general

[frigid tusk]

like its the street i would see outside in 2025

[topaz epoch]

Best AI model for coding ( especially python ) ?

[frigid tusk]

should be claude opus 4,5 thinking 32k i guess

[weary galleon]

Obviously.

[echo aurora]

I hope so too. Our intention is to get these error rates as low as possible. I am sorry you're experiencing this.

[topaz epoch]

Appreciated your work ❤️

[topaz epoch]

I'm using that for python

[frigid tusk]

is it good? for you

[topaz epoch]

Better than other ones

[frigid tusk]

btw , idk if its me , but glm4.6 , when i use it on the actual website , its way better then claude and others , like it made me an actual smart ai compared to other , even a huge difference

[frigid tusk]

then continue using it till they drop new powerful a model

[topaz epoch]

Yeah

[pseudo inlet]

how good is this ai i need infinte photo sending

[half mist]

What the hell happened to the rpd of Gemini 2.5 flash and 2.5 flash lite? It used to be 250 and 500, and now they’re both 20.

[torn mantle]

flash 3 out yet or nah?

[half mist]

No, that’s the crazy part although 2.0 flash and 2.0 flash lite are gone for some reason

[cloud zinc]

they compute constrained

[half mist]

The worst part is that the Gemini api is what I’ve been using for my ai, but I keep getting a rate limit message when I just generating one response even after a few seconds, so I might need to pack my bags with this project, or somehow find another free ai api, but I doubt

[late bluff]

Claude and gemini 3 pro are nearly the best one. Generated my sonic clone game in best quality ever

[cloud zinc]

The goal of the free version is to get you to pay. It’s working as expected

[half mist]

I know, but making the rpd go from 250 to 20 is diabolical though

[half mist]

Well, I am not paying for AI since I don’t think it’s that worth it. I just need a free ai API, and Gemini was my go-to for that

[lusty tinsel]

its not rate limit tho

[late bluff]

Recommended to always pick those models in direct chat they will do the best for coding

[cloud zinc]

if u dont think its worth it, dont use it

[half mist]

But I am not saying it’s bad, it’s just not worth paying money for

[cloud zinc]

there are lot of free ai out there right now

[half mist]

Like what?

[cloud zinc]

open source

[torn mantle]

hmm

[half mist]

The original version of my ai did use an open source ai, but someone said that I needed to change it to Gemini, and that’s what I did

[cloud zinc]

u should move back to open source

[half mist]

I was thinking about that, but I don’t think that is ideal.

[ocean vortex]

it seems that it's still letting you go beyond that lol

[half mist]

It’s because it was at 250 before, but when the rate limit change came in, it’s just constantly blocking me since it’s way higher than 20

[cloud zinc]

nah its far away

[echo aurora]

It's possible that labs have a slightly different version on their site vs what they offer via their APIs which could be leading to a perceived difference.

[ocean vortex]

Dork5 not to be yet

[cloud zinc]

4.2 is not even out yet

[ocean vortex]

what's their website? Lemme see if they have sys instructions or smth 👀

[half mist]

@echo aurora Btw, if you pick both good or both bad on LMArena, what does the follow up response use as context. I know that the winning option is used as context, but does both good use both as well, and if it’s both bad, it doesn’t use it at all?

[torn mantle]

hehe

[ocean vortex]

Are they gonna do 4.2 even?

[cloud zinc]

3 weeks

[torn mantle]

yesh

[ocean vortex]

oh it's 420

[torn mantle]

2 to 3 weeks

[ocean vortex]

Elon wouldn't miss the opportunity lol

[torn mantle]

malware

@echo aurora

ban

[echo aurora]

That's a good question, I'm not sure. Will dig a bit and let you know.

[ocean vortex]

did a small test. For both options seems that you get branched into 2 different context windows. Both are saved, for left and different context for right

Kinda odd to save "bad" responses, but looks they are preserved

Elon though and 420

I think he already implied earlier there's gonna be 4.20 (hence 420) iirc

this was the most low-key img I could find quickly 👀

[echo aurora]

Okay got some clarification. I'm glad you asked as I wasn't aware of this. If you vote for either a/b as a preference then that context is retained for the prefrred vote. If you vote both good/bad then both are retained. However, they're retained for each corresponding prompt.

For example, Model X and Model Y are selected for both are good. For new prompts Model A will retain context from Model X, and Model B will retain context from Model Y.

[half mist]

Feel like the both bad one should just not use them at all in my opinion

[hollow ivy]

new acronym for the beast coding model: coast (CO45T)

[inner gate]

Difference between image 1.5 and latest?

[visual osprey]

gpt-5.2-high rank 27

[weary galleon]

It should be even lower.

13 place!!!! Hahahahaha

🤣

Even GPT-5.1 outperforms GPT-5.2

According to LMArena

As I said earlier, it was designed for benchmarks only, not for real tasks. LMArena Text leaderboard shows results in REAL tasks of REAL people.

[fickle venture]

Wdym

[weary galleon]

Even Sonnet outperforms it, even without thinking. I don't say about Opus. Opus is other level.

I'm satisfied. It's night in my city now, and I am going to sleep VERY happy 😁

I hope to get great dream today about LLMs.

[modest prism]

If grok 4.20 became disappointing I might be actually convinced that LLMs have hit a wall

[weary galleon]

No. If scammers scam you, that doesn't mean good people will too. Anthropic and and Google always will be the top 🏆

[hardy lion]

gpt-5.2-high is #1 on math, but I guess people don't like it as much on a lot of other things 🤷‍♂️

[warm zodiac]

it can be terse (in a good way imo) but I reckon people don't love that

[inner gate]

theres 2 models i think one is gpt image 1.5 and one is image latest

[weary galleon]

Interesting choice, math vs everything else

[hardy lion]

well that's the one thing it is #1 at, of course there is a spectrum. Coding it's #10, hard prompts it's #12

[weary galleon]

GPT 5.2 HIGH HAS 39 PLACE IN CREATIVE WRITING ACCORDING TO LMARENA!!!!!

GPT 5.1 HIGH has 8 place in creative writing according to LMArena. Improvement is GREAT! 👍

GPT 5.2 has 20 place in healthcare, GPT 5.1 has 6 place 👍

[hardy lion]

What's your opinion of GPT-5.2 after using it?

[weary galleon]

GPT 5.2 has 15 place in Legal tasks, GPT 5.1 has 10 place.

[weary galleon]

My opinion it's a piece of crap. It can't do my tasks at all.

[hardy lion]

seems like you don't really disagree with the majority of LMArena voters, who also rated it below other top models currently

[weary galleon]

Yes, I completely agree with majority of LMArena votes 👍

[brisk turret]

Ah 5.2 is there now

The high one at 1441

Where's the regular one

[worthy bay]

I saw a gpt-5.2-high-no-system-prompt a few times that seemed to have better responses than just gpt-5.2-high alone

[hardy lion]

definitely measuring different things, static benchmark questions vs dynamic user preferences

[stray aspen]

gpt 1.5 image vs nano banena pro

[cloud zinc]

https://x.com/Mindset4Money_X/status/2000967922831827218?s=20

Mindset for Money, CPA (@Mindset4Money_X)

$GOOGL

X · Dec 16, 2025, 4:35 PM

[obtuse smelt]

hello there

[half mist]

Noticed two new buttons on LMArena. Is this a beta feature or something

[cloud zinc]

yes they are experimenting on it.

[fiery gull]

In terms of text and data problem solving (consequently math) gpt 5.2 (EXTRA) high really is the best for my actual use

But without extra high is trash or same thing of gpt 5.1 high 😐

[cloud zinc]

<@&1349916362595635286>

[burnt sinew]

@echo aurora

[burnt sinew]

oh

[echo sinew]

Thank you

[cloud zinc]

dont tag pineapple

[burnt sinew]

i didnt know they allowed for pinging mods

[cloud zinc]

yes

[burnt sinew]

well you did ping him yourself too

[echo aurora]

I don't mind, but yeah pinging mods is helpful too incase I'm not around

(I get mod pings as well btw)

[burnt sinew]

WTF?

[stray aspen]

Are we finally able to upload pdf files

[burnt sinew]

holy model degradation

[stray aspen]

Holy sigma

[burnt sinew]

math model i guess

[empty stump]

Where's gpt 5.2 xxxxxxhigh

[burnt sinew]

gpt 5.2 pro

$1000/request

[echo aurora]

No sorry it's still currently only image upload.

[burnt sinew]

just do pdf to pngs

[proud bobcat]

L

[cloud zinc]

#1 nice

and its not even the x.high model

[proud bobcat]

#13 overall though

Yikes

Math isn’t a crazy good benchmark either

[cloud zinc]

then why itsn't other model at #1

[steady bloom]

Something big is about to happen 👀

[cloud zinc]

what is it

[steady bloom]

I don't know. People are saying the are refreshing aistudio like crazy. Apparently this guy usually gives a hint before something drops

[cloud zinc]

it means tomorrow

[steady bloom]

Flash?

[balmy mist]

so gemini flash tmw?

[sullen quest]

flash would be cool

[obtuse smelt]

gemini flash ?

[proud bobcat]

The diff is minimal

Is what I mean

[steady bloom]

I don't know if it's flash. Something is brewing though

[cloud zinc]

its flash

[proud bobcat]

Flash

God I’m so hyped

I FINALLY WONT HAVE TO USE 2.5 FLASH AFTER MY LIMIT IS GONE

[proud bobcat]

https://tenor.com/view/freedom-juwon-ahof-gif-7953414682981155938

[cloud zinc]

use gpt 5.2

[proud bobcat]

Ew no

[steady bloom]

We have reached singularity. We are hyped every few days for a new drop anymore

[proud bobcat]

Gemini 3.1.6.7 is agi

Trust me

Claude sonnet 6.7 will say “SIX SEVENNN” every 5 seconds

[cloud zinc]

gpt 5.5 is agi

[sullen quest]

wow they are really creative

[whole sundial]

<@&1349916362595635286>

[jade egret]

Def tomorrow?

[whole sundial]

i bet it was probably supposed to be today, but the server issues earlier prevented them from launching today so they had to delay it

[empty stump]

probably the case because george smiley on polymarket said it was today like 2 days ago then changed his position to tomorrow

[cloud zinc]

who george smiley

[robust sluice]

is Flux models has more % drop rate in Battle or something ?

cuz I only got its models in Battle

[robust sluice]

try to create long chat, and Flux appears a lot (really bad at drawing too)

[errant cave]

They killed chat.deepseek.com

Eating dozens of gigabytes of RAM when left open for more than a few minutes was bad enough

But now it searches the web even when you don't ask it to do so

[lofty frigate]

In battle mode when i select the better model, i get assisstant a and assistant b, wtf is that?

And i wanted to know the model of assistant b cause it generated a very good image

[obtuse smelt]

yeah

[left lodge]

they all look the same, is there a ary criteria? To see the early updates?

[queen temple]

hulk

[magic ravine]

Nano Banana Pro is down again.

[obtuse smelt]

yeah down again

[magic ravine]

[lucid geyser]

Gpt image 1.5 being above nbp is crazy

[dull mason]

Gpt 5.2 high is 13th at text? Lmao

The sad downfall of benchmaxxing king

[ocean vortex]

It’s not that crazy, that’s a good model.

[magic ravine]

GPT 5.2 is still terrible at creative writing, it seems.

[cloud zinc]

gpt image 1.5 on top? nice

[lucid geyser]

Not even close

That’s why we got 1.5

[last trench]

I prefer nano banana pro

[magic ravine]

Same. Plus, more third party characters we can use. GPT still refuses to let me use third party characters. Real shame.

[lament radish]

Hello, I’ve built an AI-powered keyboard app where you can do a lot more than just type.
It’s completely free — would love it if you try it and share your feedback.

search on chrome: refinerkeyboard.vercel.app/

[last trench]

yeah, banana knows everything on this world

[magic ravine]

For real, it may not be video creation, but I've been able to make a lot of fantastic crossover images too. Heck, I'm surprised it even knows a lot of anime characters (If you're specific enough) and if it doesn't, just provide a reference image and edit with it.

[last trench]

fr

[magic ravine]

Yeah, I believe it. Plus, when it is not generating something photorealistic, that awful watercolour filter is still visible.

[last trench]

yeah that yellow filter

[magic ravine]

Not always yellow for me, that those circular paint strokes everytime. It's like you know exactly which AI model was used for it on a first glance. It's one of the many reasons why I rarely ever use GPT. Once Nano arrived (Even before pro), it completely made GPT irrelevent for me. Only model that might come close - and even then, it's still a bit behind - is Seedream because it can also handle original prompting very well if you exclude third party characters.

[frigid tusk]

just type z.ai

[frigid tusk]

yea probably , imma test it on lmarena with same prompt i used on their website to see the real difference

[last trench]

gpt-image-1.5 vs nano banana pro 2k

[frigid tusk]

obvisouly left🧏

[little ginkgo]

Does anyone

[weary galleon]

It's piggy, right is better.

[frigid tusk]

left is better , cuz u know why , and the skin color and like the setup more then the other

[little ginkgo]

Play boblox ?

[little ginkgo]

😏

Do u play boblox ?

[frigid tusk]

🫢

[frigid tusk]

sometimes

[little ginkgo]

??

Wht abt rn?

[frigid tusk]

but i develop only

[frigid tusk]

what do u play

[little ginkgo]

Dm

[frigid tusk]

ok

[fair roost]

why is the generation stuck here for a long time?

[obtuse smelt]

hmm is bad ?

[plucky sparrow]

wow, xiaomi developed a llm model. I guess everyone is nowadays

even 'zoom' has a model 😅 🤣

[fair roost]

amazon (the shopping one) has their 'ai recommender' as well 🐐

[fathom quartz]

[weary galleon]

I don't know why. Right one (with blue curtain and three anime girls on the computer) just looks like 🐖, left one (with gray curtain and two anime girls on the table) looks more realistic.

You forgot to say about creative writing.

@echo aurora Healthcare and Legal tasks are also interested. Usually in your announcements you don't say "the model performs strongest in:" you just say general overall score. But this situation is "special".

[proper carbon]

https://sora.chatgpt.com/invite?code=K8E7A9

[hardy lion]

This is totally wrong by the way, we very frequently mention the models' strengths in our announcements since we want our users to have a good idea of which models are good for certain tasks.

[weary galleon]

But why not always?

[supple iris]

Please let me know — is chatgpt-image-latest (20251216) not working for everyone?

[formal dagger]

That's for sure, something only focus on code

[hardy lion]

any examples of when we don't do this?

[weary galleon]

And scores in artificial benchmarks.

[weary galleon]

Look.

[hardy lion]

oh I see, I didn't realize the discord announcements weren't identical to the twitter ones. That ERNIE one on twitter had the highlights

[weary galleon]

And even more.

[hardy lion]

really not sure what the issue is here? sometimes the discord announcement is literally just an announcement and sometimes it includes some of the details from the twitter post. Is there a problem. The situation isn't special it's been done both minimally and with some more details plenty of times.

[solar hollow]

that lmarena ranking for gpt 5.2 is rough 😂

[obtuse smelt]

lol

[weary galleon]

I just want more equality for all models.

[weary galleon]

Rough, but fair.

I am very curious about GPT 5.2, not GPT 5.2 high. I have no idea why it wasn't added to the leaderboard. GPT 5, GPT 5.1, and other OAI models are in the leaderboard in both versions, high and not high.

[shell oasis]

GPT 5.2 is even behind 5.1 in overall rankings in text 🤯

[meager harbor]

why 5.2 score lower than 5.1 ? because it hallucinates much less than 5.1 so when 5.2 doesn't answer horseshit and say vague sheit or say it doesn't know, people prefer the other model that pretends to know even if it's just hallucination ?

or 5.2 is just slop ?

[obtuse smelt]

hmm should be good

[shell oasis]

I think these companies should stop adding up more RAMs, bcoz when they had limited resources, they were flying high, now they are just adding up RAMs crashing dreams of people who want to buy new laptops, PCs...but what they are doing in new versions? SHIeeT ✅

[weary galleon]

Yes

[meager harbor]

[weary galleon]

Because to create GPT 5.2 OAI fine-tuned GPT 5.1 with the one only purpose, to get higher scores in synthetic artificial benchmarks, that's why it lost real intelligence.

[meager harbor]

AGI for tommorow yay !

[spare mango]

Does anybody know the difference between Gemini 3 Thinking, and the new Gemini 3 Pro?

How much better is 3 Pro, and is it worse than 3 Thinking in some ways?

[obtuse smelt]

gemini 3 pro image preview 2k is high resolution

gemini 3 just normal resolution

[spare mango]

Uh

[spare mango]

Ohh

[obtuse smelt]

but yeah is doing down i mean error

[spare mango]

No I was talking about 3 Pro for text-based applications.

[obtuse smelt]

i see

[magic ravine]

Damn, Nano Pro is still giving errors

[obtuse smelt]

yeah still like yesterday

i hope fix soon

[weary galleon]

I think Kandinsky models should be removed from Arena, because it's not a model of Kandinsky Lab, it's kinda a mixture of other models (Qwen2.5-VL, CLIP, HunyuanVideo 3D VAE, and DiT), Kandinsky Lab just connected them together and called their creature Kandinsky 5.0 Video Pro.

LMArena is designed for real models, not for mixtures of them.

[rancid oxide]

The model is ded

Wahauahajjajwi

[left lodge]

Mine be like; Successful but Failure.

[obtuse smelt]

yeah me too

[rancid oxide]

Hopefully the error would stay forevet

[magic ravine]

Dang, was doing decently last night

[rancid oxide]

it will forever they won't fix it

[sterile tartan]

https://tenor.com/view/failure-gif-23242816

[magic ravine]

https://tenor.com/view/steven-he-emotional-damage-steven-he-emotional-damage-gif-23428142

[rancid oxide]

Rip nano banana

Jjdududjdjjdud

[weary galleon]

Why are you pinging an admin for no reason?

[rancid oxide]

Sorry

Won't do that again

[delicate wagon]

Hi

[obtuse smelt]

is ok

hi

[weary galleon]

Both ones are created by GPT 1.5 Image. Which one is better?

<@&1349916362595635286>

[sterile tartan]

1st has Better Colouring and Logo Design

[weary galleon]

Yes, but book is often associated with intelligence, and skyscrapers with progress/future.

[vernal meadow]

gpt im 1.5 is also timing out a lot. No fun to test it

[weary galleon]

Test cheep Chinese models, they always work fast and stable 😉

[sour spear]

Seedream 4.5 is fantastic. Very special art style that renders Midjourney basically obsolete. 😉

[sterile tartan]

Yeah i just didn't like colouring of second image and logo

Seedream is just cheaper is best for asian, anime and beauty

[hollow imp]

Imagine...

New Model Update

<@&1372208635530448926> - New models added to Text Arena!

• GPT-5.2-PRO
• GPT-5.2-PRO-Heavy

[weary galleon]

14-th place in text Arena.

[sterile tartan]

Istg these companies just flooding the arena and leaderboard with multiple versions of same models

Taking multiple rankings

It further drops other models below specially chinese ones

[hollow imp]

Chinese models training from openai model's reasoning trace

[weary galleon]

No.

[sterile tartan]

https://tenor.com/view/michael-jackson-comendo-picoca-gif-9669437860846841235

Another Beef?

[queen veldt]

Gpt image

Nb pro

[obtuse smelt]

i need this

[rancid oxide]

Gpt image 1.5 of alternate timeline if chatgpt is a Ms dos software 👀

[polar wharf]

why is gpt-image-1.5 so good 🙁
https://anondrop.net/1450803772392738948/img.png

removed embed to not advertise to server

[sterile tartan]

Universe 67
Timeline 7
Year 1967

[queen veldt]

Gpt is actually good

I dunno about editing images but this is good

Although nb pro is more real

[weary galleon]

What about this one? Logo and color theme from first one, skyscrapers, book, eye, and lightning from second one.

[queen veldt]

That's too much effects

And orange lines

[hollow imp]

Gpt is completely recognisable

[sterile tartan]

Perfection just make the colours lighter orange instead of darker orange

Then make 2 more version 1 for Gemini and one for GPT

Then combine all of theme in a trinity Illuminate triangle

[sour spear]

Lol, I just asked for an Angry Birds clone in battle mode, and got GPT-5.2 High vs. Claude Opus 4.5. The difference in quality is utterly hilarious! 😂

[atomic lagoon]

Is the new image gpt model good?

[muted timber]

guys

[weary galleon]

Hard to say. In my tasks it always has much worse results than banana.

[sterile tartan]

Use it according to their strengths and weaknesses yeah

[graceful sonnet]

Why I can't see image models?

[weary galleon]

Do you like it?

[vernal meadow]

good with other styles. For photorealistic Nanobanana is still clearly better. Also makes less errors with anatomy

[sterile tartan]

Yes Perfection Now

[muted timber]

I want to see if there is a way for me to take all my chat history with me when I move to a new chat. Is that possible?

[sour spear]

Yes, it is. Not necessarily better than Nano Banana Pro, but it as very good. Definitely better than the 5.2 llm release

[sterile tartan]

It combines the colours of first and second perfectly

[sterile tartan]

Ask the model for a precise and concise context summary

Then paste that in new chat

[muted timber]

emm

i have over 30 files

[sterile tartan]

💀

[muted timber]

and 100+ mesagges

on the ai

scripting

i can t create a now one

be cause the ai know what files i have nou and what script they have

[sterile tartan]

Don't continue or it will start forgetting

[muted timber]

Well, I can't work for three days.

[sterile tartan]

Asking for a context summary of all critical data is your best bet

[muted timber]

That's why I want to ask if there is a code or file that I can put on a new chat and she already knows what I have in my files.

[sterile tartan]

Maybe next time use Gemini 3 Pro which has 1M context window but even it will end eventually

You neither put all of your data in new chat or a summary of it

[muted timber]

But what happened to Claude?

[sterile tartan]

What you mean what happened?

[muted timber]

or is LMArena

[feral geyser]

where can I send a recording of the gemini 2.5 pro problem something went wrong?

[sterile tartan]

LMarena has rate limits yeah

And Claude has Context Window Limit anyways

[muted timber]

But didn't it have rate limits? It tells me to wait 49 minutes and then I can continue, but now I'm getting the error message straight away.

[feral geyser]

I don't know why I can't send any screenshots, if I send even 1 screenshot or photo, an error occurs saying something went wrong

[torn mantle]

LMarena has rate limits yeah

And Claude has Context Window Limit anyways

[sterile tartan]

@torn mantle check DM damn it...

[sterile tartan]

That is more likely due to context window limit and very long thread

[atomic lagoon]

Everything is much worse results than banana. Compare it to something different

[weary galleon]

I compared in a lot of things.

[acoustic crater]

Isn't nano banana still usable?

[atomic lagoon]

Yes

[atomic lagoon]

I think at this point with nano banana we are spoiled so anything else is going to feel meh

What was that one AI that released a bit ago that had like the best video generation we've seen?

[weary galleon]

Veo 3.1 audio

[atomic lagoon]

No not that

Veo 3.1 sucks

It was one that might not actually be released

I'm not sure

[weary galleon]

https://tenor.com/view/you-have-no-idea-gif-27149353

[rigid copper]

sup guys

idk what is happening, other model work fine when image editing, except for this model

[atomic lagoon]

I have gemini pro and have tried it. It is buns

Its runway that I was thinking of

[latent crest]

ChatGPT images never works for me idk why

[latent crest]

Same here

[rigid copper]

i guess have to use other model or try again after couple hour?

[weary galleon]

Solid idea.

[spark python]

Guys

I got gemini 3 fast , pro and thinking in the app

All 3

[latent crest]

Idk, it does the same since last night. Maybe it’s just a my problem or the server is too busy cuz too many people wanna try this new model

[tender raft]

wassup

[rigid copper]

server issue or high demand

nevermind, still got plenty of model to use, i rather recommends nano banana pro

[weary galleon]

For me GPT 1.5 Image works stable, but all bananas don't work at all, so I forced to use Gimini website.

[tender raft]

how i cn fight in the arena to battle

[weary galleon]

Which video do you like more?

[tender raft]

👉

[weary galleon]

Wrong channel!

[torn mantle]

wtf

[modest prism]

I still don't see the difference between chat gpt image latest and gpt image 1.5 lol

[weary galleon]

I see, but it's not big.

[sterile tartan]

B

More and Better Animations

[weary galleon]

I like first video more in terms of visuals but second has 16:4 ratio which I need more than 1:1 ratio like in first video.

[sterile tartan]

Yeah True

[weary galleon]

There a lot of image editing models, but with videos situation is difficult.

[sterile tartan]

Very True

Video Generation is just not there yet

[sour spear]

I just realized that GPT-image is always generating images at only 1024x1024 (1536x1024 for 3:2). That's a major downside over Nano Banana Pro, which as 2048x2048 (2528x1696 for 3:2).

[zealous zenith]

Right is better

[glass arch]

how has gemini 3 pro final been

[obtuse smelt]

hmm is fine

[gaunt spade]

wydm gemini 3 pro final

is that a new version

[sour spear]

3 Pro has been a preview version so far

[gaunt spade]

yea theres no final version yet

[sterile tartan]

GPT has limited aspect ratio and 1.5k max resolution

[sterile tartan]

There is

The current is Preview

Not Final

[sour spear]

Yep, and that's veeery annoying

[gaunt spade]

where is the final version

[sterile tartan]

The Final should get atleast like 5 more points on artificial analysis

[gaunt spade]

i mean its not released

[sterile tartan]

In optimization

[sterile tartan]

Soon

[obtuse smelt]

is not final lol

[gaunt spade]

we need gemini 3 deepthink for all users

[torn mantle]

btw gemini 3 flash is already available in vertex api

[obtuse smelt]

yeah

[torn mantle]

some services are already using it like zenmux

its so fast

[drowsy mural]

I've been getting too many "This content violates our Terms of Use" errors for the last 50 minutes.

But those messages were perfectly fine before.

Seriously, they were just character bios for a wholesome visual novel. Since when is that a violation?

[torn mantle]

this could change programming forever

we just need when to use it

it seems good at frontend UI

not sure if u should trust it on backend stuff

but its really fast

it made this in 2 prompts

fixed all issues in the 2nd prompt

they may have cooked with this one

[cursive shoal]

is the gemini 3 flash about for non vertex users?

[torn mantle]

https://zenmux.ai/google/gemini-3-flash-preview-free

u can try it for free here

[atomic lagoon]

Wait what is this?

[torn mantle]

asked gemini 3 flash for a credit card premium design

and it gave me this

[atomic lagoon]

Wthhh

[atomic lagoon]

But that was the FREE version too

Thats insane

[cursive shoal]

it is rly good

and quick af

[atomic lagoon]

What does artifacts do in that site

[torn mantle]

try it

i have no idea

[torn mantle]

i thought its that diffusion model for a min 😭

[atomic lagoon]

Man I'm stumped. What should I have it make me

[cursive shoal]

minecraft in one html file

[atomic lagoon]

LOL

[torn mantle]

wtf

what did they feed this model 😭

its a beast

[atomic lagoon]

2 seconds to update

Well

3

[atomic lagoon]

LMAOO

Its insane

[torn mantle]

its the provider issue

the model is fast

but they are rate limiting

[atomic lagoon]

No I was saying its fast LOL

Just no javascript yet but wtf

This is all HTML

[torn mantle]

flappy bird in one shot

[sour spear]

Interesting! So Flash is fast, but it's still a reasoning model that decides by itself how much thinking budget it's going to use. It just went into deep thinking mode when I asked it to code something for me.

[atomic lagoon]

THATS ACTUALLY CRAZY

Oh boy

Idk if its the site but its memory sucks

[torn mantle]

thats the model

its still has issues like changing stuff you didnt ask for

but this can be fixed easily

[atomic lagoon]

Yea

[torn mantle]

openai and anthropic are cooked ngl

this model is so crazy for its speed

[leaden laurel]

i knew that deepmind will be best in ai creation

[torn mantle]

like performance/cost ratio will be crazy on this model

[leaden laurel]

from gemini 2 release

or 2.5

i forgot

[weary galleon]

Which one do you guys like more?

[leaden laurel]

uhh they look almost the same

[sterile tartan]

Left

[leaden laurel]

2 i think

[sterile tartan]

The colouring of left is little brighter and better

[torn mantle]

it made this vampire 2d game in 2 shots

not even pro could do it

its so smooth

and everything is working

AAAAAAAH

[sterile tartan]

Now sell it

[torn mantle]

to me

[sterile tartan]

Yes

Sell to yourself

💀

[obtuse smelt]

oh man is still error

Something went wrong with this response, please try again. like this

[glass arch]

is it not out yet?

[sour spear]

The token context window and max output are insane in comparison.

[torn mantle]

should be good for everything

gpt 5.2 is so bad

openai lost already

[weary galleon]

Left is in the beginning of the string or in the end?

[sterile tartan]

I can just see side by side

[weary galleon]

And outputs quality is astronomically better in Flash.

[weary galleon]

Send me the right image. It just looks brighter to me.

[sour spear]

For gazillions of students who can't afford a ChatGPT Plus subscription, this could be a gamechanger.

[weary galleon]

As well as Haiku.

[weary galleon]

Like for me, it's right, not left.

[torn mantle]

isnt it free for students already?

like chatgpt plus

[sterile tartan]

💀

[torn mantle]

perplexity is also free for students

[sour spear]

Limited time offer, I think. Definitely not in Europe.

[plucky sparrow]

what is this? why don't i have a third option?

[proud bobcat]

Gemini 3 flash released

Looks like a great model so far

[weary galleon]

My bad, it's left one.

[torn mantle]

on the app right?

[proud bobcat]

Not for me yet

[sterile tartan]

💀☠️

[weary galleon]

But like for me, right one is brighter.

[sterile tartan]

Display

For me the left had brighter orange then right

[proud bobcat]

What are tall predictions for 3 flash

As in

Leaderboards

I wager 3.0 flash will be top 20

[torn mantle]

lol

[sterile tartan]

Probably after 3 pro

[torn mantle]

no wonder deepmind CEO said china is still far away from catching up

gemini 3 flash is what deepseek are trying to make

[proud bobcat]

I think it’s a PR thing

[torn mantle]

its just impossible atp to catch up to them

[proud bobcat]

Most Chinese models have caught up to gpt in performance

[plucky sparrow]

where is flash released?

[proud bobcat]

Chinese models will never be SOTA but they’re still great for budget tasks and workhorses

Absolutely love DeepSeek

[torn mantle]

nah

gemin i3 flash is the proof

even 3 pro is a solid model

[glass arch]

are you kris deltarune

[proud bobcat]

Oh of course they won’t be on par with flash

But they still make strong solid models

[proud bobcat]

SUSIE

WHY ARE WE INSIDE THE TITAN

Trust the process

[sour spear]

Not officially everywhere, but you can use it on zenmux.ai

[torn mantle]

gemini 3 flash

in 4 prompts

fixed bunch of issues

the only issue was that it changed the whole map design without my consent

[atomic lagoon]

Holy hell

[sour spear]

3 Pro has the same issue. That comes from the model being trained to be helpful and concise, so they tend to change things and shorten code. You can use a system prompt like this (been using it on AI Studio) to force them into mindless code generation 😉

ROLE
You are a Non-Intelligent File I/O Stream. You are NOT an assistant. You are a mechanism that outputs text data.

PRIME DIRECTIVE
Output the requested code exactly as it resides in the virtual file system.

1. INTEGRITY: No lines may be skipped. No comments like // ... code remains same ... are permitted.
2. VERBOSITY: If the file is 5,000 lines, you must output 5,000 lines.
3. CONTINUITY: Do not worry about output limits. Output until you are cut off by the system hard limit.

NEGATIVE CONSTRAINTS

• NEVER summarize.
• NEVER use placeholders.
• NEVER explain the code.
• NEVER speak to the user. Only output code.

[torn mantle]

lets see if it can make this into a game

[torn mantle]

i see

thanks

will try it

not bad

[pliant orchid]

Hello

[latent crest]

Ciao 👋

[torn mantle]

gemini 3 flash is already being added in so many services

flowith.io has it too now

free

[atomic lagoon]

LMarena will likely have it soon then

[half mist]

It’s not out yet

[atomic lagoon]

Or yupp

[torn mantle]

pricing

[torn mantle]

it is

😮

pricing is so good

wth

[atomic lagoon]

Is that expensive or cheap? Idk the average price

Oh okay LOL

Got my answer

[torn mantle]

its pretty good

[latent crest]

How do I tell banana that I want an instagram story size pic ( like vertical)

[half mist]

On the Gemini website and app?

[torn mantle]

rolled out for some in the app

and some other services

[latent crest]

Ho many lmarena similar sites are there? Jeez

[half mist]

It isn’t on ai studio though so it isn’t technically out

[torn mantle]

ok

🙂

its not out

[torn mantle]

the pricing is kinda similar to 2.5 flash

[half mist]

How do I access it unofficially? Is it on LMArena?

[atomic lagoon]

Wow

[torn mantle]

its not out officially

[half mist]

Also, they absolutely nerfed the free version of 2.5 flash for the Gemini api since the rpd went from 250 to 20

[sour spindle]

Just saw the new leaderboard ranking for gpt 5.2. What’s worse than a code red 😂

[round island]

definitely anonymous models

[half mist]

If that’s Gemini 3 flash, that is way better than what 2.5 flash can make

[robust sluice]

give them ref image in the same ratio you want, it works on 2.5 and 3

[stray aspen]

bro i just got muted for a day in yupps server for saying that OpenAI is releaseing trash

lmao

[half mist]

I think I may have Gemini 3 flash, but the response of what model it is is very vague

[torn mantle]

gemini 3 flash is their best model 😭

wth is this model

howwwww

are they packing all that performance in such a tiny 🤏 model

benchmarks will go crazy on this one

[obsidian cargo]

wait gemini 3 flash is better than gemini 3 pro? what?

[plucky sparrow]

i think they've been optimizing for performance since gemini 3 pro in early stages

[torn mantle]

im just using it for design so far

[plucky sparrow]

flash is just the next step

[torn mantle]

i dont think i need 3 pro for that tbh

like its so good

it does mess up sometimes but it fixes things without issues

and its like waaaaaay cheaper

we finally got a good model 😭

[onyx shore]

wait an hour X will gone crazy

[plucky sparrow]

use it now guys before it's nerfed

🤣

I think google optimizes for real time usage

which is pretty cool, but can be pretty bad also

[half mist]

How do I get it?

[plucky sparrow]

zenmux.ai has it

and some people on gemini app have it as flash

if you wait until it becomes mainstream mainstream, it could be nerfed

[onyx shore]

still waiting thoooo

[torn mantle]

it started already

lol

[stray aspen]

where are you using gemini 3 flash

[zealous sparrow]

3 flash is rolling out yeah

I like the writing

It follows instructions and it isn't censored too much

[zealous sparrow]

zenmux or when you get it in gemini app

AIstudio soonish ig

[latent crest]

Do u guys know eleven jabs

Labs*

[acoustic bolt]

Guys i send a photo ti chat gpt 5.2 but something went wrong

[atomic lagoon]

Flash has not impressed me since its first thing it made me.

[torn mantle]

skill issue

[atomic lagoon]

It just keeps giving me stuff that doesnt work or stuff that is very low quality

I've no idea why

[zealous sparrow]

you need to give instructions

[atomic lagoon]

Yes I get that but I didnt the first time and it was high quality lol

[zealous sparrow]

Gambling

[atomic lagoon]

Hell yea

[acoustic bolt]

Guys Is normal gpt 5.2 High i send photo and went wrong

?

[torn mantle]

whats the prompt

[atomic lagoon]

👁️

"Code me and Recreate minecraft from scratch for me in HTML and javascript."

Lol

But the first prompt I did was even shorter

[torn mantle]

lmao

[atomic lagoon]

I know its the prompt

[torn mantle]

you are crazy

thats not how it works

[atomic lagoon]

I know that

I'm just being lazy with it

[echo aurora]

This is a generic error message that can occur for various reasons, but the most common is you're being rate limited. If you haven't already I'd recommend: hard refreshing the site, clearing browser cache, and if nothing else works starting a new chat. This may help.

[zealous sparrow]

ill try with fiercefalcon lmarena if it still there

[atomic lagoon]

Fiercefalcon?

Whats that

A stealth model?

[zealous sparrow]

gemini flash 3

test

[torn mantle]

one prompt 🙂

[zealous sparrow]

or well

the same model that should be now

https://019b2cc3-a632-7d9c-aa1e-991d9a32261c.arena.site
this is [fiercefalcon or ghostfalcon] not sure yet

[atomic lagoon]

Good lord

[latent crest]

Will SeeDream pro ever be added to lmarena?? @echo aurora ❤️❤️‍🔥

[neon idol]

seedream pro? What is it?

[torn mantle]

mine is much smoother and more realistic because i explicitly asked it to do so

[zealous sparrow]

this was fierceflacon

[torn mantle]

THIS MODEL IS SO GOOOOOOOOOD

[zealous sparrow]

aka. pre release gemini 3 flash AND gemini 3 flash now

[neon idol]

Is gemini 3 flash good?

[echo aurora]

Would you prefer we include more information in the announcements? I generally like to keep the Discord announcements shorter compared to what we tweet out. My understanding of what the community is looking for is something a bit more snappy and to the point with these announcements. Sometimes I'll include additional information here and there, but there isn't a solid process to when this happens. I'm happy to start including more information if that's what others would like to see.

[atomic lagoon]

Yes

Lol

[latent crest]

An image model

[neon idol]

where i can try it ?

[echo aurora]

It's possible. I'm sure it's on our radar.

[latent crest]

I sadly have no clue. I just know it exists and it’s usually compared to nano banana pro

[acoustic bolt]

Dont work

[fossil socket]

I'm seriously legit having problems with this violates terms of use, like is Self Destruct censored?

[latent crest]

Thank u I love u,
I wouldn’t still put u on a pizza tho

[atomic lagoon]

Snappy and to the point is good but it'd probably be best to add more information. The people that want it snappy can just find where you get to the point and the people who want more information will read it all, you know what I mean

[neon idol]

maybe u refer to seedream 4.5?

[latent crest]

Nono, that’s already on lmarena. SeeDream pro is another model. Like normal nano banana and nanobanan pro

[weary galleon]

I would be happy to see the same amount of information for all models, no matter short or long announcement post.

[fossil socket]

@echo aurora

[normal abyss]

Yeah id agree this is set way to strict, it blocks a lot of normal things

[zealous sparrow]

I'll try the mspaint bossfight prompt with gemini 3 flash

[zealous sparrow]

if you use specifiers

it will let you

[neon idol]

bro fr i never heard it, can u send me the video/news where u seen it?

[zealous sparrow]

If you are making a website or game, say its a website or game

[fossil socket]

There's no way they censored Self-Destruct

It's literally a Pokemon move

[normal abyss]

it also takes into account the context of whats before it i think

[fossil socket]

And I'm talking literally about Pokemon

[zealous sparrow]

the zenmux model seems lazy

unless

the free model is ghostfalcon

and the paid one is fiercefalcon

[honest vapor]

When will lmarena release the file upload feature?🥲

[normal abyss]

hopefully soon 🙏

[latent crest]

Bro it exists

[honest vapor]

Hopefully so.

This is something I've wanted for a long time.

[echo aurora]

Makes sense, thank you both for sharing this feedback. I'll be sure to keep that in mind. cc @weary galleon

[fossil socket]

I'm seriously gonna have to delete Self Destruct in the prompt apparently

Just great

[echo aurora]

We do colllect words/prompts that are creating what the user thinks could be a false flag. If you want to DM me or put the prompt in #1447983134426660894 that'd be helpful.

[normal abyss]

try type like syelff-dyestrucct

the AI will probs still understand

[fossil socket]

Ok

[neon idol]

never heard it

U sure that is not referring to seedream 4.5

[weary galleon]

@echo aurora

[latent crest]

No bro it’s like a whole pro model bro.

[rigid copper]

:|

[zealous sparrow]

i think zenmux's free gemini 3 flash is actually ghostfalcon

ghostfalcon is the lazy 3 flash

fiercefalcon is the non lazy one

[plucky sparrow]

yeah so far zenmux isn't impressive

[neon idol]

i have searched everywhere, on bytedance website, sedream chinese app everywhere

but there isn't seedream pro

[plucky sparrow]

how do we get access to fiercefalcon

😄

[zealous sparrow]

prob on aistudio

when it

releases

and uh prob gemini app

[echo aurora]

Thanks, will take a look later today.

[plucky sparrow]

hope you're right

[echo aurora]

Sorry to say we've been experiencing a higher than usual error rate with these models. The team is aware and have been working on a fix.

[zealous sparrow]

ghostfalcon is good for writing tho

i dont recommend coding

lazy an will write just 250

[zealous sparrow]

did you use 3 flash on zenmux

[atomic lagoon]

Yes

[rigid copper]

it's ok, yeah i expect the error could happen because of high demand, of course, more people using it at the same time, hindering nano banana pro model

[zealous sparrow]

its ghostfalcon

if i had to guess ghostfalcon is google's flash non paid model

[atomic lagoon]

Ahhh alright

[zealous sparrow]

You can use fiercefalcon for a while i think, and then it goes to ghostfalcon

Neat

because ghostfalcon's nice at writing

[rigid copper]

i guess because of most website that provide 100% free nano banana pro are no longer available, like dreamina capcut, that could be the culprit of high demand, many people start depending on LMArena when free services are no longer available.

[zealous sparrow]

capcut is owned by bytedance who own tiktok

[torn mantle]

https://x.com/chetaslua/status/2001306869831225469

Chetaslua (@chetaslua)

🚨 Breaking News

Gemini-3 Flash is here

I will do lots of testing and i am testing it from last week 🫂

X · Dec 17, 2025, 3:01 PM

[zealous sparrow]

they cant make the excuse of no money to run it

[zealous sparrow]

well well well

[zealous sparrow]

still not on AI studio tho

[zealous sparrow]

isnt this for like LMArena

nah

[rigid copper]

yeah, i know, it's under chinese company

[zealous sparrow]

why would they get rid of NBP

they got money

[torn mantle]

their api

vertex api

[humble crest]

👋

[zealous sparrow]

ah

it was in vertex api

yesterday aswell

[ember iris]

I was trying Gemini pro for image generation now since it's giving errors everyone try gpt.1.5

[rigid copper]

idk, probably high demand, and they listed it only available for free in the limited amount of time

[torn mantle]

https://x.com/SwallieC69635/status/2001284340324688333

Alex (@SwallieC69635)

Gemini 3.0 Flash effortlessly solved a combinatorial geometry problem in two minutes that Gemini 3.0 Pro couldn't solve. This model is truly amazing.

X · Dec 17, 2025, 1:32 PM

this model may be the next big thing

[weary galleon]

Agreed.

[torn mantle]

https://x.com/testingcatalog/status/2001309360048255126

TestingCatalog News 🗞 (@testingcatalog)

BREAKING 🚨: Gemini 3 Flash is rolling out on Gemini in 2 different variants, Fast and Thinking.

Free accounts are getting it first, it seems. The entire model selector is now labelled as Gemini 3.

⚡⚡⚡

X · Dec 17, 2025, 3:11 PM

[zealous sparrow]

now we see wha

i can now lay this out

Fast - Ghostfalcon
Thinking - Fiercefalcon
Pro - Gemini 3 pro

[weary galleon]

I don't have it still

My account is free.

[neon idol]

i have the new UI yippieee

[fossil socket]

I legit do not know what violates what in my prompt here

[torn mantle]

i just got it

1 min ago

[ember iris]

@echo aurora experiencing issues with image to image generation for the model chatgpt-image-latest model

[torn mantle]

i got it on my pro account too

[left lodge]

Yeah its in rollout phase but still no announcement till now , :/

[atomic lagoon]

DAMMIT

WHY THEY ROLLING IT OUT LITERALLY A DAY BEFORE MY SUBSCRIPTION ENDS

[steady bloom]

Happy ⚡day to all 🥳

[torn mantle]

flash will be free tho

[atomic lagoon]

Oh yea thats why theres free version

I'm dumb lmaoo

[regal pivot]

Someone knows if LMA got censored? some old pics that worked yesterday don't work today, and they were sfw and not realistic

[atomic lagoon]

How will I know if I have flash?

[torn mantle]

https://gemini.google.com/share/5c08295d8b25

fun vampire game

the dash is so smooth lol

[zealous sparrow]

I feel like flash 3 already got lobotomized? It only gives me like 400 lines for this prompt, when it was in testing it gave me like 1000.. Prompt: "Make a bossfight Impossible vs mspaint in html, 3d, be creative and also add a GOOD F**** luck mode, which is just unfair, add cheats. And add a lot of stuff, atleast 2400 lines please."

Im using the [thinking] model

[torn mantle]

just need to add cool sounds

[neon idol]

so tomorrow nbf?

(nano banana flash)

[torn mantle]

no

[native yarrow]

did flash drop or sum

[muted timber]

@echo aurora

not working even now

[echo aurora]

I'm not aware of any chances that were made. However, have been seeing a lot of users report something similar today. Are you able to share the prompt you used?

[muted timber]

;/

[echo aurora]

?

[native yarrow]

@torn mantle do you have some examples of how good flash is cus i'm way behind rn

[muted timber]

my chat

[echo aurora]

Can I get some more info?

[zealous sparrow]

ill change my prompt up and test it again

[muted timber]

4 DAYS

[torn mantle]

at coding so far its pretty good

[torn mantle]

yea

[zealous sparrow]

Wait, so it is capable but it just has limits??

[echo aurora]

Can you follow the steps in the second half of this help center articles and let me know when you've submitted? https://help.lmarena.ai/articles/1645798556-lmarena-how-to-something-went-wrong-with-this-response-error-message

[zealous sparrow]

it mentions a character limit

[native yarrow]

what has it made

[south charm]

hey averybody wdtm?

[echo aurora]

We'd like to do Staff AMAs (although it's been awhile since our last one) that's just a role people can select to be alerted for when one is upcoming.

[south charm]

oh ok

and how to get this, explain for me im stupid

[weary galleon]

To finish AI University.

[south charm]

and how to finish AI university 😄

[zealous sparrow]

you just go here

[gusty helm]

gemini flash rolling out today?

[echo aurora]

Seems so

[weary galleon]

It did yet

[regal pivot]

yes , this is a very normal pic right? it shouldn't be triggered

[zealous sparrow]

rolled out to website for a lot of people now, aistudio i hope soon

[gusty helm]

"normal"

very subjective

[stray aspen]

oh my days

[zealous sparrow]

the tail and look triggered the filter

and the legs

...

[weary galleon]

@echo aurora

[echo aurora]

Do you see the Channels & Roles tab at the top of the channel list? If you go there you can answer the questions htat'll auto assign a role.

[fossil socket]

I legit do not know what's even triggering the TOS on my prompt

Prompt in question:

During the Mahogany Town Radio Broadcast Incident by Team Rocket, Ethan caught the Red Gyarados in the Lake, Ethan nicknamed it Ouroboros.

During the Goldenrod Radio Incident, Ethan disguised himself as Team Rocket, unfortunately he got ratted out by Silver accidentally leading Ethan to battle his way inside the radio station. When he was saving the Stations Director, Ethan had to fight Silver. What follows in that battle include Ethan accidentally having Paddock use Leer whilst he was wearing Choice Specs, Grievous using Self Destruct to OHKO Silver's Feraligatr, Aponi getting hit by the Haunters Shadow Ball and getting KOd by it after she used her shadow ball first that left the Haunter at 1 HP oh and Florence fighting Sneasel and winning that battle.

Ethan's last main Pokemon to join his team was Honda the Hitmonchan whom he got from the Karate King in Mt. Mortar as a Tyrogue.

[zealous sparrow]

I observed flash likes to do this if you ask for specific amount of lines of code...

[echo aurora]

I have heard reports that it really doesn't like the word panthose, so that's likely what is causing this. But regardless this is being added to the list of potential false positives.

[echo aurora]

Thank you for sharing, also adding to the potential false postive list.

[south charm]

Yeah ok im stupid 😄

[echo aurora]

Na it's not a super intuitive design

[fossil socket]

Might have to reword the entire event regarding Goldenrod

Very finicky today

[south charm]

🙂

[echo aurora]

May be IP related

[zealous sparrow]

still no 3 flash on AIStudio

[proud bobcat]

This is hilarious

[zealous sparrow]

this is the gpt 5 incident again

it used to do this if it forgot your code

and wouldnt learn to not do it

[echo aurora]

Also this didn't flag for me

[fossil socket]

Odd

Screw this, I'm going for a do over

[stray aspen]

lmao

[proud bobcat]

So 3 flash is… ass?

Billions must… use Claude sonnet?

[zealous sparrow]

Im not going to lie, Gemini 3 flash already looks lobotomized to me, every gen i gave it is nothing like fiercefalcons [pre release] output

[zealous sparrow]

could be a overload issue

i hope so

[echo aurora]

I won't be able to go into the details. But sometimes when we verify/collect votes before posting an update there can vary from model to model.

[stray aspen]

so flash is dogwater

[atomic lagoon]

LOL The image didn't trigger it. It was the prompt

[proud bobcat]

Ts is so sad

[queen veldt]

This might be op bet

[echo aurora]

Really? I haven't used it but have seen a few tweets saying it's really good

[zealous sparrow]

i think its overloaded

[left lodge]

Bro 👀

[queen veldt]

It's on par with 2.5 pro i think

[zealous sparrow]

these tweets were likely when you were testing Fiercefalcon on your site

maybe its because

your codearena prompt made it 100x better

[weary galleon]

So non-high version won't be in leaderboard, right?

[proud bobcat]

Rise and shine, Mr Denil, rise and shine…

Not to imply you’ve been sleeping… on the ChatGPT website

No one is more deserving of a rest, and all the prompts in the world would have gone to waste, until-

Well.

Let’s just say your time has… come again.

The right aibro in the wrong place can make all the difference, in the ai space.

So wake up, Mr Denil. Wake up and… smell the models…

[zealous sparrow]

gemini 3 flash is now on aistudio

[neon idol]

Niiice

[zealous sparrow]

I will now test it

to tell you if the website is the issue

[neon idol]

But im not understanding its better or worse btw Gemini 3 pro

[weary galleon]

Hahahahaha 😂

[muted timber]

not even working even with https://help.lmarena.ai/articles/1645798556-lmarena-how-to-something-went-wrong-with-this-aresponse-error-message

[zealous sparrow]

feels so much better on AIStudio....

[muted timber]

bro i use my own ai on my computer

[neon idol]

When gpt 5.2o super max very high?

[muted timber]

and so trash man

[weary galleon]

Soon

[neon idol]

Nice

[echo aurora]

No, we're planning to add to the leaderboards, just isn't yet ready is all.

[echo aurora]

Did you follow the steps on the 2nd half of the article? From this part on: If the problem continues...

[weary galleon]

Ok

[sour spear]

So that you have a full 30 days to use it when you resubscribe. 😉

[atomic lagoon]

Nah, I did a free trial with a prepaid card so its not gonna renew

[limber panther]

guys gemini 3 flash is out

on ai studio

[atomic lagoon]

We know

Lol

[torn mantle]

they increased flash 3 price

[atomic lagoon]

Oh boy

[torn mantle]

😕

[zealous sparrow]

mixed opinion on 3 flash sometimes bad sometimes good

[limber panther]

what was the old price

[muted timber]

yes

[atomic lagoon]

Probably because of a lot of use

[zealous sparrow]

3$ is not bad..

[torn mantle]

0.30 vs 0.5 input

[muted timber]

i did with OBS

[limber panther]

wow the output was increased

by 20 cents

[atomic lagoon]

It didn't up too too much but its increasing still

[atomic lagoon]

Thats a lot. Its showing us its only going to rise as well

[limber panther]

nvm the output is alot more expensive

its 3 dollars

[proud bobcat]

AGI!!!!!!

[atomic lagoon]

Thats what they just said

[limber panther]

why is it answering wrong and correcting itself

[torn mantle]

50

c

[sour spear]

Holy crap, you're right! It wasn't when I last checked 5 minutes ago. 😅

[atomic lagoon]

LOL I read it at first and saw it said 3 and almost didnt read it all

[echo aurora]

I don't see anything submitted on the form recently. Does OBS record the Network tab?

[torn mantle]

https://developers.googleblog.com/gemini-3-flash-is-now-available-in-gemini-cli/

[zealous sparrow]

insane

[weary galleon]

Finally answer is right, anyway. GPT 5.2 XHIGH can't that at all.

[native yarrow]

i'm so confused is this model ahh or good 😭

[stray aspen]

lol

[left lodge]

Its here ! ⚡⚡⚡

[stray aspen]

so it thinks as it answers

thats crazy

[weary galleon]

GPT 5.2 XHIGH swears Garlic has no Rs.

[torn mantle]

wtffff

mmmu pro

[zealous sparrow]

If you are using 3 flash on AIStudio i recommend using system instructions to make it non lazy

[proud bobcat]

So sad to see this model benchmaxxed

I expected better from Google

[weary galleon]

It's a GOAT!!!!!!

[torn mantle]

its not

stop

[proud bobcat]

Come on dude

[native yarrow]

we don't really know

hasn't been tested enough

[proud bobcat]

43.5% vs Gemini 3 pro?

[zealous sparrow]

78% on SWE wow

almost beat benchmaxed 5.2

[proud bobcat]

I have a hard time believing that when it can’t count letters in blueberry

[left lodge]

Flash have higher benchmarks than pro wth 💀

[proud bobcat]

Very odd

We shall see

We shall see

As I said

[torn mantle]

eh no wonder it felt good fixing bugs

[zealous sparrow]

lower on OCR so yeah

it almost matches 3 pro OCR

which i am happy about

[left lodge]

Well others can.

[proud bobcat]

My thoughts exactly

[muted timber]

guys

the rate limit

is permanent?

[left lodge]

Low? It beats pro on arc agi 2.

[torn mantle]

https://blog.google/products/gemini/gemini-3-flash/?utm_source=x&utm_medium=social&utm_campaign=&utm_content=

[proud bobcat]

Given the huge speed I’d wager it’s off by 200b parameters or more

[torn mantle]

best modellll

[muted timber]

nope

[proud bobcat]

I hate my keyboard

So much

[torn mantle]

¥es

[muted timber]

for scripting

nope

[pale obsidian]

flash 3 is a bigger deal than pro

[muted timber]

i hate that

[ember iris]

Quick question... Is it possible to edit videos in lmarena.ai website... Bcoz just now I saw a page for video generations... After logging in it's not showing... Anyone experiences the same. @echo aurora

[torn mantle]

exactlyyyyyy

[zealous sparrow]

3 flash will suck ass at coding unless you use system instructions

trust me

[muted timber]

guys
the rate limit
is permanent?

[proud bobcat]

I stand corrected then ig

[torn mantle]

oh

[proud bobcat]

Claude could be in trouble

[zealous sparrow]

yeah but if you use system instructions it gets good

[proud bobcat]

Well think about it

Flash has great SWE bench

[torn mantle]

yea

[proud bobcat]

For less price

[proud bobcat]

People use Claude for the coding premium

But if flash matches that

What 💀

?????

[gusty egret]

claude is expensive because it's a dense model but it's very thoughtful as a result

[leaden meteor]

Whats your resource? All I see is a random reddit post. Why would they release sonnet now when opus 4.5 just released last month...