuse Claude.

[cedar citrus]

bro i was in a debate i needed sources 😭

[west lodge]

google bro

you dont need claude to wear your skin

just use that brain of yours

[cedar citrus]

how are you getting acess to claude opus 4.6 in lm arena dawg

[west lodge]

aipac

bro trusdt

[cedar citrus]

bro i pay for claude im gonna get my money worth

[heady whale]

nope

does not work

[west lodge]

also slightly below is a bit of an understatemet

[cedar citrus]

uh your comparing gemini 3 flash not gemma 31b

its gemma 31b q4 vs gemma 26b a4b MoE

[west lodge]

oh wait

gemma 31b was compared against 3 flash non-reasoning

[cedar citrus]

lol

[west lodge]

yeah makes more sense now

[stuck orchid]

Has Claude Opus made a comeback?

[sweet tinsel]

It's still pretty good for its parameter count.

[west lodge]

but still

kinda rough

better off just living with the slightly slower speeds of a full dense model

or just

use qwen

LOL

itll probably run even faster than gemma cuz less params

still dense so itll be slower but

worth it

[stuck orchid]

Gemma - Top Model?

[west lodge]

nope

[sweet tinsel]

It's not SOTA.

It's just good for local use.

And excels at its domain, being edge deployment.

[stuck orchid]

They should have released Gemini 4 instead

[west lodge]

if you're coding with local models just uhh

use qwen3.5

its somehow better

[hollow ivy]

better than any other open-weights local model?

[sweet tinsel]

Not really. Gemma was always something separate and it's good to see something open source.

[west lodge]

and doesn't start repeating tool calls as quickly

[west lodge]

yeah i think gemma is more oriented towards vision tasks

not coding

[sweet tinsel]

That you can run feasible on consumer machines.

[stuck orchid]

In AiStudio, you can see Gemma thinking, without any generalizations

[west lodge]

well yeah

why would they summarize an open weight model's reasoning

[stuck orchid]

It's strange that Glm-5.1 isn't available on LmArena

[hollow ivy]

which model of gemma would you recommend for a 6 GB GPU?

(i also have 16GB main ram)

[west lodge]

if you dont mind offloading then E4B

if you DO then E2B

E4B should still be fast enough even when offloading since its low params

(effectively)

[sweet tinsel]

A small one. I think there is a Website to estimate how much you can run. But generally try to use the smaller Qwen ones.

[sweet tinsel]

I'm unsure if E4B really runs on this thing.

[hollow ivy]

is Qwen-3.6 better than all gemma models and also better than Qwen-3.5?
(at their website)

[west lodge]

it could if you offload

and run quantized

[west lodge]

yeah but the only 3.6 model right now is closed

so we cant really tell;

[hollow ivy]

is Qwen-3.6 also better than GLM-5.1?

[sweet tinsel]

Well. Depends. Qwen 3.5 Max obviously has an edge over 3.6 Plus, but 3.5 Plus compared to 3.6 Plus should be worse. Although it generally doesn't perform this well.

[sweet tinsel]

Definitely no. It really is bad at coding as an example.

[west lodge]

usually .1 increments (which everyone seems to be doing after openai started doing it) don't really provide a big ste pup

its more refinement or more RL

[hollow ivy]

so, GLM-5.1 is currently the best chinese model for coding?

[west lodge]

.5 or +1 increments usually have the best improvements (like with gemini, where the flash models outperform last gen's pro models)

[sweet tinsel]

At least for Agentic Use definitely.

[west lodge]

son what

i mark the sata port as hotplug and i can do this

[hollow ivy]

and which model would you recommend for roleplaying and realistic sandbox games, where the AI has to be an intelligent GM?
(apart from opus)

[sweet tinsel]

Minimax 2.7 works also pretty well but I believe that GLM 5.1 has an edge over it also in raw code output, even without a harness.

[hollow ivy]

(thanks for your answers so far!)

[sweet tinsel]

I don't do that with AI Models honestly, seems like a waste of time and capabilities for me, so I have no real insights on that. Prolly any non-coding optimised model with a high parameter count would work well on it.

Genuinely the creative writing outputs of large models are very good. GPT 4.5 had a very great writing style and that falls a bit under the same category.

[sweet tinsel]

I remember your name from the old days if I recall correctly. Since when are you on here active again?

[hollow ivy]

pretty long
what have we talked, back then?

i'm also in the other discord (lmsys)

[sweet tinsel]

Also back in the times where the Arena part was still under the LMSys Server?

[hollow ivy]

i used sonnet in lmsys in summer 2024

i think that was during the big Claude-Sonnet-3.5-moment

[sweet tinsel]

I at least remember you and maybe you don't recognise me because I've been a bit more inactive here recently and changed my PFP to my cute rabbit called "Suno".

[hollow ivy]

nice, thanks for remembering me

rabbits are indeed cute

[sweet tinsel]

I think I got in during the times where GPT 4 with 32k context window got released.

[hollow ivy]

(and i know of a boardgame, where they play a role ^^)

when did you start to "vibe code"?

or AI-code

[hollow ivy]

in that one, yeah, also their discord (during the Gradle.io time)

could one vibecode with GPT4-32k?

[sweet tinsel]

I tried it out over time again and again always with miserable results. I to this day don't really completely vibe around but recently I noticed that the models are actually capable now, especially with harnesses like OpenCode, so I've been doing a bit more testing over the last month.

[hollow ivy]

(i never really tried that out)
(or maybe i did , but just with bash script? i cant recall)

[sweet tinsel]

GPT 5.4 for Backend and Gemini 3.1 Pro for front-end work really well.

I mostly keep Opus and Anthropic models out of my life.

I've been a Plus subscriber since the founding days of ChatGPT and it serves me well with my GitHub Copilot Subscription that I recently got.

[hollow ivy]

what do you think about Grok?

[sweet tinsel]

Basic HTML pages worked or simple phyton scripts, not really more.

[sweet tinsel]

I actually rotate AIs a lot for trivial tasks and in that it has been performing relatively well.

The speed to quality ratio was just great with multiple agents.

[hollow ivy]

Grok-4.20?

their 4-agent system?

[sweet tinsel]

Sometimes I am stuck waiting for my GPT 5.4 to finish thinking, for up to 45 minutes.

[sweet tinsel]

Yes.

And the normal Grok 4.2 Thinking also serves me correctly.

[hollow ivy]

so, which current AI system in direct chat of arena.ai is the best for pure vibe-coding?

(of all existing ones)

if one can not code, but takes 100% of the AI code face value

[sweet tinsel]

I don't use the direct chat to try this out at all anymore as it's really unreliable. You're better off using an argentic harness and leech of free usage from Anti-gravity and OpenCode.

You can also get the GitHub Copilot Student Subscription for free if you're in that age range.

But for direct you could use Models like GLM 5 I guess.

[hollow ivy]

even for harder projects such as letting it write an AI for niche boardgames?

[sweet tinsel]

If you want Website slop don't use Loveable or such. That's just a nightmare to host outside their ecosystem, use AI Studio or v0 for that.

v0 gets you limited free access to Opus 4.6 btw

[hollow ivy]

i mean, for offline apps

not web-coding

[sweet tinsel]

5 dollars of inference monthly

[hollow ivy]

(i use Linux Mint 21.3)

i dont have paypal and i wont ever use a credit card online

there's always the danger of getting hacked/leeched

so it has to be free

[sweet tinsel]

For apps you'd need some way to build them actually. Or you could just run straight Phyton as an example.

For offline apps I really wouldn't use Arena.

[hollow ivy]

i can build apps in C/C++ here, also Java and Python

and Rust (but this is a bit harder)

[sweet tinsel]

Use something like the free agent from GLM, which has generous limits or really just run Anti-gravity and OpenCode locally.

[hollow ivy]

can i use Sonnet-4.6 for free, using that harnesses?

[sweet tinsel]

With Anti-gravity you can do that with usage limits which are restrictive although.

You can also use Opus there as a free user, but with limits that won't get you far.

[hollow ivy]

1 message per week? (for opus)

[sweet tinsel]

Token based limits. No real message limits.

[cedar citrus]

when are glm 5.1 dropping its api

[hollow ivy]

ok, that does make more sense, ofc

[cedar citrus]

they holding the model like its the holy grail

[hollow ivy]

isn't it already available in their chat?

[cedar citrus]

nope

[sweet tinsel]

Maybe when they open-source it and it's completely finished. They said that they are doing finishing touches. It's currently only a preview for subscribers.

[cedar citrus]

it will prob shake the open source model market

if its comparisons against opus

are true

[hollow ivy]

do you think GLM-5.1 will be better than Deepseek V4?

[cedar citrus]

prob

deepseek fell behind

if glm is right of the benchmarks being compared to opus then boom

[hollow ivy]

also better than GPT-5.4-high and Grok-4.20?

[cedar citrus]

idk

its what they say

we cant know for sure until they release to the public

[hollow ivy]

i think, i already chatted with GLM-5.1 at their website..

but i'm not 100% sure

maybe they do a partial rollout?

[sweet tinsel]

I don't think it's going to be this great.

There is something called "Benchmaxxing"

Common practice, ask Meta.

[hollow ivy]

yeah, unfortunately

they especially train for the benchmark, which defeats the purpose

[cedar citrus]

idk glm is getting close to opus every time honestly

[cedar citrus]

answering your question

[sweet tinsel]

But Anthropic is also moving forwards.

[cedar citrus]

fair enough

[sweet tinsel]

It's only in the coding plan.

[cedar citrus]

anthropic right now is unreachable

[hollow ivy]

ah ok, so it was GLM-5-turbo

[sweet tinsel]

In the interface for coding.

[cedar citrus]

guys

did you guys get acess to the video api in openrouter

i actually got acess

but its expensive as heck

[hollow ivy]

especially if Mythos 5 exists (and is not a fake news)

[sweet tinsel]

Let's see what they will cook up. They quietly have been at the forefront of research. Just silently. They dropped some very good papers.

[cedar citrus]

it def exists they just have to run some optimizations

[cedar citrus]

thats great

i hope they cooks

i love their pricing

[sweet tinsel]

And their main selling point will be their lost reliance from Nvidia as they use Huawei for Inference and such.

[cedar citrus]

its gonna be great for me for my api uses

but i am using nvidia nim api

[cedar citrus]

fair enough

[sweet tinsel]

1 percent of battery left and it's midnight. I think I'll head out for now.

[hollow ivy]

do you guys think, chinese models could permanently claim the second place behind Anthropic?

and ahead of Google deepmind, OpenAI and xAI?

[cedar citrus]

bye

[cedar citrus]

google deepmind is dropping great papers

i think they will be ahead of gpt and xai right now

anyways im heading out

im going to eat

im sending you a message

so i dont lose a contact

eh whatever

its blocekd

[hollow ivy]

sorry, but DMs are globally off for me

[cedar citrus]

add me if you want

[slate hare]

Gemini 3 Pro or Opus 4.6 is better

[hollow ivy]

(due to past discord shenanigans, i had to do that)

[cedar citrus]

opus 4.6 for sure

[cedar citrus]

dont worry i dont do that

you can add me if you want

im nice

anyways byee

[hollow ivy]

ok, i mentally added you
we can always talk in #1472698608980856842

(or here, in general)

[sweet tinsel]

That's Backend. For the love of god trust the autistic GPT 5.4 and 5.3 Codex with that kind of task, it really is great at it.

That message disappeared now I guess.

[hollow ivy]

i heard, 5.4 isnt that great in coding

[sweet tinsel]

But this message is for you @slate hare.

[hollow ivy]

and neither is 5.3

[slate hare]

Thanks

[hollow ivy]

but 5.5 should be

[sweet tinsel]

You heard wrong. It works wonders at back-end, just for the front-end it delivers slop.

[hollow ivy]

ah, ok

[sweet tinsel]

Others share this opinion from what I saw.

[hollow ivy]

even in sizable C/C++ programs?

[sweet tinsel]

Most likely yeah, although I don't really code in that language.

[sweet tinsel]

There still are limits but it's generally strong and has a 1M Token context window.

[slate hare]

So GPT 5.4 is better than Opus

[hollow ivy]

no, i bet not

[sweet tinsel]

Task relative but it's generally very good at back-end.

[hollow ivy]

maybe better than Sonnet-4.6, but never Opus-4.6

[sweet tinsel]

They really improved their coding capabilities lately.

[hollow ivy]

less hallucinations?

better understanding of the code?

and better understanding of the user intent?

[slate hare]

I want someone who is excellent in understanding the codes and knows how to deal with it

[slate hare]

I think so.

[hollow ivy]

so we await these models:

• Grok 4.5 or 5
• GPT-5.5 ("Spud")
• Claude Mythos 5
• GLM-5.1
• Deepseek V4
• Gemini-3.2 or 3.5

[slate hare]

I think Claud would be the best.

[hollow ivy]

yeah, CM5 is the goat

Spud might be second

[sweet tinsel]

Actually it's really very good. As an analogy Codex is the autistic senior Dev and Opus the over-eager developer which doesn't reach Backend capabilities of 5.4 that often.

[hollow ivy]

which is best of these? and which comes second?

• (current) Minimax
• (current) Kimi
• (current) Mimo
• (current) Qwen
• (current) Deepseek

[sweet tinsel]

Mythos will most likely never get released to the general public in its current large state. Maybe when it's smaller or they simply use it to distill Opus 4 from it.

[slate hare]

1)Kimi/2)qwen/3)deepseek...

[hollow ivy]

and GLM ahead of them, right?

[sweet tinsel]

Are you using it for coding? Qwen is absolutely criminal at coding, it's dumb like bread and Deepseek is outdated.

[slate hare]

All of them are from China and they will be close

[slate hare]

I know

[sweet tinsel]

I would say

1. GLM
2. Minimax (2.7)
3. Kimi
4. Mimo
5. Qwen
6. Deepseek

[hollow ivy]

wow, i didn't know, Minimax was that good

and GLM rivals Gemini 3.1 pro? or maybe even Sonnet?

[sweet tinsel]

It really is, 2.7 can be a real machine, albeit I yet have to wait on its OS release.

[sweet tinsel]

Definitely not in front-end, for which I use Gemini but possibly for back-end as both have been equally bad in my experience.

[hollow ivy]

and back-end coding is harder for AI, right?

[sweet tinsel]

Maybe GLM could be better than Backend than Gemini with 5.1

[hollow ivy]

as it can not be done as fast in python

[restive crypt]

yo everone

wsp

[sweet tinsel]

Depends really on how much data the model got and which.

Codex is the worst at front-end.

[restive crypt]

why isnt opus 4.6 misiing

[hollow ivy]

is Python the best language for front-ends?

[restive crypt]

on web

[sweet tinsel]

Do you mean that it is missing?

[restive crypt]

nah i cant see it on website

[hollow ivy]

opus shy

[sweet tinsel]

When I mean front-ends, I mostly mean the web. You mean GUIs for apps, right?

[hollow ivy]

GUIs yeah

[sweet tinsel]

It's gone from direct chat, try out your luck in battle mode.

[restive crypt]

forver??

[hollow ivy]

probably, yeah :/

to save cost

[restive crypt]

nooooooooo

tf

[restive crypt]

yeah ik its expenivee

[sweet tinsel]

Ooooff, something that I rarely use AI for, I used Opus 4.5 for that last time and it worked fairly well, I have honestly nothing more to say.

[restive crypt]

broo

chat isnt giving that felling that it was giving before

[sweet tinsel]

#general message

[hollow ivy]

i don't understand, why everyone wants to do web-dev
it is way inferior, because slower