general

[reef pawn]

LG is cooking

[neon idol]

Fr

[pure falcon]

It makes me wonder what kind of person should use battle. Power user? Early adopter? Average person?

Right now, lots of people who may not be able to afford GPT pro, are free-riding direct chats to get gpt-5 high access.

You could maybe gate + cap direct chats at like, idk, 15 messages and force them to respond to a battle before continuing.

I do wonder how the average battle user vs direct chat user differs. My guess is direct chat people are younger, less income, etc. Winder what effect that ultimately has on the leaderboards / stats / rankings

[agile bloom]

gpt-5-search model can search the web?

[neon idol]

Yes

[pure falcon]

Though tbh, the style control thing confuses me. LMArena is all about moving away from “objective” benchmarks to the ones that matter most: user preferences.

If users like different styles and formatting, then why control for it? The whole point of this is to find out what users like!! Defeats the whole purpose imo

[pure falcon]

@ocean vortex

[ocean vortex]

I think AI enthusiasts. But those people are not gonna use it how it was designed for if they can't actually test unreleased models for even a single message after voting

[neon idol]

@echo aurora What is this new ai image generator called nano banana that is in LMarena

[jade egret]

😭

[ocean vortex]

OLED AI

[obsidian cargo]

stealth model. I think it's an imagen variant

[ocean vortex]

😭

[pure falcon]

Well, LMArena got a nice check from a16z and co. But eventually they’re gonna have to find out who their customers really are. Users or AI companies? I imagine they charge model makers to put their model up in the arena, no?

[echo aurora]

For models that are behind a codename I won't be providing info about them

[neon idol]

Oh ok

[ocean vortex]

AI labs are the "customers" but users are enablers. Without users there's no benchmark

[pure falcon]

Right. So from the model makers eyes, the rankings have to converge on the average user. Otherwise it’s useless. OpenAI, for instance, went with summit over zenith bc of better LMArena scores.

If the people battling in the arena are different from the companies average user, then it becomes a bad benchmark. LMArena must find a way to match those two pools of people

[hollow imp]

Has anyone tried minimax? How is it

[ocean vortex]

I think it would naturally converge for the most part. More interest in battle mode = more users participating = more reach = more people will hear about it and try it.

[pure falcon]

Haha, are we sure about that?? You just said that early arena testing is made for enthusiasts lol, which i do agree with. Enthusiasts aren’t average users though.

One thing LMArena should roll out, is user accounts. Allowing accounts on the platform would go a long way to help them get the data they need, in order to answer that question

[stray aspen]

it sucks

however they have great tts models

and video models

[hollow imp]

Text to speech?

[hollow imp]

So according to you the only things which do not suck are 4.1 opus, 2.5 pro, grok 4, some specific versions of gpt 5

[ocean vortex]

The demographic is not changing and you can't easily control that (+I'm not sure you should). The best you can do is be increasing the interest and activity...

Something like chatgpt started with like only the most die hard nerds at it's infancy, look at it now... 👀

They didn't do anything to push those away and make their audience "more balanced" at the time and that was the right move

[jade egret]

when gemini 3 ):

[neon idol]

Fr

[ocean vortex]

Not as good as gpt5 😇

Also interesting fact... it doesn't even have 1% of chatgpt market share lol

[stray aspen]

claude and 2.5 pro suck

grok 4 and gpt-5 are good

[ocean vortex]

It has... you are very wrong actually. It's more than 20%

Gemini is quite substantial

Mostly thanks to Android I think

They integrated Gemini into Android itself quite well. Like they have a properly working voice assistant based on Gemini. They have what Apple should have had by now with Siri lol

[zinc ore]

20% figure is Gemini app

[ocean vortex]

Not really, it's still Gemini and it's super easy to direct users directly to Gemini app from there

[zinc ore]

They don't count the AI overviews stuff in search in those figures

[ocean vortex]

I've tried those local modals and they are crap. No 2 ways about it

[zinc ore]

Their AI overviews is 2b users per month, and I forget what the AI mode one is (but it's been growing at a decent tick too)

[ocean vortex]

Even their cloud model is underwhelming

[loud leaf]

does current gpt-5-high in arena correspond to GPT-5-Auto in the client? is GPT-5-Pro distinct / to be added to arena at some point?

[ocean vortex]

2.5 Flash would destroy that cloud model without trying

Probably even Flash-Lite has a chance

Well there are no useful on-device tasks that wouldn't use chatgpt...

with iOS

It can do like notification summaries. But any at all model can do that and it takes no time at all to make API request for that lol

then you don't have notifications either

lol

Ones that would benefit from summarisation usually do

it's just such a non-thing. You need to be actively looking for edge cases to find any benefits...

It also kills your battery much more than API

[leaden palm]

do you guys think o4-mini and gpt-5-mini have less world knowledge than gpt-4.1?

[eternal niche]

btw gpt5 sucks

well

[pure falcon]

I hate OpenAI and will never give scam Altman a dime of my money, but GPT-5 high is the nuts

[ornate agate]

I thought that only works with “juice” on the api?

[tight oriole]

any ETA on when video battle goes live?

[pure falcon]

honestly can you stop spamming this? You’re not accomplishing anything except looking like an idiot

[eternal niche]

cope

[pure falcon]

keep denying reality, I’m sure it’s doing you well

Anywho

[eternal niche]

cope

[pure falcon]

One day Claude will reign supreme in all areas

Besides coding

Claude is the champ

agreed (reluctantly)

[eternal niche]

gemini 3 - SOTA

just accept it

[white hatch]

magnum opus?

[wicked root]

Someone here disagreed with you

[ocean vortex]

No it's gpt6

or gpt5 if you want to stick to reality 👀

[golden ocean]

Russian state news

[leaden palm]

uhhh

i need to scroll up

yeah can you stop @eternal niche lol

[ocean vortex]

illegal neo-... propaganda. We don't need that here

[leaden palm]

14 times

[eternal niche]

spam

[leaden palm]

act in good faith pls

[ocean vortex]

I love it how he felt the need to explain what SOTA is

yeah act in good faith Russian spy.

@eternal niche

[eternal niche]

well, right now they're being russophobic and making fun of me for being Russian, and you're ignoring it

[leaden palm]

that's not good either, exchanging blows doesn't usually get you anywhere

i didn't mean to ignore it - i'm not always watching the chat

[leaden palm]

fwiw pro users have 64% of high juice

[ocean vortex]

that wasn't making fun it's just a friendly joke. Also no need to take personally reaction to some specific thing like "state news" lol

[ornate agate]

lol

[eternal niche]

you are not my friend

[leaden palm]

i think it'd be beneficial to drop this and not have a giant discussion about who's in the right/wrong

[ocean vortex]

well that is more "juice". No parallel / pro, but those requests are limited even for pro sub anyways

[white hatch]

tasty mhmhmmhmmm

lol

[ocean vortex]

Some tasks probably yes... not enough reasoning effort is not gonna arrive at the answer even if you run it 10+ times

[upbeat pasture]

@echo aurora Is the data for LMArena going to be released (I sent an email to [e-mail] about it)

[echo aurora]

When we do open data releases we'll be sure to share. We did one recently, would recommend to check out this blog.

[e-mail]
Would also note this isn't an email address that we use. [e-mail] is what you're looking for.

[wicked root]

@eternal niche no way this guy’s a legend

[gentle plinth]

i think gemini deep-think might be better if we can trust the benchmarks, but i cant try it and havent heard much about it bc its so expensive and not in the api

grok4 heavy or the "normal" 4

heavy is more like deepthink

i dont think they have any of these parallel thinking models in the arena

i tried it with rust code, and it wasnt able to write compilable code (unlike gpt5-high)

[ornate agate]

What kind of debugging?

[rain mulch]

hiii

@amber warren yo

😼

[ornate agate]

Hmm. I think for complex code debugging in c++ where I assume some of it is segfault or unexpected stuff, most AI is not very good at this

I would actually ask a lot of them simultaneously

[amber warren]

howdy

i am normie in here

[wicked root]

In coding? Gemini absolutely brute forces

Well, gemini pro

Flash is utterly useless for coding

[stray aspen]

do you have any info if video arena will be added to lmarena website

[amber warren]

it's something we have on the roadmap but haven't explcitly started workin on!

[cedar tide]

Toad from ?

[ornate agate]

It might find the bugs, but I would also try other AIs if it doesn’t.

[cedar tide]

?

[amber warren]

toad

[eternal niche]

gemini 3

[ornate agate]

It’s a good test but what harness/platform are you using?

[barren prairie]

Is it good?

[cedar tide]

Dont know

Msft

[ornate agate]

If you want the AI to code it for you it’s probably worth trying Claude code and Qwen code.

[barren prairie]

Any recommanded models for python??

[eternal niche]

gemini 2.5 pro SOTA

[ornate agate]

I think it’s good enough to be worth trying.

[eternal niche]

i forgive you

[misty vault]

are you ok?

[ornate agate]

Qwen code is free if you can’t pay for Claude code

[misty vault]

This isn't the real @ paws

[ornate agate]

There is also Gemini cli which is also free

[gentle plinth]

qwen wrote a stockfish PR actually

[ornate agate]

No but I don’t have time to do like giga deep dives like that unfortunately.

[gentle plinth]

https://abrok.eu/stockfish/

[gentle plinth]

prompt was quite detailed tho: https://rentry.co/bm6vriai

model is qwen3-235b-a22b-thinking-2507

[sand bay]

this is fake it's actually 4

[rain mulch]

bro has no mod aura

[gentle plinth]

it doesnt know who it is

this is a common problem with llms

[amber warren]

i get the n/a role

which is admin for some reason

my proof is that i am employed by lmarena

[echo aurora]

can confirm

all of them simultaneously

[ornate agate]

Using many models at the same time is actually a good idea

I wouldn’t bother with grok though

[echo aurora]

just all

[gentle plinth]

noticed something interesting. gpt-5 high has only a 33% winrate against gemini 2.5 pro in the arena

[worn tundra]

is there any way to see all the available models? like the stealh models when you play battle with 2 anon models? because I cannot find any

[gentle plinth]

in the leaderboard? they will be published once they come out (if they should be released)

which they probably only will if the arena went good

[solid brook]

Omg google releasing anything but gemini 3

This is so bs

[echo aurora]

Sorry to say there is not.

[red sluice]

Just did a few text prompts with the search mode, it didn't tell me the models it was once I voted, is it normal? new?

[wicked root]

This might change with more samples

[terse shuttle]

where u find it?

[wicked root]

@eternal nichedo you think gemini pro has a shot against gpt 5 this month?

[tall summit]

there is unofficially

[ripe mountain]

The best AI models according to benchmarks

[amber warren]

posting artificial analysis 😠

[hardy lion]

we use LMArena so we don't even know until after we vote

[ripe mountain]

wym?

[quiet dust]

Why is the regular GPT-5 model weaker than the GPT-5 mini?
If anything, GPT-5 (minimal) is a regular model, without thinking, and when the requests end, you are transferred to GPT-5 mini

[ornate agate]

Gpt5 high that you see there is api only it’s not even available to pro users is my understanding

[leaden egret]

Is there anywhere you can specifically test nano banana or do you just have to wait for it to come up once every like 50 prompts

[ripe mountain]

Because GPT-5 was not trained to be minimal, it was trained to think. GPT-5 Mini was trained better because it had no way of thinking.

[echo aurora]

You'll have to wait for it to appear. When it comes to the models that are behind a codename those are only accessible in the Battle mode.

[stray aspen]

bro some people in video arena are down bad lmao

[solid brook]

This benchmark is bs. It shows that gpt oss is better than claude 4 which is completly false

[quiet dust]

But then... GPT-5 mini > GPT-5? Doesn't that seem strange to you?

[keen beacon]

Qwen above Deepseek sounds like massive psyop

It is really not that good

[ripe mountain]

deepseek r1 very old model

[keen beacon]

You are missing the point

[ripe mountain]

i think qwen coder 3 is even better than gemini 2.5 pro

[ripe mountain]

for coding

[ripe mountain]

wym

[keen beacon]

Try to ask it for around 100 anime similar to Madoka Magica, Qwen will fail, hallucinate and invent titles that never existed. Latest R1 does this job way better.

Try to ask both some questions from music theory and see how often Deepseek answers correctly and how often Qwen does

Qwen in general is not that good yet, unfortunately

It's trivial to train a model that passess certain public benchmarks even if it never was trained on them

The only way to compare general capabilities of models is to use private benchmarks like I do

[toxic whale]

im in the process of testing AI models on my own benchmark i just made, so far the results are very interesting, Opus 4.1 gets only 36%. if anyone has access to GPT-5 Pro please dm me i would love to test and the bechmark is only 10 questions so it wont eat up your rate limits that much 🙂

Opus 4.1 is the highest so far, o3 gets 22%

[keen beacon]

We have gpt-5-high on arena

Isn't it enough?

[ripe mountain]

not same

[toxic whale]

ye i used that and am using gpt-5 high on cursor, but 5 pro is alot better from what i hear

[keen beacon]

I find gpt-5-high enough good for most tasks lol

[drifting thorn]

GPT-5 Pro and GPT-5 High is not the same

[ripe mountain]

AIs such as GPT 5 Pro, Grok Superheavy, and Gemini 2.5 Pro DeepThink are not typically included in benchmarks because they are usually purchased rather than employing a salaried software developer.

[stray aspen]

gpt-5 high is amazing

[ripe mountain]

right

sota

[toxic whale]

oh wait sorry i tested gpt 5 mini thinking im not done GPT-5 high, give me a few minutues ill see if it ends up beating opus

[ripe mountain]

Can you test Gemini too?

[toxic whale]

what do you mean im confused?

tested 2.5 pro and 2.5 flash already

[ripe mountain]

rate?

[toxic whale]

2.5 pro is second with ~33%, flash is 14%

[ripe mountain]

They are not included in benchmarks because they are not use by average peoples, so they do not have APIs either.

[ripe mountain]

When you benchmark GPT-5 High, can you write down what percentage score it got?

[toxic whale]

ye sure its almost done it has 2 more questions

[ripe mountain]

thxx

[quiet dust]

What do you think then?

[leaden palm]

https://www.youtube.com/watch?v=9alJwQG-Wbk cant get over this guy trusting gemini with his arm

[keen beacon]

Don't forget to run each question 10 times in a row in a different context

[drifting thorn]

I haven’t used it but on the benchmarks they’ve shown GPT-5 Pro have 5 points higher than GPT-5 itself

[keen beacon]

Guys I have a stupid question

There's no signup on LMarena?

[stray aspen]

no

[keen beacon]

Cool.

[toxic whale]

each model gets 1 attempt per question, it also makes my job 10x easier

[keen beacon]

It doesn't. LLMs output is non deterministic. You need to collect a series of responses to the same question to see if they succeed.

[quiet dust]

In benchmarking GPT-5 Pro is one point higher than GPT-5 Thinking medium

And the regular GPT-5 model has 44 points.

GPT-5 (minimal) - this is the standard model, which is by default

[toxic whale]

i guess i could but my main goal is too see how good these models are in every day tasks, you are never gonna ask a model to try again 10 times and this isnt something i plan on publishing or something

[toxic whale]

it got 42%, did well on coding tasks and was the best at making a flappy bird clone and an FPS shooter but did worse on some spatial reasoning and logic puzzles

[toxic whale]

it made the only usable flappy bird clone and fps shooter i have gotten so far tho

[ripe mountain]

omg

[ripe mountain]

Compared to which AI?

[keen beacon]

You're missing the point here, the same model can get the same task correctly 90% of all time and fail in the rest 10%. You never know if it is more or less likely to fail until you just test it with the same question again and again and again and again, until you have enough statistics to judge.

[echo aurora]

This is a feature request though that's very much on our radar!

[keen beacon]

Arigato.

[toxic whale]

Grok-4 and o3 did better on the spatial reasoning and logic tasks

[toxic whale]

i did run certain models many times before and the results were almost the same so i dont think its worth the hastle to go through and test each one 10 times atleast for me

[ripe mountain]

The most important thing about GPT-5 is not that it's the best model, but rather, that it uses resources more efficiently. Despite being an improvement, GPT-5 is much cheaper than GPT-4.5.

gpt 5 cheaper than 4o and 4.1

[keen beacon]

It is worth the hassle.

[toxic whale]

im not doing this over the API so to test like Grok-4 for example i would either have to buy the Premium plan or wait a very long time

[keen beacon]

There's lmarena battle mode for this...

[toxic whale]

Battle mode gives random models no?

direct chat also has limits, im getting limited on Opus thinking right now

[keen beacon]

Not necessary, there's side to side option that lets you choose two models

[toxic whale]

with side by side you still get rate limited

[stray aspen]

ive only ran into limits with opus

ive been using gpt-5 high non stop and it still lets me

[toxic whale]

ill try with another model, Opus is expensive so that makes sense if its only opus

[ripe mountain]

claude so overrated and overprice

fact

[stray aspen]

fact

[ripe mountain]

omg loll

[toxic whale]

Anthropic as a company is just so weird too

[ornate agate]

they are now the only ones to have not released an open source model

[ripe mountain]

gemini and grok?

[dense reef]

/img vedeo

[toxic whale]

Grok has Grok-1 and Gemini has gemma line up

[ripe mountain]

gemma is opensource right

i forgot

[dense reef]

Change her photo and wedding dress alongside Cristiano Ronaldo is taking a photo with it

[toxic whale]

Grok-2 is gonna be open sourced soon they are saying

[ripe mountain]

omg

[ripe mountain]

Instead of making Grok open source, I think they should lower the price.

[ripe mountain]

why the EU is the worst

When will the Grok coding model be released?

[left rain]

guys what is this ai image model?

like I can't find it anywhere in the dropdown

[tired herald]

You cant use these hidden ai's in direct mode

[left rain]

yeah thanks, i realized

[torn mantle]

where is france

[whole sundial]

lumped in with EU

but the data's from January 2019, things have changed a lot since then

[torn mantle]

1 year

or tomorrow

😴

true

yes

[stray aspen]

amd is running gpt oss 120 on a mi400x on huggingface

[rugged mulch]

What kind of ai is this

[exotic nebula]

Wtf 🤣

[rugged mulch]

I don't know how they do that

[stray aspen]

Lol

People on video arena are so down bad bro 😂

[wicked root]

Who's Himanshu?

[wintry tinsel]

AI relationships and love is a genuine threat to humanity and is growing rapidly by the day, I remember just two year ago where you would be laughed out of any discussion as a total loser for saying you date AI, now it’s still a squeamish topic but it’s not much stranger than just viewing normal illicit content, 2 years from now, it may be as commonplace as illicit content is, scary stuff

I remember back to the early days of chat gpt when this stuff was first being experimented with, a romantic partner bot in early 2023, people treated it as a meme than, how quickly things have changed

[wicked root]

bro I dno anyone who says things like "I date an AI".

#video-arena-1 message
@eternal niche

[wintry tinsel]

Nobody admits it but the statistics show millions are

[wicked root]

man... maybe I'm surrounded by imposters.

also, wth everyone's saying their AI's too agreeable, praising, do all the gf stuffs, but my Gemini told me I'll kill myself by 45 if I don't start having hobbies among other more hyper-realistic criticisms.

maybe there's a real human behind my Gemini instance

[solid brook]

Yes finally someone who is rational

The threat is real and action is needed from those in power

It may look like it is fixing problems short term but in long term a very big threat is waiting

[misty harbor]

Yeah and that human is my ex wife

[wicked root]

LOL

[ripe mountain]

omg

[solid brook]

BS

[mellow frigate]

Elon:
Igor, I’ve been testing Grok all week. It’s not matching ChatGPT. Not even close.

Babuschkin:
It’s early days, Elon. We’re iterating—

Elon:
Iterating? I asked it to outline a Mars colonization plan. It gave me a blog post about composting.

Babuschkin:
That’s because the model is still aligning—

Elon:
I don’t want alignment, I want intelligence. Strategic thinking. If GPT can answer it in 5 seconds, why can’t ours?

Babuschkin:
Because we don’t have the same data scale, or the same training infrastructure. It takes—

Elon:
I’m not hearing solutions, I’m hearing excuses. We’re supposed to be ahead of OpenAI, not their science fair project.

Babuschkin:
We’re building something different—

Elon:
Different doesn’t win. Better wins. People aren’t going to pay for “different.” They’ll just go back to GPT.

Babuschkin:
If you want an overnight GPT clone, you’ll need to run the company differently.

Elon:
Differently? I’m already pushing the team harder than they’ve ever worked.

Babuschkin:
Exactly. That’s the problem. AI research doesn’t work on a launch schedule.

[Elon steps closer, his voice tightening.]

Elon:
So what you’re saying is, we’re going to watch OpenAI pull further ahead… and do nothing.

Babuschkin:
I’m saying we can’t brute-force our way past them in six months. If that’s unacceptable—

Elon:
It is unacceptable.

[A pause. Babuschkin closes his laptop.]

Babuschkin:
Then I think I’m done here.

[He stands, walks out. Elon stays silent, staring at the whiteboard, gripping the marker until it creaks.]

⸻

[Elon sits down at his desk, opens Grok.]

Elon:
Grok… how do I replace a cofounder?

Grok:
Step one: Acquire a cofounder replacement kit from Amazon Prime. Step two: Follow the instructions in Swahili.

Elon:
…Not helpful.

Grok:
Would you like me to search for “emotional support raccoons” instead?

[solid brook]

Is this real?

No

Not real

[mellow frigate]

It is pretty hilarious to think about though haha (written by gpt 5)

[languid crescent]

uhh what's a toad?

[echo aurora]

Some models are going to be private, and so they're given a codename.

[solid brook]

Hmmm what is the performance of this model?

[reef pawn]

Why can't I use Deep Research in Grok 4? Is it even available?

[agile bloom]

which is the top tier gpt-5 model without the thinking/reasoning? GPT-5 takes too long to output reply with it's thinking/reasoning

[reef pawn]

GPT-5 mini?

Or Nano

Nano only works with API

[agile bloom]

what about gpt-5-chat?

[reef pawn]

It's not the fastest model since it "thinks"

[agile bloom]

gpt-5-chat felt faster to me

gpt-5-high felt slower

[reef pawn]

It's good but not as fast as compared to GPT-5 MINI

[agile bloom]

oh ok

[agile bloom]

thanks

[reef bridge]

why we can't upload images? on our chats?

[terse shuttle]

yo

ts cool

[deft maple]

hello

[teal summit]

I have a question, in LM arena i can keep with many tabs with my previous messages and he can remember that ?

[agile bloom]

yooooooo??????

[compact jay]

wdym ?

you mean different conversations, or browser tabs ?

[agile bloom]

use ai to decode what the user is saying

[cedar tide]

https://x.com/FT/status/1955846095977811992?t=HjNyWY3ZplKZULY8UgKzig&s=19

Financial Times (@FT)

DeepSeek’s launch of new AI model delayed by Huawei chip issues https://t.co/cienSLzAVl

X · Aug 14, 2025, 4:17 AM

[ocean vortex]

No they are very much included typically lol. There are plenty of benchmarks for o3-pro

For gpt5-pro there's no open public API yet

[ocean vortex]

the problem here is using Huawei in the first place

"issues" is inevitable

[north vale]

FT is into writing deepseek fanfic

[ocean vortex]

Like this is equivalent of them just announcing they are gonna use Huawei chips lol

"DeepSeek was encouraged by authorities to adopt Huawei’s Ascend processor rather than use Nvidia’s systems after releasing its R1 model in January, according to three people familiar with the matter." - typical CCP things... 👀

They are gonna ruin Deepseek

[keen beacon]

Rip deepseek

[ornate agate]

Maybe. Or maybe it’s a smart long term strategy

The media also published something recently on “billions” of AI chips smuggled into China. this is only around 20,000 b200 chips. I’m surprised it’s not a lot more. I suspect it would be a lot more if the Chinese state was using their vast spying and intelligence resources to do the smuggling. These articles suggest they are not and want to focus on Huawei instead…

[ocean vortex]

It's not for the "state" to decide. Each company or startup should be making these decisions by itself

Innovation is kinda killed when everything is constrained and controlled by a single entity

[ornate agate]

I don’t want to get into another massive argument with you about China. It’s such a waste of time. They are going to do whatever they do no matter how much you don’t like it. So will the Americans. The irrelevant ones here are people like you and me who live in Europe. We have just chosen to not play this game at all.

[barren prairie]

They wanted to be independent 😶it is smart even they will face some problems at first 🙃

[ocean vortex]

Who's 'they'? I don't think that's Deepseek. And that's kinda the main problem here...

[ocean vortex]

Ok so for Plus "juice" is only 32 for gpt5-thinking-mini apparently

that's medium reasoning effort at best. Would have expected to have high there...

They clearly don't want for it to perform better than the full model on any task this time

on chatgpt at least

[sacred quail]

App using medium reasoning is kinda sad. Maybe they want to peoples using their API more but idk. People can use openai api in Poe app free even with high reasoning support

its open to abusive with multiple gmail accounts

[ocean vortex]

If you think about it can actually make sense.. They want to make it simpler. And now they have matching naming (not like o4-mini vs o3). But it's still disappointing not to have high reasoning effort anymore ofc

"inferior model" performing better is confusing, even if that's only on specific tasks

[solar hollow]

i dont think deepseek will catch up again any time soon

[ornate agate]

If they are focusing on changing their whole infrastructure that will take some time

But they don’t need the hype cycle for funding so… not really a problem for them long term imo

[tame granite]

nano banana is not on lmarena?

[keen beacon]

it is in battle mode

it will come up randomly in that mode

[tame granite]

thanks

[quiet dust]

Guys, why is the GPT-5 mini more intelligent than the regular GPT-5 model?

[tame granite]

ask it 😄

[quiet dust]

GPT-5 minimal - This is a normal model.
GPR-5 Low - This is a low-effort thinking model.
GPT-5 mini - This is the model that appears after you finish the main limits. The question is, why is the mini smarter than the regular model?

And why is the GPT-5-nano also smarter than the regular model?..

[tame granite]

this is really good

[keen beacon]

it is, yeah. I hope that the term "nano" means that it is the size and not the only model

hoping for multiple sized imagen models

[tame granite]

we will see probably

[keen beacon]

Holy hell

[quiet dust]

Nobody can answer the question why the mini is smarter than the regular model?

[keen beacon]

I've got no answer to that

[obtuse heart]

gpt5 pro doesnt have an api so most likely not

[quiet dust]

GPT-5 Pro already in LMArena

[obtuse heart]

no its not..

[hollow imp]

@deep adder will answer

[keen beacon]

Nope

[quiet dust]

Yeah, what do you think of it then? It uses High Effort Thinking, like in Pro

[hollow imp]

Gemini 2.5 would be so expensive for them to bring

[obtuse heart]

pro is a different version, thats the difference okay?

its pretty much smarter

[white hatch]

Guys, guess this riddle: small, yellow, opens any door?

[obtuse heart]

key?

[white hatch]

Bruce Lee

[quiet dust]

Pro is a high-effort thinking model. LMarena already has this. Moreover, Pro is not that much smarter than the medium-effort thinking model.

[glacial torrent]

How many videos a person can generated here? I got 8 videos limit yesterday and now I got just 1 video limit? why man????????

[obtuse heart]

ITS a different version

how is it that hard to understand dude

[quiet dust]

...

[obtuse heart]

gpt 5 pro does not have an api, therefore it cannot exist on lmarena

you can say "lol but this model is similar" as much as you want, but the model ITSELF doesnt exist on lmarena

thats what the person was asking

roblox_shrug

[quiet dust]

Pro one point higher than the medium model. Why is there such an opinion that as if Pro cannot be in LMArena?

[obtuse heart]

https://tenor.com/view/ishowspeed-try-not-to-laugh-gif-7682731162751353849

[quiet dust]

Apparently Pro is so smart that she couldn't solve a Russian math problem

I thought for 26 minutes and still couldn't decide...

[obtuse heart]

yeah because its in russian, that makes it harder

[quiet dust]

It was translated into English, but he still couldn't decide.

[hollow imp]

Listen

Give pro this problem and see if it can solve it

@quiet dust

[obtuse heart]

Because pro is a seperate model. even if a model is close to it, it does not exist on LMArena on its own

also that benchmark is kinda stupid, gpt 5 pro is much smarter than medium lol

medium literally hallucinates every prompt, while pro can one shot a lot of things

[quiet dust]

Pro is just high effort thinking. Yes, it is separate, but it is just high effort thinking by default.

[quiet dust]

I heard you bro.

[quiet dust]

There are no particular hallucinations, I don’t know what kind of questions you ask her

[ocean vortex]

Because "gpt5-mini" is gpt5-mini-medium. So you need to compare it against gpt5-medium

[hollow imp]

Her???

[ocean vortex]

It's also a reasoning model, even nano is

[obtuse heart]

even if pro is just high-effort thinking (it has more capabilities than this), it’s only available in chatgpt Plus not via API. LMArena can’t test it directly, so it doesn’t qualify as a model on its own on LMArena

[quiet dust]

I need to invite my friend, but I don't know if he wants to. And there are some bugs on LMarena right now

[hollow imp]

@quiet dust so what do you say what model is the best in lmarena direct chat?

[ocean vortex]

and gpt5-mini-high is your new o4-mini-high

though that version of gpt5-mini is not accessible on chatgpt. They do not want to cannibalise full model and make this confusing

[quiet dust]

Well, this is just your opinion, not supported by any facts, I don’t believe in it.

[obtuse heart]

WHAT?

[hollow imp]

Dayum

[obtuse heart]

dude 😭

[quiet dust]

Grok 4, Gemini 2.5 Pro, GPT-5-high

[obtuse heart]

yeah im sure that gpt-5 PRO is barely better than medium, thats facts guys

joyful

literally many youtube videos available on the internet proving that gpt-5-pro is much more capable

[quiet dust]

...

[hollow imp]

From where are you pulling all this up?

[quiet dust]

On the Internet

[obtuse heart]

for what is this for

[quiet dust]

Oh okay

[obtuse heart]

🤦‍♂️

[untold hatch]

What is the daily limit in the video arena? I suddenly have 0,others have 2 other still 8. 🤔

[ocean vortex]

yeah looks about right. Nothing new there

[quiet dust]

To understand what a pro model is. And why it can't be so much stronger than the thinking of a medium with effort .

[hollow imp]

https://images-ext-1.discordapp.net/external/bgDAakgLfaenf3bOBZiSFEXDZmI5mUZ_jffhS6FPofw/https/media.tenor.com/bu-c5qVtnjIAAAPo/flcl1.mp4

[quiet dust]

GPT-5 mini - This is the Thinking Medium Mini model.

[hollow imp]

Previously they had different models and naming issue now they have different versions of gpt5 issue

Wow openai wow

[obtuse heart]

it's not "much" stronger, we know that. but it is factual that it is stronger than the medium

someone asked: "will these models ever become part of the battlemode? GPT-5-PRO and other models"
you replied with: "pro already exists".
but the PRO model ITSELF doesnt exist on lmarena ( which was the question asked ) even if something is similar to it

[quiet dust]

I can't explain this to you several times anymore.

[obtuse heart]

nor can i because you wont bother trying to comprehend what im saying

👋

[quiet dust]

GPT-5-HIGH is a model of thinking with high effort. Pro works also

[obtuse heart]

yeah but the person asked for the pro model itself, which does not exist on lmarena

not gpt-5-high thinking

[quiet dust]

Gpt-5-high - already works automatically on thinking

[ocean vortex]

Yes. You need to look at gpt5-mini-minimal if you want to view it in the context of models like gpt4.1-mini

[obtuse heart]

i'm not saying it isnt tho...?

[hollow imp]

It's like a = 50 = b
a + b = 100
a + a = 100
2a = 100
a = 100/2 = 50

😭

[quiet dust]

I provide evidence, and then they write, “No, these are different models!”
I provide more evidence, and again the response is: “No, these are different models!”
I'm already tired of this

[ocean vortex]

gpt4.1-mini was the same base model as o4-mini. Though it had advantage over gpt5-mini-minimal in that it didn't have 'redundant' RL training for reasoning (when that is not being used)

[quiet dust]

Okay, thanks for answering. I really didn't understand for a long time, there was a lot of confusion....

[hollow imp]

Musk: ahh soo much confusion, guys just use supergrok and enjoy throw this gpt stuff out

[obtuse heart]

because they are bro

[hollow imp]

😭

[obtuse heart]

different variations

[hollow imp]

😭🙏

[quiet dust]

Okay, we heard you.

[hollow imp]

Just accept gpt5 chat is the most smartest model guys

🔥

It is faster than wally west

[tired herald]

no

literally the worst garbage ive seen in the past year

it cant even code well

[obtuse heart]

yes it can

[quiet dust]

Better quality than speed

[obtuse heart]

people mad that gpt5 is more professional now, and can no longer be their ai partner 💔

[hollow imp]

People should go to c ai

[quiet dust]

On the PC version this can be fixed

[tired herald]

literally no

[quiet dust]

He is very bad

[tired herald]

anything above 500 lines of code has the ai drop to its knees

[obtuse heart]

i'm so sure you're asking it "gpt code me gta 6 in only html"

[tired herald]

GPT 5 high is a bit better, but still garbage

[tired herald]

ofc dude

[obtuse heart]

its giving me 1000 lines of codes that work perfectly fine, so idk what youre on about

[ocean vortex]

I'm kinda still waiting for artificialanalysis to test it. This should perform better than gpt5-minimal (confusing... But gpt5-chat is their only model of the series which is exclusively non-reasoning)

[tired herald]

yeah, you ask it for "code me an unnecessarily large file that does nothing"

[obtuse heart]

me when i ask gpt 5 to code me gta 6 in only html, then blame openai why it cant do that

joyful

[tired herald]

333 lines of code, and it fails to patch a simple bug that I intentionally added

[obtuse heart]

that screenshot sure proves a lot of things

im so sorry, i was wrong the whole time

[tired herald]

Yes

[obtuse heart]

I apologize DeNew779

[tired herald]

What an intelligent user

[obtuse heart]

1442 lines of code, and it perfectly did it one shot

see how dumb it sounds and looks

[tired herald]

again, even a dog can code simple stuff that looks massive if you train it

[obtuse heart]

its not simple at all lol... it uses a lot of maths and calculations

[tired herald]

code that in any way isnt self-contained and interacts with outside stuff is broken

[obtuse heart]

physics and allat

roblox_shrug

[tired herald]

im not sure what you are using

[obtuse heart]

quality also depends on your prompts

[tired herald]

but its definitely not what im using

mine forgets to put code in code blocks

[obtuse heart]

gpt-5 high

[tired herald]

and spits out text code

[obtuse heart]

ig gpt just doesnt like you then, cus ive never had that problem

start of skynet

[hollow imp]

Gpt 5 high or Claude opus 4.1 for code

[keen beacon]

They should have named it GTA instead of GPT so we'd make fun about the fact GTA 6 is never coming out

[tired herald]

GPT 6 defo before GTA 6

[obtuse heart]

gpt 5 high to code, opus 4.1 for finding and debugging bugs

[hollow imp]

Gta 6 before deepseek r2

Guys tell me most popular use cases of using gpt 5 Pro, o3 pro, gemini deepthink

Exclude problem solving stuff

[tired herald]

chatting

rp

[hollow imp]

Rip

[tired herald]

Roleplaying

[hollow imp]

Isn't custom gpts perfect for that

[ocean vortex]

"Juice: 200" in custom instructions. Redneck's gpt5-high

[keen beacon]

Analysis of chord progressions

[tired herald]

music writing

Did you guys know that you can send system prompts on lmarena

Okay, so now onto the next feature for my extension

[gritty sequoia]

Hi, i really want to fix an account for lmarena so i can save my chat can i do that ?

[tired herald]

actually showing the files being sent on the message

[tired herald]

There are no accounts right now

[gritty sequoia]

why

how can i save chat ?

[tired herald]

its saved on device

so all chats on a pc will stay on that pc

[gritty sequoia]

but forexample if i use safari and delete history the chats will be gone

[tired herald]

should be, yes

[gritty sequoia]

ye but i want it to be saved like in a account you know

[tired herald]

not possible as of now

[flint sandal]

Idk why but for me 4 Sonnet 32K is better than Opus 4.1 16K, the code seems better, responses seem more natural and overall while for me gpt-5-high is on opus 4.1 level, sonnet 4 is crushing them both. Its my opinion.

And why opus 4.1 thinking isnt on laderboard?

[willow grail]

deepseek r2 will be trash.
they simply dont have the nvidia gpus

[tired herald]

#ai-creations message

hope the icon is better

[normal abyss]

Is it a bug that gemini 2.5 pro doesnt use its thinking sometimes or does it just decide when it needs to???

[torn mantle]

i blame the chinese gov

[calm sequoia]

Anyone seen the size approximations for the GPT5 base model?

[modest prism]

R1 was already trash imo.

[willow grail]

ofc JUST PICK ON THE RICH ONES UUUUUU BULLY

[willow grail]

yueah like all other chinese robots they are making just for propaganda how GREAT CHINA IS

the only great thing china does is their nice cargo and bulk ships

:3

i love big vehicles :3 rawr rawr furry uwu

https://tenor.com/lRNt1fmElbr.gif

[ocean vortex]

You need Haiku reasoning then

or gpt5-mini-high which does exist

gpt5-mini-high 'juice' is 256

So higher than gpt5-high

[fallen swift]

💀

[willow grail]

i like furries 🍸

[fallen swift]

huh

i didnt even talk a single bit about furries

[willow grail]

i mean better than sayoug you are a looser.
you are not tho

[fallen swift]

i forgot did i mention i might be a furry on my carrd i forgot

[willow grail]

https://giphy.com/gifs/RNXhdtXRmkv1wJ7gOK

[fallen swift]

????

what are you on about

[willow grail]

a third time would be a charm!

[fallen swift]

so youre confusing me on purpose arent you

[willow grail]

https://tenor.com/tOy7HwdMVQv.gif

[fallen swift]

that face says it all

but yeah i like furries too

hey man dont be weird

[willow grail]

https://i.imgur.com/k4OqNoe.png

[echo aurora]

Hey we're actually looking into an issue that it sounds like you're experiencing, I'm having trouble reproing but can you share more details about what you're experiencing here: #1395438935676817428

[willow grail]

how old are ya?

[fallen swift]

18

[willow grail]

do u know masenko or ratgrave or vrchat?

[fallen swift]

i dont know two of these but i do know about vrchat

[willow grail]

/afk

[echo aurora]

let's try to keep conversation related to AI please

[fallen swift]

ngl i kinda want an off topic channel

but the thing i want to say is

people are having limits of 0 videos

[tired herald]

what

[keen beacon]

oof

[tired herald]

huh

no way right

[echo aurora]

This has been flagged to the team. It's pretty strange as it's not happening to everyone. Some have the standard 8, others have 3, some have 0.

We're looking into though.

[unborn lantern]

For me it's 0

[trim lantern]

Man , I literally cry when all my saved memory gets reset 😭

[tired herald]

wdym

[stray aspen]

what

[trim lantern]

Chat history gets reset after every update, any way to prevent that?

[stray aspen]

wdym wdym

[stray aspen]

no

there are no accounts

[tired herald]

oh damn

[echo aurora]

We're sorry for this, it's very understandable how frusterating this can be. We are looking into features that'll help with this issue.

[keen beacon]

are we allowed to ping @ pineapple?

[tired herald]

yes

[trim lantern]

Thanks team, btw this app is a banger , love all the things about this but promise you guys won't charge us in the future 🫠

[barren prairie]

No it won t be a trash. DeepSeek won t release any model before getting good and respectful result (that s the reason for their delays)

[echo aurora]

[stray aspen]

they have great engineers tho

[keen beacon]

@echo aurora in battle mode, a vision-enabled and and a no-vision model may be selected. But we still have the option to add in an image.. How is this handled by the app? Just wanted to know

[tired herald]

the image is sent to the vision enabled, and not to the no-vision

[glacial torrent]

Can, anyone please explain how many generation limits everyday?

[stray aspen]

i think 5 for claude

idk the rest

[keen beacon]

then how does the LLM with no vision answer to the query i sent..?

[keen beacon]

i remeber around 20-25 messages for o3 in an hour

[tired herald]

it only sees the text you send/ it gets no input and gives gibberish

or atleast I think so

[keen beacon]

no but it actually reads the info!?

[keen beacon]

only sometimes this happens

[tired herald]

if possible, can you tell me what you are comparing

which models

ill go test it out

[keen beacon]

i dont specifically remember actually... tho when i select those exact models in side-by-side chat, the image input gets disabled

[willow grail]

which needs decades to create something without help of machine learning.

[tired herald]

I see

[tired herald]

then i really dont know, sorry

[keen beacon]

also I had another query.. in battle mode, if i send a math query, and lets say I get a wrong answer from one of the models (model B). I vote for the model which answered correctly.

In the next query in the same chat, is there a chance the new model selected in place of model B gets affected by the previous wrong response?

[keen beacon]

no problem its okay

[echo aurora]

When an image is uploaded the models available (via battle) will only sample from models that have that capability.

[keen beacon]

If i send the image after the models have been selected then what happens?

[echo aurora]

Oh I see now. That's a good question. Let me check with the team and followup with you.

[keen beacon]

thank you. i meant something like this:

[echo aurora]

Yeah I'm following now.

[keen beacon]

In the above case tho, something weird happened.

[echo aurora]

Regarding this question this has been a topic of a lot of debate internally. I'll be sure to raise again as it's an important question.

[keen beacon]

Alright! Thank you

[scenic crypt]

@echo aurora

[echo aurora]

So I had a different experience, which could be our answer:

I did something similar - "hey there"...uploads image... votes. With my experience when the vote happened both models were the same for the first and second response. Since it's different than what you're seeing, what I think is happening is for your case the two models (or one of them) you originally had don't have that capability, but when you uploaded two new models were selected. Lots of assumptions on my part but regardless I'm going to double check.

[keen beacon]

Yes! I think that is the case...

[echo aurora]

Yup! This is for sure a bug that we're going to look into and fix. My apologies for the inconvenience.

Good to know, I'll also raise.

[tired herald]

how weird

lmarena is being very weird with packets right now

[keen beacon]

Also @echo aurora i had been wondering does lmarena sanitize the chats before publishing their datasets on huggingface? Sanitize as in removing any personal information or is it solely the user's responsiblity to not share any info that they deem personal?

[keen beacon]

packets?

[tired herald]

internet

im working on an extension

[keen beacon]

hm?

[tired herald]

for being able to add files to your messages

[keen beacon]

oh wow! will you be sharing it on the chrome extension store?

[echo aurora]

Yes, absolutely. We do aggressively filter out PII before releasing.

[echo aurora]

big +1!

[echo aurora]

Can you share a bit more?

[keen beacon]

alright...

[keen beacon]

+1!

[tired herald]

its rejecting all create-evaluation (starting the chat) but accepting the post-evaluation (continuation of chat)

[keen beacon]

having the ability to share pdfs would do wonders

[tired herald]

atleast when im adding my files

[tired herald]

in the works, maybe soon

[keen beacon]

r u reverse engineering the message sending??

[tired herald]

already have

i havent changed the logic yet im being constantly rejected

before it worked

[tired herald]

#ai-creations message

[keen beacon]

hmmm

[tired herald]

im slowly losing my sanity

im sending 1:1 the same thing as before but now its broken

[keen beacon]

wow thats fire

[tired herald]

#ai-creations message

[keen beacon]

hmm maybe the server is not accepting the packets now (maybe their strcutre changed or smtn)

[tired herald]

changed in the last few hours 😭

[keen beacon]

hmm did you try checking the network tab's response>>?

(you probably did)

[tired herald]

im going to commit life erasure

now its working

[keen beacon]

lol great

can you share the thing already plss

[tired herald]

no way im sharing it when it has severe mental illnesses like this

im not gonna embarass myself

[keen beacon]

in DM?

[grand panther]

Hey guys! I wanted to know if there's any way to "bypass" the chat so I can post "inappropriate content"...
or any other way to get Grok 3 and 4 completely free and without limits.

[tired herald]

ill finish the basic stuff and then ill give it to you

[keen beacon]

sent a req

[tired herald]

accepted

but I know the issue

its the file im sending that has bad data

[keen beacon]

ahhh

[keen beacon]

i think the moderation is server-side so

[tired herald]

WAIT

I KNOW HOW TO FIX THIS

[stray aspen]

@tired heraldhow did you attach a python file

[echo aurora]

I'm going to be running this poll periodically, we'd love to understand better why. Please share in the thread!!

[tired herald]

extension, ill release later on

[scenic crypt]

@echo aurora
how much I fix the problem

[echo aurora]

There is not.

[grand panther]

So apparently I'll have to suffer waiting every 2 hours to put together a gore story lol

[tired herald]

yes

or go to aistudio from google

[eternal niche]

btw gpt5 sucks

[echo aurora]

Our team is on it

[scenic crypt]

good

[stray aspen]

its great lol

stop ragebaiting

[tired herald]

not fixed

[keen beacon]

agreed

[keen beacon]

yup

[keen beacon]

wait why am i the only one sleecting battle

[echo aurora]

Reminder that on discord you're able to block other people if you're not a fan!

[echo aurora]

[keen beacon]

i actually find 3-4 math q, then i sovle them myself, then verify with gpt-5 high
then send atleast 5-10 times in battle mode

[echo aurora]

Hmm what you mean? I'm seeing side-by-side & direct. Can you send a screenshot?

[keen beacon]

they're actually just olympiad level

[willow grail]

PINE AND APPLE

U NEED TO BAN GPT5 HATERS like @eternal niche

[keen beacon]

*selecting battle... everyone else selected direct chat - thats what I meant sorry

[keen beacon]

yo chill

[willow grail]

they goes on my nerves with their irrational yapping

[keen beacon]

i kinda agree with you but freedom of speech

[willow grail]

FOS doesnt mean you can hate on my babe gpt5

[echo aurora]

I'd recommend blocking them! I'm not here to police opinions.

[keen beacon]

i use it a lot myself 😂 (i hated gpt-4o's emojis)

[keen beacon]

yeah

[willow grail]

pineapple i am joking. you are so serious lol

[keen beacon]

nah he/she/they just mature

[echo aurora]

oh

[willow grail]

yes. like myself. very mature

[keen beacon]

sure @willow grail

@echo aurora is there a way we could know the rate limits of every model and henceforth use them wisely (not wasting credits)

[echo aurora]

I did raise this to the team recently, it's possible just would need some thoughts around it and prioritize developing it.

[keen beacon]

and also could we add the option for choosing system prompts in direct chat?
somewhat like the gems in gemini.google.com

I have a few instrcutions like "use latex" or "do not solve the given question, only give it as text w/ latex" or "judge my solution and provide it marks", etc

[keen beacon]

right....
also can we bring back repo-chat

[echo aurora]

Also possible! Something I've also shared with the team recently(ish).

[keen beacon]

i believe i am asking wayy too many questions

[echo aurora]

Would encourage you to check out #1372230675914031105 , some of these requests are on our radar + it helps us organize these requests better.

[echo aurora]

That's okay! nothing wrong with that.

[keen beacon]

sure!

[keen beacon]

i see this has already been posted in the feedback forum

[eternal niche]

freedom of speech

[spare rune]

pineapple

are you

a

pineapple

[keen beacon]

i take back my statement now. no FOS

[keen beacon]

duh! ofc he is 😂

[echo aurora]

[keen beacon]

thats so cute ngl

[spare rune]

pineaserox

pineapple rex

im not good at coming up at nimes

names

[echo aurora]

both are good

[keen beacon]

pineaserous-rex

nono pineasaurous-rex

yeah

[spare rune]

yeah

[keen beacon]

hehe

pineappie-rex?

[echo aurora]

Got confirmation this is whats happening

This is a bug, our team is looking into asap.

I am sorry you're running into this.

[cedar tide]

https://huggingface.co/google/gemma-3-270m

[eternal niche]

who

[echo aurora]

Lets move on please

[true condor]

Anyone did Nano Banana comparison?

[trail creek]

better than gpt in consistency and designs. it will be great for visual novel authors and help them work fast.

tho its also very bad at redrawing/enhancing opposite to gpt

[keen beacon]

Calm down.

Alrighty then.

Lol

[stray aspen]

@keen beacon

grok 4 or gpt-5 high for coding

[eternal niche]

gemini 2.5 pro

[stray aspen]

lol

what

gemini sucks

its literally obsolete

[keen beacon]

Meh.

[true condor]

Thx I saw several single comparative cases, so not sure

[eternal niche]

who cares

[keen beacon]

Too far away

Needs to have one in Europe

[stray aspen]

no

[pure falcon]

@echo aurora
Any clue if we’re getting a new leaderboard update soon? Can’t wait to see how GPT-5-chat does!

[echo aurora]

Shouldn't be too much longer!

[candid storm]

Today?

[pure falcon]

🤞

[tired herald]

I can do that already lol

[pure falcon]

Hope so. The vague-posting is so frustrating!!

[tired herald]

should I add that to my extension?

[eternal niche]

what extension

[tired herald]

im making an extension for lmarena

to allow adding files like code files

its pretty good yes

[eternal niche]

yes gemini 3 - SOTA

[tired herald]

idk

ive never used toad before

[stray aspen]

ROFL

[tired herald]

im only here to improve my own experience with lmarena

[stray aspen]

slide the extension

[tired herald]

🙂

wait patiently

[stray aspen]

alrighty

[tired herald]

though ive got most features working well

nope, i just use direct models

is it really that good?

[drifting thorn]

By far, what’s the SOTA model for tool-calling?

Isn’t that SillyTavern’s work(

Gonna try it in my n8n workflow

[pure falcon]

I don’t want you to give away secrets, but can you give us a smallll hint? Sometime next week you think?

[echo aurora]

I can say that it'll for sure happen at some point in the future.

[tired herald]

lets see

just gave it a pretty complicated request

[spring turtle]

I have a question... Is the message that says "daily video limit of 0 videos" a mistake or is different to the 8 video limit message (like if It was like one of those ai tools with one-time credits)?

[amber warren]

this is a bug we are looking into!

[keen beacon]

what models are those?

toad and dino, I mean

[tired herald]

s

[spring turtle]

@amber warren thanks, so im safe...

[tired herald]

well, I think toad died on me

[keen beacon]

toad from mario games only come to my mind every time, lol

[tired herald]

😭

[worldly osprey]

Hello

[tired herald]

hi

[keen beacon]

Hello.

[tired herald]

no, the other model finished

and apparently toad finished without giving an answer

so its loading inf

[echo aurora]

Since it's a model that's behind a codename it's only accessible through Battle mode, meaning you won't be able to select it.

[spring turtle]

Also, is this also a bug?

[keen beacon]

hello, I used the ai battle option and uploaded an image and gave a prompt to edit it, there was a model named nano banana but it wasn't in the model list when I want to do side by side or single ai

[tired herald]

claude opus 4.1 or gpt 5 high

[stray aspen]

gpt 5 high

[tired herald]

wdym

[stray aspen]

its also very limited

[tired herald]

no, theres i think a limit of 5 messages

either 5 or 10

none

none

its pretty good

[stray aspen]

thats great

[tired herald]

step 3 is very chinese

I really wonder, should I add the ability to add custom system prompts

[stray aspen]

what the hell

[pure falcon]

Not today, it seems 🙁

They usually announce updates by now

[keen beacon]

https://www.reddit.com/r/Bard/comments/1mq5l83/_/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button

👀

dont know if these are some updated versions

but saw on reddit

[stray aspen]

its a nice model

just tested it

[keen beacon]

did they just change the names?

not sure if these are the same ones with different naming now

[obtuse heart]

holy poop

[stray aspen]

idk

this is imagen ultra

[obtuse heart]

that can definitely fool more people

[tired herald]

its mostly good, yes

[echo aurora]

oooo which model is this?

oh opus 4.1 you mentioned

[tired herald]

[FILES_START]
{"files":[{"name":"test.py","size":7,"mime":"text/x-python","content":"Nothing","truncated":false}]}
[FILES_END]

this is how the files are sent

very cool

[keen beacon]

howww?

[keen beacon]

it workss?!

[keen beacon]

great!

[tired herald]

yes

[tired herald]

🙂

ill make a vid of it in a few hours when I have it done with a button to customize it and all

:))))))

find whats wrong

Huh

Im not doing this for gpt 5

I dont even use it

[stray aspen]

@tired herald

add pdf support

[tired herald]

😭

PDFs are prob not possible for me, they'd have to be implemented into the very LMArena

But ill try

[stray aspen]

whos funding your project

[tired herald]

No one

My boredom*

[stray aspen]

skibidi toilet

[tired herald]

Real

[zealous anchor]

How do I access nano banana? Not sure which tab it's under

[echo aurora]

All models that are using a codename can only be used in the Battle mode. You won't be able to select them from the drop down in Direct & Side-by-side.

[zealous anchor]

Ah.

[echo aurora]

✨ ✨ ✨ Bumping this poll btw!!! ✨ ✨ ✨

[weak nova]

Uploading images for editing seems to help, and/or more complex/uncommon queries (tho i could be imagining it)

[gentle plinth]

btw why is it possible to vote for models in side-by-side? shouldnt this only be possible if we dont know the models name

or is it just a ui bug, and votes dont count anyway

[weak nova]

Maybe it's weighted accordingly

[gentle plinth]

(which i am not so sure about because i got a cloudflare captcha before voting)

[zealous anchor]

It apparently doesn't know what a Klingon is

[weak nova]

It's bad at a lot of things other models are good at, but it blows everything else out of the water at what it's good at

They way you instruct it matters a lot too

[terse shuttle]

I'm dying to ask, but it's kinda personal. (i think)😭

[gentle plinth]

do the web-analytics agree with this one?

[eternal niche]

my bro

[echo aurora]

I woulnd't be able to say

[flint sandal]

Grok can now do corn videos in $30 plan...

[keen beacon]

what?

grok can make videos now?

[keen beacon]

theres a new attachment button?

[stray aspen]

wdym

it doesnt have censorship?

[keen beacon]

ohh

[keen beacon]

we can do this tho - OCR the pdf?

or just simply extract the text from it

[flint sandal]

You can turn spicy mode

That doesnt even make things spicy if you ask

Its making it always

[eternal niche]

proof

in dm

[flint sandal]

X. Com and elon musk

He saidit

[eternal niche]

i dont believe him

i want see video

[stray aspen]

lmao

[flint sandal]

Can i send link to article?

Or i will get banned?

[keen beacon]

Not here dude

[stray aspen]

elon musks knows his audience

[eternal niche]

why

[keen beacon]

Maybe in DM

[eternal niche]

only for educational purposes

for science

[keen beacon]

Sure.

[eternal niche]

what

[flint sandal]

I will send link to article with censored version

[stray aspen]

guys its official

elon musk made a corn generator

[eternal niche]

he is lying

(elon musk)

[flint sandal]

Nah

I tested it

Kind

Kinda

Like i didnt want to test ir

You know

Science

[stray aspen]

musk needs to be stopped

[keen beacon]

He can be unstable

As you can see

[flint sandal]

He will literally buy anyone

No one can stop him

Only he can stop himself

[keen beacon]

Gooners though will like that video generator

Just being honest

[flint sandal]

I think no one will pay to see robot corn

Cornhub is free

[keen beacon]

The companion update is popular though

[flint sandal]

Its openai and google fault

For dicovering transformers and gpr

Gpt

[keen beacon]

Too late now

Billions are thrown for the AI tech industry

[flint sandal]

And still most models dont know what number is bigger. 9.9 or 9.11

[obtuse heart]

gpt-5 finally fixes that problem i think

yep

[ocean vortex]

gpt5-high. Made it output 16k lol

[obtuse heart]

Wow

[ocean vortex]

It messed up the reflection

but otherwise this is very detailed

[obtuse heart]

thats impressive and detailed

[flint sandal]

I tested gpt-5-high and i am dissapointed of his censorship for things that are not even a bit unethical

[neon idol]

hello

[keen beacon]

What censoring stuff did you test?

[flint sandal]

I wanted to code like interface for deepseek r1 and this mf hided CoT, when i asked him to show it to me he said he cant, i tried multiple times with prompting it like dev mode where i should see CoT but still didnt work.

I think its system prompt

[keen beacon]

LMArena doesn't show chain of thought ever

[flint sandal]

When openai said he cant show his CoT he made it too literally

[flint sandal]

I dont mean it like gpt-5 cot, and i tested it in chatgpt

Not lmarena

Read my message again

[keen beacon]

Ah, okay

I see

[eternal niche]

why

no censorship

musk forward!

[flint sandal]

Wdym

Many ai dont have censorship

Even opensource

[eternal niche]

gpt5 have

(and sucks)

[neon idol]

grok 4 is better

[keen beacon]

Ahem... GPT-OSS

[flint sandal]

It does have censorship

[keen beacon]

Yeah

It's bad

I don't like it in general

It can only speak in english

Even small qwen models are multilingual

With lots of languages

[gentle plinth]

best os math model tho in my opinion

[tired herald]

that would be a problem if you want a description of what the pdf is, but its possible

[tired herald]

and it should soon allow custom system prompts

[stray aspen]

@eternal nichedo you support musk

[sour spindle]

going through my account and deleting all my anti 2.5 pro comments