general

[shrewd citrus]

and since 5.5 is only on codex rn I think the exact same thing is happening

[frosty lava]

wasn't it like days ?

i don't remember honestly

but what matter is that it will go public

at some point

[loud verge]

https://cunnyx.com/i/status/2047382400112660608

leo 🐾 (@synthwavedd)

Perspective helps!
︀︀
︀︀GPT-5.5 underperforms Mythos on:
︀︀- SWE-Bench Pro
︀︀- HLE
︀︀
︀︀It is basically on-par on:
︀︀- GPQA Diamond
︀︀- BrowseComp
︀︀- OSWorld-Verified
︀︀
︀︀It is better on:
︀︀- Terminal-Bench 2.0
︀︀
︀︀All while being more token efficient, smaller and cheaper than Mythos (and actually available!)

Quoting leo 🐾 (@synthwavedd)
︀
GPT-5.5 benchmarks are out
︀︀
…

CunnyX · Apr 23, 2026, 6:29 PM

[main nexus]

Gpt 5.5 out?

[indigo knoll]

Gpt 5.5 is available on free plan Chatgpt?

[frosty lava]

No

[frosty lava]

yes its out

[grand raft]

but not on free plan

[grand raft]

i need the 5.5

[loud herald]

Kimiiiiiiiii yessss

[zenith steppe]

Kimi yes kimi no , is it jailbroken or just like that?

[loud herald]

Its jailbroken but its so easy to do so

Thats the reason I like chinese models

They always have low guardrails

[zenith steppe]

They are like that naturally not even by design , lol

Isnt kimi a steal from claude?

[vale quest]

Kimi is also disabled on arena.ai

Because everyone spent their balance

[whole sundial]

it's not?

[vale quest]

Now try to use it

[whole sundial]

no kimi models work

@echo aurora here is a trace id for you:
:19f5165f-0c6b-

also btw i was able to get the reason why it failed, i guess arena is broke

Your account org-3768766e50c242e2ade5fc3b3b783831 <ak-f4h9btz5i7s111b3pub1> is suspended due to insufficient balance, please recharge your account or check your plan and billing details

i can donate my moonshot ai key to you guys if you need it /s

[primal orbit]

is there still any chance to get opus 4.7 thinking in battle mode?

[whole sundial]

i think we can blame moonshot ai here for not having an auto-topup function

i looked and i didn't find one, unless i am missing it

[whole sundial]

also btw @echo aurora if you want a simple message to say to the devs, just say "Kimi models are failing due to Arena's Moonshot AI platform account being out of balance, you have to top it up to fix it. I don't believe Moonshot has an auto topup function, so you'll have to check on it often."

also btw i really do have a moonshot ai api key, it still has some balance on it and i have some experience with the platform

[void shore]

i just have a feeling that the people using kimi are gonna drain account balance really fast

so it makes sense that it would go broke

[whole sundial]

yeah they've made it more expensive this time around, probably to make up for the price cuts and speed improvements made during kimi k2.5's time

[void shore]

and its also open source

[whole sundial]

but once again i blame moonshot for not offering automatic topups

[void shore]

if anyone has that much gpu power

than download it and run it locally for others to use

[loud herald]

Well they wont focus on putting guardrails up thats why and I approve haha

[sly cedar]

Is it still able to use or no?

[loud herald]

Chinese models distill from anything so yes

[whole sundial]

you'll have to wait on arena to refill their moonshot account

[sly cedar]

I mean, if i use it, what notification it will be?

An error?

[whole sundial]

it will be a something went wrong error

[void shore]

long story short, no.

their account balance is empty at the moment, so no requests can be sent through

until it gets refilled, you'll just get an error message

:3

[whole sundial]

the message i sent was from arena now sending partial trace to users, I extracted it and got the message

[sly cedar]

Does anyone have thoughts on openmythos?

[urban herald]

i need it :3

[loud herald]

Seems fine to me

Any company doing this though would 100% be doing a pay as you go

Not credit system where they have set numbers

[sullen sable]

.

[inner relic]

guys

deepseek v4 is here

https://api-docs.deepseek.com/

there's deepseek v4 pro and

uh

flash

[grand raft]

#1372229840131985540

[vale quest]

Bro is not a mod

[grand raft]

yeah

[echo aurora]

Okay thank you for the heads up, looking into and flagging

@whole sundial are you sure this is the case? I'm not getting any issues with Kimi models.

[limber crag]

Hey weren't there more tape models?

What happened to the rest

@pineapple

[echo aurora]

Sorry to say I can't go into details about codenamed models

[limber crag]

No issues!

Can i suggest a feature?

[whole sundial]

they may have already fixed it

[limber crag]

How about we can like vote on other's battle mode generations?

[whole sundial]

are you getting responses?

[limber crag]

Like it can be a scrollable thing

[echo aurora]

Yeah, have tried out a bunch of and they all seem good 👍 What makes you think it was balance related and not some other error problem?

[echo aurora]

That would be super cool. Some kind of social aspect to it where others can vote. Would be really interesting to see those leaderboards too.

This idea has been something we've kicked around a bit.

[limber crag]

Ohh any plans on working on it then or you cant talk about it

[inner relic]

how these dudes are not excited about deepseek v4

bro

ok i am posting nothing her

[main nexus]

This real???

Dude

[echo aurora]

Nothing that I'm able to share.

[inner relic]

yes t's real

[main nexus]

Deepseek v4 better cook my meal or else this model sucks

The hype better be good

Gotta be better the opus 5

And gemini 3.5

[whole sundial]

arena returned an error from the moonshot api to my console that said the account was out of balance

[obtuse smelt]

hmm well is max generating image is 3 not much more ?

[inner relic]

https://x.com/ValsAI/status/2047513613750202452

Vals AI (@ValsAI)

DeepSeek v4 is now the #1 open-weight model on our Vibe Code Benchmark, and it’s not close.

It leaves the #2 (Kimi K2.6) in the dust, and even beats out frontier closed source models like Gemini 3.1 Pro.

X · Apr 24, 2026, 3:11 AM

[obsidian cargo]

It's already on direct and side by side mode!

I hope it stays

[minor bloom]

DEEPSEEK!!!!!!

[wicked talon]

Wtfff

No one told me about this

[wicked talon]

Bruhhhh

[grand raft]

but the announcement

[echo aurora]

🐳

[echo aurora]

incoming

[minor bloom]

Wait

[minor bloom]

Why can't I upload images?

Or files?

To deepseek

[stray aspen]

Deepseek is out

[vernal raft]

I'm legit confused

I don't see it on arena

Can't tell if those are all ai from gpt 2

[stray aspen]

https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro

[vernal raft]

Omfg can't distinguish ai from real

[rigid copper]

awwww.... :/

[wicked talon]

Wait did it literally just come out?

[minor bloom]

Yes

[wicked talon]

Oh

[minor bloom]

Like 10 minutes ago

[wicked talon]

I didn't even realise

[wicked talon]

30*

I was wondering why it wasn't on deepseek app

Probably will be released later today 🙂

For now gotta use arena.ai

[minor bloom]

I wanna see benchmark results

Its probably on the same level as Gemini 3.1

[wicked talon]

It's gonna benchmaxx

[minor bloom]

But worse than opus

[wicked talon]

Wouldn't surprise me if it's close to Claude

[minor bloom]

And 5.5

[wicked talon]

Nah

[vernal raft]

Please tell me this model can kick anthropic from the throne

[stray aspen]

5.5 sucks

[grand raft]

lets see if this model is good

[minor bloom]

Need to compare it to kimi

[stray aspen]

Wait for mythos to crush it

[rigid copper]

@echo aurora what was that actually mean? like maximum of 10 attachment per chat?

[stray aspen]

Im still waiting for the spud

[minor bloom]

Only 5 people have access to that

[echo aurora]

Correct

[vernal raft]

Yeah 100€ per 1m can stay there overpriced for enterprises

[wicked talon]

Why is flash actually fast asf

[wicked talon]

Mythos is ass bro

Claude is definitely over hyping it

Glazing

[vernal raft]

Please someone tell me that this model is actually running on Huawei chips

[frosty lava]

AND overpriced

[minor bloom]

There is literally no way to know this

[frosty lava]

but we already know about the price

[wicked talon]

Correct but I believe it's going to be good but everyone is over hyping it so much

I mean wasn't 4.7 a downgrade from 4.6?

[frosty lava]

who want a good model that you can only use 2 time a month !

[wicked talon]

True

[frosty lava]

cause of the price and usage

[wicked talon]

Deepseek will probably smash qwen

And Kimi

Kimi doesn't have image generation

[minor bloom]

Claude is unusable without paying

[inner relic]

https://x.com/deepseek_ai/status/2047516922263285776

DeepSeek (@deepseek_ai)

🚀 DeepSeek-V4 Preview is officially live & open-sourced! Welcome to the era of cost-effective 1M context length.

🔹 DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world's top closed-source models.
🔹 DeepSeek-V4-Flash: 284B total / 13B active params.

X · Apr 24, 2026, 3:24 AM

[minor bloom]

Even sonnet

1.6 T??????

Holy

[wicked talon]

Bigger then Gemini lmao

Grok is 2T though I think

[minor bloom]

No

Its 0.5 T

[rigid copper]

pretty sure once i hit it, I have to create another new chat :)

[minor bloom]

Grok 4.4 will be 1T when it comes out according to Elon

[wicked talon]

Bro why is deepseeks knowledge cut off may 2025

Fake V4?

[minor bloom]

I knew my blue whale wouldn't disappoint

[wicked talon]

No way deepseek kept this model from us for this long

[wicked talon]

It's disappointing me now

I gotta make some random bull htmls then try to make it code a speedtest server

[pseudo magnet]

damn

[fickle venture]

DEEPSEEK IS back hell yeah

[echo aurora]

That's the case if you'd like to continue uploading more files. But you should still be able to prompt.

[vernal raft]

Nsh

[pallid crypt]

deep seek v4 lets go

[vernal raft]

Idk

[wicked sage]

DEEPSEEEEEEEEEEEEEEEEEEEEK

YESSSSS!!!!!!!!!!

[stray aspen]

Deepseek v4 is so cool

[pallid crypt]

they scored worse then kimi tho sadly 😢

[wicked talon]

Never speak again please

NEVER

Leave this server

[inner relic]

nah bro

[minor bloom]

Its worse than kimi in coding

[inner relic]

i think deepseek v4 is master at roleplay

check

[wicked talon]

Roleplay 😛

[fickle venture]

@echo aurora what's the context limit is it 1M?

[stray aspen]

Nah it isnt

[minor bloom]

But it has a lot more knowledge than kimi

It doesn't look like it's reasoning improved

[wicked talon]

Yesss

[minor bloom]

From 3.2

[wicked talon]

They updated there website

[empty stump]

I wish it was multimodal but this is impressive

[wicked talon]

It is ain't it?

[stray aspen]

No we got scammef

But deepseek is so cool

[wicked talon]

what

what

[rigid copper]

It will more likely to hit if I do image generation in side-by-side mode

[wicked talon]

am I in rage bait right now

[stray aspen]

Deepseek is so cool

[echo aurora]

We typically will do w/e the default API setting is.

[heady kite]

How is Gemma 4 31B ahead of Kimi 2.5?? Big model size difference

[stray aspen]

Kimi 2.5 sucks

[echo aurora]

Guessing this is rate limit, but will check this Trace and keep you updated.

[stray aspen]

And its old

[wicked talon]

Kimi 2.6 out now

[inland quest]

Non thinking pro version over the thinking in leaderboard looks like mistake, especially if you look at it rating at all, cuz it looks lower than expected for new models.

[heady kite]

Also why is gpt-oss-120b reasoning: high not on there

at least for code arena

[wicked talon]

I'm gonna cry myself to sleep

Goodnight chat

[echo aurora]

I take that back, this is a caused by a bug that was flagged to the team earlier today.

[echo aurora]

🫂

[inland quest]

Deepseek v4 Pro on leaderboard performs nearly as non thinking GPT-5.4 LOL.
Wtf
And thinking version as gemini 3 flash?

[whole sundial]

i'm already disappointed by v4 pro, it fails some of my world knowledge test questions
one of them both glm 5.1 and kimi k2.6 gets right, another one can be correctly answered by grok, claude, gpt, gemini, old and new glm and kimi models, and even hy3, a brand new model from tencent that is the same size as v4 flash but yet it gets questions right that v4 pro (5 times the size) can't

[desert pendant]

gng what

[inland quest]

Having probably x3-4 size of parameters

[fickle venture]

It beats opus 4.6 Max even Opus 4.7 cuz it suck

https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro

[vernal raft]

I was expecting much more NGL

[stray aspen]

Deepseek v4 is so ass

And no vision

[desert pendant]

BRO

YALL ARE NOT LYING

😭 I JUST WOKE UP RN

[stray aspen]

It sucks anyway

[balmy mist]

is deepseek out?

[stray aspen]

Even qwen is better

[whole sundial]

i agree, i would rather use hy3 preview than either v4 flash or pro

[inner relic]

ok but I think deepsee kv4 peak at roleplay

creative writting

[feral kernel]

what up with this?

[desert pendant]

gng

i know what i will use it

[desert pendant]

arena gang when they see a new model for 0.1 sec:

[feral kernel]

damn

[whole sundial]

tencent hunyuan completely re-did everything with a guy from openai and their hy3 model is pretty good for its size (around the same size as v4-flash), same company that made the awful hunyuan dense models, the 4b has the world knowledge of smollm2-360m with awful reasoning and tool-calling, it came out in the middle of last year
if they made a 1t+ model it would smoke v4-pro and every other open model

[balmy mist]

ughhh dont tell me its another bust

[whole sundial]

i really hope arena adds that model, it would probably be top 10 open

[desert pendant]

gng i feel so happy

time to test deep seek

[dusk dragon]

So uh what model are we looking at right now that's the best for world knowledge. Thinking Gemini 3.1 still?

[stray aspen]

what in the disappointment is deepseek doing

[dusk dragon]

Yeah, why did deepseek drop v4

Isn't it kind of bad timing

[whole sundial]

yes

[inner relic]

they do lot of experiment

deepseek v4

[gleaming wraith]

What was deepseek v4 called before it was revealed what it is?

[dusk dragon]

Deepseek v4 though is not like v4 performance

[inner relic]

so eh they're focused on writting and code

[dusk dragon]

It's like 3.5

Also anyone know the best place to get really good usage with Gemini 3.1 pro anymore

[desert pendant]

OH YEAH

[velvet furnace]

gpt5.5 is good?

[desert pendant]

happ

[feral kernel]

yep

[soft river]

Is DeepSeek on the app?

[stray aspen]

guys deepseek is NOT back at it

woke up and turned on my pc at 10 pm just to be disappointed

great

[inner relic]

bro

are you only focused on code?

[desert pendant]

gngImtired

[dusk dragon]

Gemini just needs to drop Gemini 3.5 and just destroy every model like the goat it is

[stray aspen]

yeah and math

[desert pendant]

if i get a good code bro

[inner relic]

ask deepseek to solve a impossible math

chinese are good at mathh

[stray aspen]

it just butchered my code bruh

[desert pendant]

someone who actually use roblox to also test deep seek or another AI

great

; D

im risking a good code for deep seek v4 code guys

[balmy mist]

how is deepseek?

was it worth the wait?

[barren sable]

there were some rumors/drama that the deepmind guys use claude at work rather than any internal models/checkpoints. Doesn't seem great for google.

[stray aspen]

no

it sucks

[balmy mist]

lol

[inner relic]

i think uh

it's decent

writting is good

[stray aspen]

i wanna see where artificial analysis will place it on their benchmark

[desert pendant]

2 errors

ehh

[inner relic]

and this guy think deepseek v4 sucks bc It can't do a lua code perfect

but yeh i agree with him

it sucks at code

[stray aspen]

nah its just bad

[inner relic]

sometime

Mimo v2.5 did better at one shot prompt

[desert pendant]

if Deep seek suck at the code now

[stray aspen]

mimo is actually decent

but nothing crazy

[balmy mist]

im going back to sleep smh, so its not the best open source?

[stray aspen]

no

[heady kite]

Did you guys mention that Deepseek is okay at writing or no?

[inner relic]

idk you go check for yourself

[stray aspen]

i just woke up in the middle of night to a trash release

[desert pendant]

should i use mimo if deep seek fails guys?

[stray aspen]

no use claude 4.7

[desert pendant]

happ

[inner relic]

wth bro claude 4.7 is so xpensive

yeh

use mimo v2.5

i think cheape

[stray aspen]

gemini 3.1 pro then

[desert pendant]

HOLY MOLY BRO

[inner relic]

rok

[balmy mist]

bruhh

[desert pendant]

I AIN'T RICH

[inner relic]

dont be disappointed

i think it's good at creative

and writting

[balmy mist]

thats not impressive tho

[desert pendant]

if deep seek fail me bro

[inner relic]

I already told you guys, deepseek is focused on

[stray aspen]

theres no way it was gapped by glm 5.1

[inner relic]

writting and code

[stray aspen]

thats like all the way down in artifiical analysis

[balmy mist]

we have models that are good at that for free

[inner relic]

everything is code sop now?

[desert pendant]

HE DID IT

[stray aspen]

at least we got 1 million context

[desert pendant]

IN THE THIRD ATTEMPT

[balmy mist]

i just dont see the point of this launch

[desert pendant]

LETS GOOO

[balmy mist]

??

[desert pendant]

deep seek did it

in the third attempt

[inner relic]

does this mean, deepseek adapt to error

each attempt

[desert pendant]

he always

did that

now i will use roblox (cuz literally im using godot right now)

[stray aspen]

deepseek front end is so bad

[inner relic]

try in three attempt

[balmy mist]

i guess its just the 1 mill context that we care about?

[earnest rover]

deepseek v4 is one of the best **||overhyped ||**model

[stray aspen]

cap

its 5.5

[inner relic]

wth it's Deepseek v4

[balmy mist]

fr they waited so long for this lol

[stray aspen]

im gonna try mimo

2.5 pro

seems like its insanely decent

[stray aspen]

the spud

they tricked us into thinking it was the spud

they love marketing campaigns bruh

[balmy mist]

this is interesting: https://x.com/ValsAI/status/2047513613750202452

Vals AI (@ValsAI)

DeepSeek v4 is now the #1 open-weight model on our Vibe Code Benchmark, and it’s not close.

It leaves the #2 (Kimi K2.6) in the dust, and even beats out frontier closed source models like Gemini 3.1 Pro.

X · Apr 24, 2026, 3:11 AM

thats a bigg gap tbh

[desert pendant]

to be honest

[stray aspen]

slopmark

[desert pendant]

for me deep seek is doing great (yet)

[vernal raft]

Mythos

[balmy mist]

you dont trust that team?

[stray aspen]

not bad for a first shot

needs some fllow up prompts

its better than gemini

and i like that little detail of adding the server location

[heady kite]

How many tokens did it output for that?

[desert pendant]

trollface

to be honest im kinda dissapointed with some deep 4.4/4v stuff

[stray aspen]

is this thing fr

im never using it again

it gave me html

[desert pendant]

be specific dude

[inner relic]

can claude do this

claude sonnet 4.6 dumb as hea

[stray aspen]

guys did mimo cook

[desert pendant]

ok this is great

it just need some adjustments

[stray aspen]

mimo is better than deepseek lmao

[balmy mist]

dont you dare say that

deepseek is my friend

[stray aspen]

its better for frontend

[balmy mist]

what is deepseeek better for?

[stray aspen]

for feeling disappointed about interrupting your sleep for a slop release

[balmy mist]

lmaoo that was a good one, imm go to sleep on that note

[inner relic]

ask mimo to mae

make

advanced npc ai shooter

and i said

can claude solve this

br

bro

[desert pendant]

mimo easily go past 300 lines from code

dawg

happ

[thick pawn]

Anyone else having trouble with gpt image 2 not completing jobs at the moment?

[desert pendant]

ok mimo is kinda great

[wicked talon]

Reddit should definitely make an ai app

[night moat]

It's been more than six hours

[thick pawn]

Yup, image 2 is definitely playing up at the moment

[sly cedar]

Deepseek seems on logic god mode

I used this model, and its cooking

Far better than the previous model i used

[wicked talon]

Deepseek is taking over?

[vale quest]

God damn deepseek

Deepseek basically pulled a meta

Wdym

[inner relic]

lmarena users

can you stop

yapping

[wicked talon]

Deepseek v4 is literally benchmaxxing

[inner relic]

and focus on someone else

[wicked talon]

First of all we ain't lmarena bro

Second of all leave

[inner relic]

Rude

[wicked talon]

You were rude

[inner relic]

though buddy you're too focused on code

[wicked talon]

Mate leave no one wants you here

We are here to talk about ai

[inner relic]

uh ok

this is general

i dont want to argue with you eh

[wicked talon]

Bruh what's this servers main purpose

[inner relic]

just yall to acknowledge that deepseek is smart at some task

not just code

[wicked talon]

It's good at code too lol

[inner relic]

ok ok

[inner relic]

yeah but mimo 2.5 looks better at code than deepseek

[wicked talon]

Lmao that's like saying Kimi is better than opus at code

[inner relic]

yeah bro stop being ignornant

and you're still rude even I didnt want t argue with you

[wicked talon]

If you have a problem tell a mod

[vale quest]

Well idk abt that

Deepseek is better at debugging and deep tasks

Mimo is good at structured small tasks

And being fast

[loud herald]

It is

[velvet furnace]

can we use gpt5.5 in battle?

[surreal zephyr]

5.5 mogs all

[sly cedar]

For me deepseek is better at following instruction context and debugging which is quite amazing

[loud herald]

If 5.5 was here we'd have a ping

[vale quest]

Never used deepseek i just see reviews

Glad it worked out for you

[surreal zephyr]

At code security?
5.5>5.4>5.3>5.2>opus 4.5> opus 4.6 >>> opus 4.7 >>>>> gemini 3.1

[surreal zephyr]

(Actual tests, prepared& reviewed by all 3 models together, anonymously)

[sly cedar]

I reviewed it, its like gemini 3.1 pro but on steroids on context & debugging or even creating elements, i'm glad i can use gemini 3.1 pro but on steroids

I used it for roblox project

[vale quest]

Nicee

[sly cedar]

Honestly i used to make roblox game with old deepseek model, but one of the generated scripts actually creates bold ui design, and i like it, but back then deepseek was infant

it can't even handle codes pretty well back then imo

Deepseek had big glow up rn

[sullen creek]

yo do u guys think they are adding gpt 5.5 to direct

[sly cedar]

If its cheap = yes, elseif its expensive = yes > later remove

[river moat]

How now how to use Gemini 3.1 pro for free

[sly cedar]

Deepseek v4 for now

Imo

or Kimi 2.6 atleast

[vale shell]

hello. Will Gemini 3.1 pro be available on arena ai?

[sullen creek]

glm 5.1 is closest to opus rn

[sly cedar]

https://tenor.com/view/dj-gif-24173332

[surreal zephyr]

Hows deepshit v4

[sly cedar]

Deepshit v4 cooked my code pretty well

[wicked talon]

Ai studio

[wicked talon]

I've cried over it 🙂

[sly cedar]

Deepshit is now deepreal

[wicked talon]

Yesss

[grave peak]

How its deep?

[sly cedar]

Because it can thinks deep

https://tenor.com/view/thinking-of-you-christopher-walken-considering-it-deep-thinking-deep-thoughts-gif-4077808712528826392

[toxic whale]

Deepseek v4 pro does about same performace as Sonnet 4.6 Thinking, or Opus 4.6 on like Low in my testing

[river moat]

How to use DeepSeek v4?

[karmic temple]

Hi 👋

[river moat]

Hi

[karmic temple]

Can you make a video of the picture I uploaded above?

[river moat]

This?

[shrewd citrus]

woah v4 is here

[karmic temple]

Yes

[river moat]

Okay

What’s you need?

For a video

[karmic temple]

Yes, I really need the view of our village. This is the school.

[river moat]

Describe what is needed?

[karmic temple]

The video will be drone style slow motion.

[river moat]

Ok bro 1 second

[karmic temple]

You don't understand Bengali.

[river moat]

Sorry no

[karmic temple]

When will my video be made?

না তুমি পারবে না তৈরি করে দিতে

[whole sundial]

Video Arena is only available on https://arena.ai/video, if that is what you are here for.

[river moat]

It’s generating

Sorry

I can’t

[toxic prawn]

/ image-to-video I want to the this movie chale and play ho

[brisk turret]

Deepseek launch looks like a dud but when you factor in cost, it's a killer

[wicked talon]

It's bad

[toxic prawn]

/cinematic slow zoom, 4 friends watching movie in dark theatre, screen light flickering on faces, dramatic mood, realistic camera movement generate this video

[half veldt]

is this a bot?

goofy ah bot

[spring oar]

GPT 5.5 better than opus 4.7 ?

[knotty fable]

And no new version number on Seedream, but they've done something, more responsive prompting and better results.
My bet that it's a hidden update to counter GPT2.

Which one is better? Matter of taste - but it really need a goal photo to tell.

Tudi & Seedream left, and GPT2 right.

Funny thing is that while Seedream have added fake noise to make images more "photographic", it's seen on the GPT2 image.
While the same noise now is much smaller on Seedream at left - and I've done a dozen in the last hour to get her pose and dress right so it is consistent.

[jaunty dawn]

Can a model's name be changed? I noticed in a chat that a model previously called deepseek-v3.2 was renamed to deepseek-v4-pro

[civic plaza]

Is the website down? It’s using forever to load

[wicked talon]

V4 is the new model

They took 3.2 away

[compact flame]

Hey chat

How good is gpt 5.5 after testing

For me it seems great

[surreal zephyr]

haha i wish

atleast those were removed and replaced with

[sterile tartan]

What's this

[surreal zephyr]

deepseek v4 is actually deepseek v3.2-experimental

🤣

[sterile tartan]

💀 💀 💀

U Sure

[surreal zephyr]

yeah i am

v4 flash is v3.2 exp

literally says

[sterile tartan]

It says that fir API Replacement doesn't it?

[compact flame]

Hey pro

[surreal zephyr]

it says renamed 3.2-exp to 4-flash

but 4-pro is new

[compact flame]

How good is gpt 5.5?

[surreal zephyr]

best

🔥

[sterile tartan]

Well atleast pro is new

[sterile tartan]

Really Good

[compact flame]

Seriously?

[surreal zephyr]

yes basically they were testing 4 flash under name of 3.2 exp

or

pr move

not me to know

¯_(ツ)_/¯

[compact flame]

I guess chatgpt finally beaten Claude after all these months

[surreal zephyr]

it was always better at thinking just needed more specific prompts, and it wasnt best at ui

[orchid olive]

when can I see the V4 or 5.5

[compact flame]

Well when a crab whistles on the mountain

But deepseek v4 is there tho

[sterile tartan]

@surreal zephyr exactly which models are available on Deepseek Web/App?

[sterile tartan]

Because it's cheaper

[compact flame]

Why do you correct me

I know it's cheap

[sterile tartan]

Well u didn't said that before

[brisk turret]

"Price is blended using a 3:1 output-to-input ratio: (3 × output price + 1 × input price) ÷ 4. This reflects typical usage where output tokens cost more and are generated in higher volume."

Petition to add a slider to the pareto graph for input:output ratio

[sterile tartan]

Just incase you didn't knew

[compact flame]

Kk

[sterile tartan]

K

[surreal zephyr]

no idea i dont use deepshit

i need multimodality

[sterile tartan]

Ufff

[surreal zephyr]

models without proper multimodal reasoning are unreliable imo

total cost efficiency 5.5 vs 5.4

5.5 up to 7x more token efficient is wild

so up to 3.5x cheaper

[sterile tartan]

No

Because it's more expensive

[surreal zephyr]

2x higher cost per token yes
but uses 7x less tokens

[sterile tartan]

Is doubled

[surreal zephyr]

so 3.5x cheaper total

[sterile tartan]

Wait

[surreal zephyr]

(just dont spam xhigh when medium and high do fine then you can save 3x the quota)

[sterile tartan]

You are Absolutely Right

Very Sigma Calculation Bro

[brisk turret]

where 5.5

[vernal raft]

In battle mode

[river moat]

What ai is the best for a school

What is this

[light sleet]

pineapple will get u 😈

[light sleet]

bro is roblox exploiter 👍🏼 👍🏼

[surreal zephyr]

uhhh

[light sleet]

or wait

[light sleet]

u in punkteam because of vpn 😡

same bro same

[sterile tartan]

https://tenor.com/view/shut-up-dont-talk-close-mouth-grab-lips-mad-family-gif-21732788

[surreal zephyr]

uhh yeah totally

🤔

[river moat]

Who know what ai is the best for school

[shrewd citrus]

gpt 5.2 search

[river moat]

Thank

[river moat]

And coding

I forgot say

[restive charm]

This problem can be solved
This session has reached its token usage limit. Please start a new chat to continue.
Trace ID: 76f18173-373d

[tidal sierra]

from what point in tim

can an ai

[flint sandal]

lets go

[tidal sierra]

just cancel a chat

[tranquil burrow]

your

[robust marsh]

pls fix ur fckin captcha😭

[flint sandal]

lets see if the $200 pro subscription was worth it, and yeah im releasing this game on steam to get my $200 + tax back😭

but i heard 5.5 pro is really good at game-making

i will use like meshy ai and add real 3d models to the game

and see what will happen then

half way there

[astral cobalt]

arena ai actually crash

[flint sandal]

extended pro😭

[restive charm]

Look at this problem

[flint sandal]

the results are interesting but does someone have a great pc to run it? because on my m2 mac it runs at 5fps😭 please

[waxen seal]

https://youtu.be/K6eSZdl6w7Q?si=aVtXBFnq8f8Lm45O

[light sleet]

Send

[surreal zephyr]

thats why i use gpt

gpt has no "end conversation"

so it actually listens to you instead of ending himself when hes lazy

[flint sandal]

check dm

or

anyone?

[rose tendon]

gemini 3.1 flash lite preview .. is it a temporary issue or ?

[light sleet]

nvm I'm lagging hard too

[bronze abyss]

guys wasn't claude opus 4.6 available in the LLM chat what happened to it?

[flint sandal]

its too expensive to give free access to opus to everyone with generous limits

its still in battle tho

[bronze abyss]

ooooh
thanks

also have anyone tried chinese models like kimi?
if so what's your review about it

[flint sandal]

new deepseek is trash, kimi-k2.6 is the smartest one in my opinion in how it reasons in every domain, and glm-5.1 is okay but i dont like glm cause it has the same problem that gemini has like if you tell it to change one button it will additionaly change the full page, even if i ask it to not do that.

[bronze abyss]

thx

[flint sandal]

but i recommend to pay for claude or chatgpt, they are way better

i just bought cgpt pro for 200 bucks u know😭

flex

because paying for glm or for kimi that arent SoTA is i think a waste of money

[bronze abyss]

i dontdo complex tasks I'm not even into coding I'm a med student

[flint sandal]

gpt-5.5/gemini 3.1 will be great for you

[strong ferry]

Ngl, so far GPT 2 has been pretty impressive. I asked for this prompt:

Create an illustration showcasing details about the differences between Bigfoot and the Abominable Snowman. On Bigfoot's side, it describes it as being either male or female, brown fur and more man-like in its face, looking almost like a Neanderthal. It is aggressive only when provoked and can be found in the woods of America. On the Abominable Snowman's side, it is mostly a male species with white fur and a more ape-like face, bipedal with large feet like Bigfoot, and is less aggressive. It is a creature that prefers solitude and is known to save some of those who wander in the blizzard in the Himalayas. Some theories suggest it may be a Tulpa created by the Tibetians.

And it's shockingly good with the text. Even Gemini struggled when you asked for too much. This is consistent.

And here's a map I asked for my fictional island

[bronze abyss]

yes and no
i like to use claude analysis level for the deep content and also his special commands like oods and L99 actually make difference for me

[flint sandal]

yeahh i recommend you to buy claude max for 10 bucks monthly, look at the promo codes online and u will be happy with that

[flint sandal]

YEAAHHH NOW PUT THAT TO 5.5 PRO AND MAKE AN AAA GAME OUT OF IT😋

wtf

[surreal zephyr]

codex made factorio copy and now playing it

[bronze abyss]

transactions here and complicated in africa
also 10 bucks is waaaay more than it worth here

[strong ferry]

I wish lmao

[flint sandal]

nahh 10 bucks for this much usage is really good, in api you would pay like 150 bucks for this much usage

[flint sandal]

plus you have all the special features

can someone mute him please?

[bronze abyss]

I'll look into that

[surreal zephyr]

<@&1349916362595635286>

thanks

[flint sandal]

wowww pretty fast moderation here

[surreal zephyr]

🤖

faster than ai

🔥

[grave peak]

Fair enough

[surreal zephyr]

tbh ai moderation here would be peak

auto delete scams & video requests

[rose tendon]

any takes on this ? it's really hard for me to start a new chat at this point

[flint sandal]

AND BTW WHERE IS SORA I BOUGHT PRO AND NO SORA HERE? ://

[tidal sluice]

Been using Deepseek V4 for a while and it doesn’t improve. After a few exchanges, it loses memory. When I point it out, it doesn’t even remember forgetting, so things get messy. Eventually it only recalls the very first question, so it’ll hit me with “So what you meant is this!” — bringing up ancient history even though we’ve moved on.

[stray aspen]

It sucks

I guess ill have to switch to mimo 2.5 pro

[knotty fable]

https://youtu.be/_193U2aNaeE
This is bloody nutz.

[surreal zephyr]

Deepseek v4 being worse than kimi2.6, gpt 5.4 is just funny

[split topaz]

Hey..

[surreal zephyr]

Deepseek v4 is actually v3.2 exp

😂

[light sleet]

💀

yet they said it's gonna beat 5.5

lol

[grand raft]

what??????????????????????????????????????????????????????/

[light sleet]

where are the deepsleepers?

[grand raft]

https://tenor.com/view/spongebob-you-what-what-did-you-do-spongebob-meme-spongebob-squarepants-gif-19389913

[split topaz]

API of V 3.2 got recently updated so they might have been shadow releasing for a while.

But yeah it's not that useful looking at the benchmarks. I can only hope that this being a preview would signal improvements later on

[earnest rover]

so anyone knows whats the rl for gpt image 2 in chatgpt for free users (OFC)

[storm dust]

yo guys

did you witness the kimi logo redesign?

it looks different now

[compact flame]

Why do you ask

The API is not even out yet bro

[tulip parcel]

How’s gpt 5.5?

[flint sandal]

i am testing 5.5 pro and im quite pleased

but nothing revolutionary

just like a gpt-5.3/5.4 situation

[proud bobcat]

The whale has awoken.

[proud bobcat]

Well the point is is that it’s a reliable workhorse

It’s not supposed to be super duper ultra intelligent

And for what it is it’s an extremely competitive model

[flint sandal]

i would rather use qwen 27b than the new deepseek

tbh

[proud bobcat]

How come

[flint sandal]

deepseek seems to have the gemini issues

and glm issues

qwen 27b doesnt

[proud bobcat]

That’s

That’s a very broad statement

What are these issues

[flint sandal]

.

[dusky hedge]

Have you guys found a way to use claude opus for free?

[proud bobcat]

Again DeepSeek is meant for good, fast intelligence

It’s not supposed to be SOTA

[flint sandal]

its not good or fast

faster than other open chineese models ye

[proud bobcat]

I’d beg to differ?

It’s faster than Gemini 3 flash for me and Claude sonnet 4.6

[flint sandal]

whats ur provider

[proud bobcat]

I use the app and openrouter

It’s been quite nice

DeepSeek has NEVER let me down any time I’ve asked it something

The one time it did was because I didn’t describe something correctly

Which is insanely impressive for a lower tier model

[flint sandal]

i mean flash is good as the fast cheap model

but pro is supposed to be good and SoTA like thats the point of pro

[proud bobcat]

We will have to see its intelligence score

It’ll probably be equal to muse spark

[flint sandal]

i would rather have a really slow model that is SoTA and is good

[proud bobcat]

That’s your preference then

Nothing wrong with that at all

I personally fave Kimi K2.5 and K2.6

[flint sandal]

but still with fast models that arent that good you spend more time fixing and iterating so slower models are actually faster to work with

from my experience

[proud bobcat]

It’s more about how it thinks

For example I can tell you DeepSeek will always provide you decent code

It may not be opus quality

But it will work

[wispy light]

why am i not able to login in lm arena website

[proud bobcat]

Every time

[surreal zephyr]

Deepseek v3.2-exp*

Renamind model is wild

Deepseek geniuely has most overhyped open source models while having worst ones

Qwen has best models by far from open

Qwen 3.6 27b solos deepseek v3.2exp aka v4

And if you need price to perf then gpt 5.5 is still best

[indigo knoll]

Is Deepseek 4 all that? Or just overrated?

[limber hound]

we'll see actual results in few weeks just as it always happens

[void shore]

They benchmark maxed it

So it seems good on paper

But people are saying it isn’t the greatest when it comes to programming tasks

[limber hound]

tbh not total disappointment, wanna test it on few thousand hundreds context

[pastel ember]

The only benchmark worth trusting is arc-agi, the rest is just benchmaxxing and pattern matching. If DeepSeek doesn’t at least hit gemini 3 flash level on arc-agi, it’s a flop. At that price, nobody’s gonna want it.

[void shore]

I’ll test it

[limber hound]

Engram sounded so promising

[void shore]

And see what happens

[indigo knoll]

Is Gemini 3 Flash still the best non thinking model rn?

[limber hound]

Due to constraints in high-end compute capacity, the current service capacity for Pro is very limited. After the 950 supernodes are launched at scale in the second half of this year, the price of Pro is expected to be reduced significantly

[pastel ember]

Yep. Especially in terms of price and intelligence.

[limber hound]

hopefully that's true

[indigo knoll]

Second is GPT 5.3 or what?

[pastel ember]

I’d say GPT-5.2, but that’s already much more expensive than Gemini 3 Flash.

[indigo knoll]

What, so 5.3 which is a newer version is worse than 5.2?

[pastel ember]

Maybe GPT 5.4 Mini, but I haven’t tried it yet.

[tranquil burrow]

Guys am i able to ask y'all a question?? When will you be able to use ChatGPT Image 2.0 ? I know its in the Leaderboard but we cannot use it yet (as of my knowledge)

[stray aspen]

mimo 2.5 pro is so good

[pastel ember]

Just open https://arcprize.org/leaderboard and check out the price/intelligence. It shows how well a model can actually think. Chinese models, as expected, are all at the bottom.

[stray aspen]

mimo 2.5 pro is great for front end

[proud bobcat]

Respectfully

I have actually never seen a lower iq argument here

You do realize V4 is a completely new dataset

[stray aspen]

it sucks tho

[proud bobcat]

They updated the models with the new weights

And removed the old ones

[stray aspen]

mimo is way better

[proud bobcat]

Oh my god bruh for the last time DeepSeek isn’t supposed to be SOTA

It’s the reliable workhorse

[stray aspen]

its supposed to be the poor man's 5.5

[proud bobcat]

In benchmarks DeepSeek V4 outperforms 5.4 xhigh pretty often

I don’t know why that means it sucks

[stray aspen]

it wont outperform the spud tho

[proud bobcat]

ITS NOT SUPPOSED TO 😭

If you want SOTA you go for Kimi, Claude, GLM

DeepSeek is for rapid deployment

[stray aspen]

kimi is NOT SotA gang

[proud bobcat]

My guy what.

[stray aspen]

mimo 2.5 pro is

[proud bobcat]

???????

You used it for one prompt and said like

“Yeah this is SOTA”

[stray aspen]

nah

i tested it yesterday

[proud bobcat]

Mimo is the exact same philosophy as DeepSeek

[stray aspen]

but its smarter

and does stuff correctly

[proud bobcat]

Ehhh in my testing not really

[stray aspen]

and we get 1 million context

but its actually smart

and gives you complete coding projects

not just a 100 line template like gemini n stuff

[proud bobcat]

Well in Gemini’s defense here it’s always been a pretty ass model

I just like DeepSeek because it’s reliable

[stray aspen]

deepseek needs vision

[proud bobcat]

Yeah

Multimodal coming soon

As per their post

[stray aspen]

send me

[proud bobcat]

Hold

[stray aspen]

if it gets vision its better than mimo

[stray aspen]

great

[proud bobcat]

Again you can prefer what you want

But I think a lot of people conflate that new model must be SOTA

I personally love V4

[ionic vigil]

I love that it doesn't run inference on nvidia

[compact comet]

it's insane how deepseek just keeps being nerfed intentionally and it still manages to perform near SOTA

[frosty lava]

then no, if you look for price to performance open source model or small team model are the best

[stray aspen]

how do you konw they are nerfing it

[frosty lava]

Deepseek is working the hardest on architecture improvement

that's definitly true

[compact comet]

they literally are not allowed to use nvidia

[frosty lava]

and they say what they achieved and innovated publicly

so everyone can technically replicate it

[compact comet]

they have the best engineers in the world probably

no questions

[stray aspen]

why

[frosty lava]

it's profitable also for other ai companies they will just use those technics

[stray aspen]

i think claude does

[compact comet]

blud asked why

[proud bobcat]

Wait

[frosty lava]

i don't think its the best in the world but i think other ai companies just doesn't care and are just scaling

with compute power

[proud bobcat]

DeepSeek v4 pro BEATS Kimi K2.6 in swebench verified???

[stray aspen]

kimi is ass bro

anything beats it

[proud bobcat]

All your opinions are dogwater bro

[stray aspen]

they aint

im just saying the truth

[proud bobcat]

Kimi wipes the floor with Claude opus 4.7

It’s not always about benchmarks

Bros laughing while opus 4.7 won’t even read documents, listen to instructions, and takes shortcuts always

Not the mention the stealth price hike with the new tokenizer leading to 35% higher costs for an already expensive ahh model

So you’re getting worse performance with Claude while paying more premium

[proud bobcat]

DeepSeek’s architecture is actually insane

Running a 1.6T parameter model at such a fast speed?

Holy

[frosty lava]

honestly they will keep going like that and keep reducing compute power necessity, faster training, and it'll just be profitable to everyone

[stray aspen]

@proud bobcatare you running deepseek v4 locally

[frosty lava]

other ai companies will just steal the idea to implement on their but its normal honestly

[sonic wigeon]

how's deepseek guys
anyone tried it

[stray aspen]

its mid

bad for frontend

[sonic wigeon]

hmm

[stray aspen]

and its bad for Lua coding

but its great for math

[frosty lava]

but deepseek doing the dirty work for architectural improvement

[sonic wigeon]

its not multimodal either eh?

[stray aspen]

not yet

but the ywill add vision later

[sonic wigeon]

interesting

[proud bobcat]

Oh yeah dude I have like 3 servers laying around just for this

Totally

[stray aspen]

its not that expensive lmao

[proud bobcat]

I’m not wasting money to run local ai

I don’t need it

I just like keeping up with releases and benchmarks

[frosty lava]

maybe if at some point we will be able to compress the model so much (like 99%) without loosing quality we will be able to run T model locally lol

[frosty lava]

to run a 1.6t model locally and at decent speed ?

it's expensive bro

[rocky geyser]

;-;

[stray aspen]

not more than 15 k tho

[frosty lava]

yes so its expensive for most people

[rocky geyser]

15k if your lucky and want slow speeds and thats still expensive- 😭

[stray aspen]

wdym thats like just 10 months of work

[rocky geyser]

without food, rent or anything else

[frosty lava]

bro your ragebaiting honestly

[proud bobcat]

It’s only 15k guys

[stray aspen]

yes

[frosty lava]

you can't save 100% of what you get

[sonic wigeon]

15k is the avg yearly salary of the top 5% middle class people where i live

[frosty lava]

anyway

[stray aspen]

im not dude

its the truth

unless you live in some third world country

[proud bobcat]

I’d rather buy me and my future husband a cottage somewhere than use that money for ai slop

[stray aspen]

wdym ai slop

[sonic wigeon]

"some third world country"
that is literally the entire world bro.
wake up.

[stray aspen]

its not

[sonic wigeon]

europe and NA is just a small part

to dismiss 4-5 billion people like that is a crime

[stray aspen]

working at mcdonalds in canada can get you more money than other countries

[frosty lava]

your ragebaiting cause you know its impossible to save 100% of the money you get every month

so it won't be 10 month

[proud bobcat]

As much as I nerd out about AI I’m never using that for anything serious

[frosty lava]

but much more

[proud bobcat]

I’d rather spend 15K on something useful

A used car I can drive

[sonic wigeon]

97% of people cannot do that.
the ones that could and are are getting kicked out or constantly racist-ed

[proud bobcat]

An audio setup

[sonic wigeon]

either way we're getting off topic

[stray aspen]

guys lets stop talking about this before the night fury warns us

[proud bobcat]

Point is is that llama 4 maverick is the best model and you’re all wrong

😎

[soft river]

So sad that the new model isn’t in the web/app yet

[proud bobcat]

DeepSeek?

[stray aspen]

it is wdym

[proud bobcat]

It’s been out for a good month I’d reckon

[soft river]

No it’s not

[stray aspen]

what model are you talking about

[soft river]

It’s DeepSeek v3.2

[stray aspen]

we have v4 gang

[soft river]

Not 4 yet

[proud bobcat]

https://tenor.com/view/oh-my-god-bruh-jjk-jujutsu-kaisen-higuruma-higuruma-hiromi-gif-6772381213627884903

Dude.

It’s been V4

[soft river]

Bruh it’s not 😂

[proud bobcat]

For a month

[soft river]

Not at all

[proud bobcat]

Jesus Christ.

[stray aspen]

what in the ragebait are you doing

[proud bobcat]

I think ai might be giving us brain atrophy

Genuinely

People will open the DeepSeek app and see “instant” and “expert” modes and still say ts

Just ask DeepSeek what it’s knowledge cutoff is

[soft river]

“Glazer” and you can’t differentiate them

Crazy

[proud bobcat]

Ts is ragebait

The first one is V4 flash

The second is V4 pro

What is there not to get

It’s been like this for a month my dude

Today it got released for api access

The ragebait is INSANE

[fiery gull]

v4 is good, but the qwen 3.6 27b bruh, how this run in my pc?

[proud bobcat]

The power of dense models

[soft river]

Nuh uh

[proud bobcat]

I can’t wait till we get a good 32B dense model from Qwen

[fiery gull]

waiting my 3.6 2b

[proud bobcat]

https://tenor.com/view/off-with-his-head-cat-cat-meme-gif-2379878183170633577

[soft river]

That was only a change in the interference

[fiery gull]

my pc only run the 2b in 15t/s ;-;

[tired mantle]

Excuse me, where is GPT 5.5 on Arena? What's the name of the model there?

[fiery gull]

the 27b is only 4t/s ;-;

[proud bobcat]

Deadass?

[surreal zephyr]

v4 is same model as v3.2 it was renamed lmao

[proud bobcat]

They changed the weights my dude

They deprecated 3.2

[surreal zephyr]

nah they made v4 pro (thats actual new one)

[fiery gull]

I tested the v4, is more fast and has more context

[surreal zephyr]

but they renamed 3.2 to 4flash

[proud bobcat]

4 flash is a completely diff model

It’s 285B parameters

[fiery gull]

flash v4 is another model

[proud bobcat]

3.2 exp was 671B

[surreal zephyr]

bro it kept the score on arena

[fiery gull]

but really like the v3.2

[surreal zephyr]

its NOT a new model

[fiery gull]

bruh, Idk about lmarena

I'm talking about my expecience in deepseek.com

[proud bobcat]

Yeah

Also

[split topaz]

According to official benchmarks Deepseek V4 Pro scores 154 points MORE in comparison to Claude Mythos in codeforces rating. Only 3.5 points behind Mythos in BrowseComp, strange.

[proud bobcat]

It’s right here cuh

[grim cliff]

Is that a new AI?

[stray aspen]

@proud bobcatwhy does mimo 2.5 pro thinking process look similar to deepseek's

[surreal zephyr]

ctrl+shift+r

[proud bobcat]

They learn from the best

😎✌️

[proud bobcat]

I just checked this

Right now

[split topaz]

That is my source.

[surreal zephyr]

worse than qwen 3.6 27b

xD

[proud bobcat]

It’s not SOTA but I’m loving it for math work and number crunching

It’s so peak

[fiery gull]

yep ;-;

[stray aspen]

mythos is cookign

they will destroy the spud

[fiery gull]

qwen 3.6 27b > v4 pro

in my docs tests

27b > max

lol

[proud bobcat]

How so?

[grim cliff]

Why is it so bad?

[surreal zephyr]

is the mythos in the room with us?

[fiery gull]

27b is better that max 3.6 🤣

[grim cliff]

I mean what model is good in like Science and creativity

[fiery gull]

qwen is horrible making big models

[proud bobcat]

Mythos glazers when they don’t even have access to the model and still hype it

[stray aspen]

it will be soon

aand it will ccrush the spud

[split topaz]

Probably yeah, also research.

[proud bobcat]

I

[grim cliff]

?

[grim cliff]

Can you name some models maybe

[proud bobcat]

What?

I was referencing the other dude