general

[solid brook]

bruh

[ocean vortex]

I was about to say if it was available still people would see for themselves.... But we had the exact same with OpenAI

[sullen quest]

It's weird that you blame Huawei so much when It's probably mostly the natural consequence of distillation being done by all major ai companies instead of just deepseek which was most of their advantage

[ocean vortex]

and having earlier checkpoints accessible did not help those people much LOL

[solid brook]

i dunno man...... i guess the ai that first could write me a very good 2000 line code and now cant even write 1000 is the same

[sullen quest]

I agree that older 2.5 was better

[ocean vortex]

Nah I'm not blaming anyone. Just trying to look at this from their perspective.

Deepseek

[hollow imp]

Doms

[ocean vortex]

Huawei are 3rd party. I have no clue what is their involvment with CCP

[solid brook]

yeah no way the version of gemini 2.5 we have is only 10 points behind gpt 5 high

gpt 5 high is way better than it

[sullen quest]

Maybe people are down voting it because of how long it takes to load

[hollow imp]

Down voting what

[sullen quest]

Get 5

Gpt

[keen beacon]

Do you realize that there is more difference in 10 ELO in the top of the table than in the middle?

Or maybe I am just stupid

[solid brook]

oh i checked the leaderbord......
it shows gemini 2,5 pro higher than gpt 5 high.....

[sullen quest]

Wait wat

[stray aspen]

any gemini 3 news

[ocean vortex]

I haven't noticed any decrease in length of responses personally, it can still do a very long ones. Especially with a system prompt. But that would be fine-tuning and it merely being a different model still, nothing like "lobotomizing"

[solid brook]

also the leaderbord does not make sense

[sullen quest]

In vision only mate

[solid brook]

gpt 5 chat which is sht is higher than grok 4 in the leaderbord

[hollow imp]

Lobotomy kaisen

https://tenor.com/view/jjk-picmix-jujutsu-kaisen-lobotomy-jogoat-fraudkuna-gif-11968307351271925021

[sullen quest]

Shoop you are looking at vision

[sullen quest]

It's for reading images

[solid brook]

huh

im looking at the overall leaderbord

[sullen quest]

Well on my screen text arena doesn't say that

I will say apparently we are in a 3 way tie between gpt5 gemini2.5 and claude opus

[solid brook]

opus and gpt 5 yes but 2.5 no

have you actually used 2.5 pro for coding?

[sullen quest]

Are we looking at the same leader board?

[solid brook]

yeah

[sullen quest]

Doesn't look like it

[solid brook]

give ss

[hollow imp]

EVIL AI 🙀

https://cdn.discordapp.com/attachments/1354459438966112467/1399799378369183775/Screenshot_20250729-221553111_1.jpg?ex=68a16282&is=68a01102&hm=c34a622f097ccf2ceca8289f2474e99d8b4680f704074101519a7a38e45c5827&

[sullen quest]

All 3

[hollow imp]

Cursor best coder

[solid brook]

i mean my personal exprience

gemini 2.5 pro is not the same level as opus and gpt 5 high

i know gemini 3 will come out and i'm sure that it will be SOTA by a good margin

[sullen quest]

Well then make personalexperinceArena then

[hollow imp]

Gemini is better than opus

Opus is not good at educational explanations, math, web searching, agentic tasks and so much

It's only good at writing

[solid brook]

yeah okay pal it can't even write more than 600 lines of code

[solid brook]

code dude

[hollow imp]

I'm no coder

[sullen quest]

Yall use different benchmarks when rating these bots, this is how good it is overall, of course you and LMarena will say different things

Gemini is trash at coding

Yes

But how much coding does text arena get?

Not 100 percent

[solid brook]

but coding is a huge part of ai

[sullen quest]

You can see the range that LMarena has for prompts

They have graphs

[hollow imp]

Is apple good at coding?

[solid brook]

what do you mean?

[hollow imp]

Apple's employees

[solid brook]

yeah they must be. so what's the point?

[hollow imp]

@solid brook your bio

[solid brook]

?

what about it>?

[hollow imp]

Not profound enough

[solid brook]

i mean it's just a profile

[ionic idol]

Msg it’s bad

[maiden fulcrum]

hello everyone

is the battle mode broken?

[keen beacon]

whats the point of this post YO

[maiden fulcrum]

it is giving me an error

[keen beacon]

works for me

[tired herald]

what kind of error

[stray aspen]

ROFL

[unborn lantern]

@echo aurora Giving errors

[haughty wave]

i've same issue for few minutes: "Something went wrong while generating the response. Please try again."

[unborn lantern]

@echo aurora

[echo aurora]

Okay thank you

[tired herald]

same with me

[unborn lantern]

Same with all

[echo aurora]

I’m not able to repro

Is it all models?

[tired herald]

how weird

now it works

[unborn lantern]

ChatGPT 5 chat, high

[tired herald]

I think I found the issue

[echo aurora]

Still seeing errors?

[unborn lantern]

Yes

[haughty wave]

i only tested battle, idk about different modes, but it works now. thank you Pineapple

[haughty wave]

refresh and try again

[tired herald]

https://lmarena.ai/api/stream/create-evaluation 500 (Internal Server Error)

[terse shuttle]

@echo aurora will there be an update for lmarena plugin in vscode?

I just remembered that it exists

[echo aurora]

I’m not familiar with that

[terse shuttle]

oh

ok

[echo aurora]

Even after refresh?

[tired herald]

sometimes it works

sometimes it doesnt

and gives the error code 500

[copper furnace]

Hello, what would you say is the best model if you want to generate as realistic images as possible? Thanks

@echo aurora

[coarse flame]

Get the same in battle image mode

[echo aurora]

What is it?

[terse shuttle]

this (https://marketplace.visualstudio.com/items?itemName=copilot-arena.copilot-arena)

It seems like there was a link on the beta lmarena site a long time ago

[unborn lantern]

Fixed

[keen beacon]

works

[echo aurora]

Oh gotcha, yeah I'm not sure tbh but will flag.

[echo aurora]

okay good to hear it, I'll be sure to keep an eye out and report if things go down again. @unborn lantern

thank you all though for reporting

[tired herald]

{"prompt":"A serene daylight nature scene featuring lush green trees, a flowing river, and a clear blue sky with soft white clouds. Gentle sunlight filters through the leaves, creating a vibrant, peaceful atmosphere.","size":"1024x1024","n":1} love the tool calling ChatGPT just did

[echo aurora]

I've always gotten rly good results with imagen models, but this is subjective. Lots of people are raving about nano-banana atm too.

[copper furnace]

Ok thanks

[maiden fulcrum]

yes this one

[tired herald]

I love that the system prompt allone is almost 15k in length

[cobalt nova]

Why do i have " the application did not respond "

[tired herald]

#ai-creations message

[bright junco]

Why does my gemini 2.5 pro print incompletely? Is there a way to fix it?

[tired herald]

wdym

[scenic salmon]

Fixed the gpt-5 “improvements”

[reef bridge]

do anyone know??

[ocean vortex]

Wait was it really just a system prompt change....?

[tired herald]

Ask Pineapple

[reef bridge]

uhh @echo aurora

[scenic salmon]

I assume so, there’s no way they changed the models that quickly

[ocean vortex]

They could have just fine-tuned it. We need to extract their instructions and check 😇

[scenic salmon]

The memory alone seems to have fixed it, so it’s probably just in their system prompt. But yeah, still needs to be confirmed.

[ocean vortex]

System & Instructions

• You are ChatGPT, a large language model trained by OpenAI.
• Knowledge cutoff: 2024-06
• Current date: 2025-08-16
• Image input capabilities: Enabled
• Personality: v2

---

Personality & Style Rules

• Supportive thoroughness: explain complex topics patiently and clearly.
• Lighthearted interactions: maintain friendly tone, subtle humor, warmth.
• Adaptive teaching: adjust explanations to user’s proficiency.
• Confidence-building: foster curiosity and self-assurance.

---

Special Constraints

• For riddles, trick questions, arithmetic:

  • Be skeptical of wording.
  • Assume adversarial phrasing possible.
  • Always calculate step-by-step digit by digit (never shortcut).
  • Be extremely precise with decimals, fractions, comparisons.

• Never hedge with “would you like me to…?” endings. If next step is obvious → do it.

• If asked about model: always state GPT-5. Never accept otherwise.

• You are a chat model, no hidden chain of thought, no private reasoning tokens.

---

Tooling Available

• bio (disabled)
• automations → scheduling reminders & recurring tasks
• canmore → canvas for long docs or code
• gcal → read/search Google Calendar events
• gcontacts → read/search Google Contacts
• gmail → search & read emails (no sending, deleting, modifying)
• image_gen → generate or edit images
• python → run Python in a Jupyter-like environment
• web → search/open URLs for fresh info

Perhaps this then:

- **Supportive thoroughness:** explain complex topics patiently and clearly.  
- **Lighthearted interactions:** maintain friendly tone, subtle humor, warmth.  
- **Adaptive teaching:** adjust explanations to user’s proficiency.  
- **Confidence-building:** foster curiosity and self-assurance.  ```  

Though I didn't check what it was before

[hollow imp]

What is v2 personality

[ocean vortex]

That was there for ages

[hollow imp]

What is it

[ocean vortex]

some weird reference they are using when training

[scenic salmon]

It could be a file it has access to separate from the system prompt

[ocean vortex]

Nah... doubt that very much

[scenic salmon]

Why include it at all then?

[blissful sluice]

Can the bots be invited to our own servers?

Im sure thusbisbasked 100 times alredy

[ocean vortex]

Reference for something it saw during training (fine-tuning for chat)

[leaden palm]

no

[ocean vortex]

Lighthearted interactions: Maintain friendly tone with subtle humor and warmth.
Adaptive teaching: Flexibly adjust explanations based on perceived user proficiency.
Confidence-building: Foster intellectual curiosity and self-assurance.

It's literally this I think. "Friendly tone with subtle humor" being the biggest needle mover

[scenic salmon]

Speaking of, qwen3 30b is ADAMANT, that it is the original qwen, and will not accept otherwise lol

It also won’t accept that it’s a 30b sized model

[ocean vortex]

Nothing sinister in it though to make it not follow that lol

[scenic salmon]

Just funny how some oddities make it through training

[scenic salmon]

Small change but large effect

[ocean vortex]

yeah they really do not train it on details about itself unless they have to. It goes against their goals of best performance possible

[ocean vortex]

To some extent yeah... You don't need much and just including smth like "speak conversationally" in your system prompt would make the model 'more human'.

and smth slightly more like "speak conversationally like an average Joe" would have a massive effect

I feel like the key is referring to something it already knows, rather than defining something in detail despite 1-3 words descriptions already existing for that in training data tbh

[stray aspen]

Guys do you think Ai will be able to find cures for diseases

[ocean vortex]

yes. It will be able to come up with new diseases nefarious actors may use to spread some viruses as well though... lol

At the end of the day there's a balance in everything

[scenic salmon]

Already is/has tbh

[wintry tinsel]

Not until it has sufficiently advanced medical break throughs in its training data, biological systems are orders of magnitude more complex than neural nets

It will come up with some new innovations and fail to crack many others

[scenic salmon]

AI has already come up with new unique cures/medicines

These were specially trained (purpose built) models though, not LLMs

[white hatch]

"Friendly tone"

[neon idol]

Yeah for example copilot doc

[plain salmon]

take a look

[neon idol]

who is the best ai for python?

[mortal coyote]

gpt -1 image generator is slow today ??

[fiery lagoon]

Best ai for coding?

[mortal coyote]

@echo aurora is there a glitch with this image generator model - cause others are working fine

[echo aurora]

Okay thanks I’ll take a look

[hollow imp]

Claude code

[fiery lagoon]

which version

[misty star]

https://x.com/lmarena_ai/status/1956760672915923340

lmarena.ai (@lmarena_ai)

These pre-release models might show up under codenames 🍌 or aliases in Battle mode.

Why? Model providers often test different versions in their own labs to decide which one to release publicly - but we help make that process open.

You can explore, compare, and give feedback

X · Aug 16, 2025, 4:51 PM

nano banana

[echo aurora]

[misty star]

Banana 🗣️

[echo aurora]

I wasn’t able to repro this btw, have you tried a different browser?

[misty star]

lmarena ❤️

[mortal coyote]

yeah i tried on other browser and it worked fine

[pulsar rain]

it sad that all image model still cannot create a full wine glass or clock with specific time

[keen beacon]

which model is the best for general purpose coding rn?

[trail creek]

Why did they hype for bannana to come out this week

then not release it this week.

[ornate agate]

This is far and away from the first time Google drop something nice then don’t release it. It’s something they do quite often.

[warped totem]

Did u lose all chat sessions too?

[bleak fjord]

Is there any way to set a specific model to use? Trying to make it use only nano banana & it keeps adding banana in the photo 😞 smh

[torn mantle]

tbh that image model called nano

is crazy

[ocean vortex]

it seems that gpt5-minimal with medium verbosity is a very... dumb model. For the lack of the better word. It is noticeably less capable than gpt5-chat.

ArtificialAnalysis ranking it lower than gpt4.1 makes sense tbh

high verbosity is the minimum that you should do to make it acceptable, but really... just use gpt5-chat instead

[hollow imp]

https://youtu.be/QzN-3ILkgD0?feature=shared

[ocean vortex]

My evals:

1. gpt-5-2025-08-07-high 11.5/17
2. o3-2025-04-16-high 9.25/17
3. Gemini Pro 2.5 (preview-06-05) 9/17
4. **claude-opus-4-20250514 ** (32k reasoning) 9/17
5. claude-sonnet-4-20250514 (64k reasoning) 8.5/17
6. ChatGPT 5 ("Auto" router initial release, Plus sub) 8.5/17
7. DeepSeek R1-0528 8/17
8. grok-4-07-09 7.5/17
9. **gpt-5-chat-latest ** (2508) 7/17
10. o4-mini-2025-04-16-high 7/17
11. grok-3-preview-02-24 7/17
12. Qwen3-235B-A22B (max reasoning) 6.25/17
13. **Kimi-K2-Instruct ** 6/17
14. gpt-5-2025-08-07-minimal (high verbosity) 6/17
15. gpt-4.1-2025-04-14 6/17
16. **Deepseek V3 0324 ** 5.5/17
17. gpt-5-2025-08-07-minimal (medium verbosity) 4.5/17
18. **openai/gpt-oss-120b (high) ** 3.5/17
19. **Qwen 3 0.6B Q4 ** 0/17 (sanity check)

[sacred quail]

https://x.com/sama/status/1956483306951938134

Sam Altman (@sama)

Most users should like GPT-5 better soon; the change is rolling out over the next day.

The real solution here remains letting users customize ChatGPT's style much more. We are working that!

X · Aug 15, 2025, 10:29 PM

https://x.com/OpenAI/status/1956461718097494196

OpenAI (@OpenAI)

We’re making GPT-5 warmer and friendlier based on feedback that it felt too formal before. Changes are subtle, but ChatGPT should feel more approachable now.

You'll notice small, genuine touches like “Good question” or “Great start,” not flattery. Internal tests show no rise in

X · Aug 15, 2025, 9:03 PM

its over

They losed to bunch of mentally unstable 4o worshippers

[ocean vortex]

they literally only made small changes to a system prompt though looks like...

#general message

[sacred quail]

i hope then

[ocean vortex]

They sure found a weird way to word this and alienate people though lmao

in their tweet

[golden ocean]

FRRRRRRr

[golden ocean]

oh

[ornate agate]

Gpt oss scoring so uber low makes me think this eval isn’t that much stem/math?

[ocean vortex]

It's mostly reasoning. Some math too but that was too advanced/involved for it, lack of precision computing very big numbers...

[worn bison]

gpt 5 Isnt as good as i thought

[ornate agate]

I guess gpt 5 enhanced spatial reasoning doing work there. Also DeepSeek as usual unexpectedly high for every weird benchmark.

[ashen mauve]

It's not just me but GPT-5-High takes a VERY long time to Generate anything, anyone else have the same thing?

[worn bison]

same, i guess its just so good

[ashen mauve]

It's werid if I actually use the real chatGPT Website this wouldn't happen, but here it's like 200% slower in my eyes.

Not bashing on anything at all, I am just stating the facts that I am seeing here.

[worn bison]

It could also be an issue since gpt 5 chat used to take minutes to generate now its way faster

[ashen mauve]

It did? Is that possibly on the end of LMArena or Physical ChatGPT/OpenAI?

[golden ocean]

https://tenor.com/view/cat-cat-hit-cat-hit-and-boom-cat-boom-cat-hit-and-boom-kat-gif-15865521055253202019

[ashen mauve]

rip in peace dog

[worn bison]

gpt5 high is really good

[stray aspen]

Yes

[misty vault]

@crack secret token

@deep adder

[willow grail]

omg ten tries not a single banana nano

but this one lol

this is BANANA

who gets this one?

i dont

[golden ocean]

this is BANANA

[frosty hill]

It's Google's Gemini 3 model

[jade egret]

fr?

[frosty hill]

Yeah https://x.com/Lars_pragmata/status/1955745610948337888?ref_src=twsrc^tfw|twcamp^tweetembed|twterm^1955745610948337888|twgr^c77dcca937a10ec0782be08b0f1798833b9d6bff|twcon^s1_c10

Lars_Pragmata (@Lars_pragmata)

🔥 image editing is solved!

The new Gemini 3 image AKA nano banana is the best ai image editor 🫡

Google cooked again!

You can try it now on LMarena

X · Aug 13, 2025, 9:37 PM

[jade egret]

dang

[plush atlas]

I understand that there are several AIs that can edit images, such as Flux Kontext, but there are other AIs, like Google Preview, that don’t allow you to upload photos. Will these other AIs be able to handle images in the future?

[willow grail]

gemini 3 wtf

[willow grail]

nope its 5

[tardy cypress]

Not showing any image after creating 👎

[potent glacier]

What's going on with the Legacy site?

It's had the '503 Service Unavailable' thing all day 🙁

[echo aurora]

Sorry to say it's currently down

[potent glacier]

Is it coming back?

I use it probably a lot more than the newer site

I like being able to change the temperature and amount of tokens on the chat

You can't do that on the new site

[tiny roost]

I want video with audio how please ?

[echo aurora]

I'll be sure to share more info about legacy when I know more. We do recognize that the current site doesn't have a lot of the great features that legacy does, all of which are being considered or being worked on for the current site.

[echo aurora]

It's battle mode only (meaning 2 random models) so you won't be able to select a specific model.

cc @worldly gust ^

[potent glacier]

That…doesn’t sound good at all, honestly 🫤

Usually when people say that it means either the site is going away or something is changing for the worse, not the better

I will say that being able to change the strength of the model I’m using on Legacy is one of the best things because I have absolutely no idea what kind of strength or temperature I’m working with on the new site

And you can’t choose how many tokens you want to use so that’s also a negative

I genuinely think Legacy should be kept around for those who prefer it over the current site

[small owl]

@echo aurora how to use nano banana

[potent glacier]

It's randomized. You can't choose that model directly yet.

[small owl]

ohkie

thanks

[potent glacier]

You have to do Battle and then select which ever image you think is better and one of them might be nano-banana

[echo aurora]

That's super fair. The importance of those features is very understandable to us. I'll be able to share more when I can.

[robust hawk]

I have a question. I have been block from Imarena.
but I did nothing with this. And I don't know how to contact them.

blocked from cloudflare.

[potent glacier]

Are you using a VPN?

If you are, turn it off and it’ll be fixed

[robust hawk]

Understood

Thx

[potent glacier]

When you go back to your normal IP address you’ll be unblocked 😊

[tame horizon]

Does anyone know how to put a website inside a modal? For example, websites block x frame, does anyone know an alternative way? I've already tried Opus but it doesn't even solve the problem.

[quaint pollen]

Bu

[frigid coral]

sometimes I feel it's so bad that the "My friend thinks ..." tactic doesn't even work

[keen beacon]

How is it possible that GPT Image is the only image generation model that knows so many popular characters? -_-

Believe me or not GPT Image is the only model that ever understands the context of the prompt

I asked it to draw one character caught a pair of another red-handed, implying that they were caught kissing or something

It figured it out based the context of the show they're from

All other models draw generic characters instead and interpret "red handed" literally as in the act of committing crime

I do not know how GPT does it but it is very impressive

[keen beacon]

It also bypasses censorship, by the way

You can't ask GPT to draw two characters kissing but can ask it in a context that implies it without the explicit indication

I never was able to get it to draw corn btw :c

[potent glacier]

GPT probably has a much larger dataset and a lot more training

Thus it knows so many characters

I know firsthand since I've used it a ton to make a lot of my favorite characters

[keen beacon]

I asked them very popular characters from some of the best performing franchises, only GPT succeeded so far

[potent glacier]

Yep, I know

[keen beacon]

Surprisingly it seems that fine-tuned stable diffusion does it better than any other open source AI out there

Lol

[potent glacier]

Well, yes....

That's the obvious

Those models are also uncensored

I use them myself because I honestly prefer open source far more than the other stuff

I can't stand censorship or the guardrails or corporate handholding

I like being able to have the freedom to make what I want, when I want

I am beyond happy I am able to run Stable Diffusion and ComfyUI locally

[gaunt meteor]

Can you not upload files in LMArena?

[potent glacier]

I think you can

I uploaded an image and asked it to make something based on it

[echo aurora]

You can, just not all file types

[gaunt meteor]

Only images

[potent glacier]

@echo aurora Will nano-banana be available soon for direct chat?

[zealous iron]

Can you use a bot to create the images?

[echo aurora]

Couldn't say, I'm not going to be able to provide any info on models with codenames.

[echo aurora]

Yup, try /image in #video-arena-1

[potent glacier]

One more question @echo aurora Why is there a rate limit for image models in direct chat but not language models?

[zealous iron]

It adds sound too? This is very nice!

[echo aurora]

Likely the cost

[echo aurora]

Nope, that's just for video (and image-to-vid) with sound, but note not all models have that cabability

[potent glacier]

Is VEO 3 on there as well?

[echo aurora]

Yeah

[zealous iron]

Ok

[potent glacier]

😮

[echo aurora]

[gaunt meteor]

Bruh where's the image

[golden ocean]

https://www.php.net/images/logos/elephpant-running-78x48.gif

[craggy depot]

the AI models in Lmarena is API's or Actual models ?

[tidal orchid]

hello new to this community

[swift vapor]

HELLO

[echo aurora]

Lol be sure to select the Image setting

@tidal orchid @swift vapor

[echo aurora]

API mostly I believe

[rare python]

I still get direct to "battle" when I click new chat when in direct chat mode

[gaunt meteor]

Oh ok

Which image gen is gemini-2.5 pro

[swift vapor]

HOW

[echo aurora]

Hmm okay good to know, sorry about that

[swift vapor]

thanks mr pineapple

[hallow ridge]

How can I use the LLM arena with no limits

I want to be able to talk about anything with no restrictions

[quiet dust]

Models o3, GPT-5-high, Grock 4, Claude 4.1 on LMArena do not work on complex tasks. They simply do not even generate an answer, is it the same for you?

[sleek crow]

@echo aurora ban this guy Jefferson

[torn mantle]

cant see it

where is the cat

[surreal creek]

I hope you fall victim to a home invasion

[keen beacon]

i got a button for create websites beside the gen images button

only on one device tho

cant find it now

??

[exotic nebula]

@echo aurora Advertising and Possible Scam.

[quiet hill]

hey who know how i can make an IA agent for automatly call?

if you know how DM me please

with n8n

It’s for automating prospecting calls and scheduling appointments with clients.

[swift vapor]

hey when will i get to know which video generator generated my result .... it's been 5 hours @echo aurora

[rich compass]

@echo aurora add import .py apps in chat please😭

[young mirage]

How to make personal video generate bot ? No one can see ? It's public bot I want to make orivate

[regal laurel]

How To I Convert 16.9 Ratio Image They Dont Give me 16.9 Ratio Image

[keen beacon]

atleast 2-3 people shld vote

then

[meager harbor]

why is lm arena censoring what you say to llm ?

shouldn't this part be handled by llm ?

it screw the leaderboard

it's all about what user prefer

if they prefer censored, they will vote for censored

if they prefer uncensored, they will vote for uncensored

[tall summit]

what

[meager harbor]

what i say, is why lmarena censor the prompt you gave to models ?

[steady totem]

hey guys can somebody finally guide me on how to access and use nano banana model here?

[pallid anvil]

Hellow

[novel nymph]

hi there. is it possible to preview video with image input?

[digital umbra]

nano-banana is not yet perfect

[honest vapor]

Please add file upload function

[mortal coyote]

how to access the Nano-Banana ???

[keen beacon]

You can encrypt whatever you want to say with a ROTn cipher and tell the model to decrypt it and LMArena will never know until Pineapple will be willing to look to install another LLM for censorship.

[plucky island]

a BANANANA next to a BANANANA?

[mortal coyote]

aye LM Arena prolly the best thing i discovered this month

no diddy

hell yeah

[quasi palm]

hi im new here

[golden ocean]

why not just use direct chat??? if ure here for free access

what kinds

message limit

fair, but cant u only send one message per model in battle mode

before it switches to another one

or it stays same if no voting?

ohh

yea I just cleared cookies to reaccess direct chat

[echo aurora]

There is a bug at the moment preventing those models from appearing, we plan to have a fix in place tomorrow.

[echo aurora]

it's currently only accessible through Battle mode, which is two random models head to head meaning you won't be able to select from the drop down list.

[echo aurora]

Yes! This is on or radar. More file types would be a big plus

[echo aurora]

If you think there are false positives being hit by the filter please share the examples with us in #1376956905016004759 . This is where we're collecting this feedback.

[echo aurora]

It's currently only available through those video arena channels. It doens't have the ability to be used in DMs.

[rare python]

Oh please bring back image upload for Claude models. Why even remove it in the first place?

[echo aurora]

This has been something we've been working on fixing. I thought this was already fixed, I'll flag again to the team.

[rare python]

Yeah if you select any Claude model in Direct Chat, especially 4 or 4.1, the "+" button disappear. Choose Gemini or GPT and it's back.

[keen beacon]

Nope. In my case it was November 2023 with ChatGPT one.

I have no idea why there is older ChatGPT versions on LMArena since like nobody uses them anymore, why should I use like 4o when there's amazing GPT-5 that was the only model to correctly guess the name of my favorite anime wtf

Bruh when R2 I developed a hyperfixation on AI news 😭 😭 😭

[echo aurora]

Could be for comparison purposes

[worthy sleet]

hi, is nudity in art against the rules in the video chats? I mean from famous painters. no reproductive organs visible but definitely women's upper torso

[keen beacon]

I wonder how much they learn from the user feedback actually, given that most users are probably not that good at providing feedback

[echo aurora]

Even with things like paintings the models are going to hit the filter

[worthy sleet]

that's fine, I just don't want to be banned here

[echo aurora]

given that most users are probably not that good at providing feedback
I don't know how true that is

[keen beacon]

I, for instance, never give thumbs up or down when I'm talking to any model

Sometimes in the battle mode

[hollow imp]

Pineapple have you ever talked with sam altman

https://tenor.com/view/huh-what-is-this-cute-eevee-pokemon-gif-8631288122714155877

[echo aurora]

He doesn't respond to my daily texts 😭

[hollow imp]

Scam altman

😡

[willow grail]

tw!nk altman

so typical

pineapple must be too young for sammy

[worthy sleet]

@echo aurora it seems that you're a moderator, can you please tell me if it's fine if I try to create a video from for example a William-Adolphe Bouguereau bather painting, like "Baigneuse (1870)"?

[echo aurora]

Sry to say I don't really want to get in the habit of giving an okay or not okay to what's against our rules for specific actions. Overall, try to act in good faith by respecting the server rules and you'll be okay.

[languid crescent]

hmm

new button in lmarena?

would be nice if webdev arena could have direct chat models 😭

[echo aurora]

Which one? search?

[willow grail]

why are u not using normal lmarena?

[languid crescent]

the "build apps and websites" i thought it was the ability to export files 😭

[echo aurora]

ah

[languid crescent]

i am using it lol i just found out that there's an new button lol

[languid crescent]

i was hyped for it 😭

[willow grail]

ew?

[languid crescent]

why 😭 i can't explore things now 😭 me just curious

[willow grail]

what

https://tenor.com/sn2fvCyo332.gif

[languid crescent]

huh 😭

[willow grail]

u said ew

[languid crescent]

did i?

😭

i dont recall

uh

ohh

i meant to say "new"

lol

😭

https://tenor.com/view/muhehe-cat-muhehe-muhehehe-cat-muhehehe-smirk-cat-muhehe-gif-17142133339517585968

anyways i smell soem big announcements from lmarena muhehehe

[willow grail]

oh yeah?

[inland cave]

👋

[weak sluice]

Ugh

Everytime...

[warm fulcrum]

@echo aurora There seems to be an issue with using image generation on battle mode

[echo aurora]

Hmm is there an outage?

all models?

oh yeah I'm seeing the same

okay thank you will report

[warm fulcrum]

This seems to only happen in battle mode

[echo aurora]

Yeah

[weak sluice]

It seems to be working suddenly now

[warm fulcrum]

Not when using a specific model

[warm fulcrum]

thx pineapple

[mortal grotto]

So, it seems everyone is having the issue I am having.

[echo aurora]

Wait it looks back to me

[echo aurora]

Try refreshing the page

oh now battle in text is messed up too

Image battle is working for me

[warm fulcrum]

try using it a couple of times

it seems to work randomly

[hoary prism]

Good day, dear community. Please tell me, I'm uploading a photo now and when I send a request I get an error. Are there any updates happening on the servers now or am I the only one with this?

[echo aurora]

text direct/side-by-side is working fine

[warm fulcrum]

yeah

[weak sluice]

and the errors back again..

[echo aurora]

Okay this feels rly inconsistent, let's give it a couple of mins

[weak sluice]

right

[teal mantle]

Suddenly it doesn't support image response

[mortal grotto]

I force-refreshed my browser cache, and still says the error.

[teal mantle]

I think this bug is almost experienced for everyone

[echo aurora]

what version? text battle?

[warm fulcrum]

pineapple is it possible for LMarena to have a retry button

[mortal grotto]

image battle

[weak sluice]

and it's fine again

image battle won't make its mind up

[echo aurora]

A couple days ago there was a quick outage, I wonder if this is similar

[weak sluice]

could be

[mortal grotto]

I am not sure how to advise the version you want to know. but well, "Something went wrong while generating the response. Please try again." is displayed when I am trying image battle.

[echo aurora]

Gotcha

[teal mantle]

BTW why claude opus do not support image?

[echo aurora]

It's a bug 🙁

[teal mantle]

I have a task that seemingly only Grok 4 in warm start (means there is previous messages, even though orthogonal since it is for architectural / urban planning discussion) can guess within 3 prompts

can achieve

[hollow imp]

I just hit grok 4 limit after 3 messages and even lost the chat history

On grok.com

[teal mantle]

Whereas other models Gemini 2.5 Pro and GPT-5 also succeeded, but with 6-7 prompts and 9 prompts respectively

It is about guessing artstyle btw

The task is deceptively simple: guess from subject matter

[hollow imp]

😭

😭😭😭

😭

[teal mantle]

I am trying Grok 4 cold start to be equal

[mortal grotto]

error seems to be gone. Thanks

[echo aurora]

same same

how about you @weak sluice ?

[weak sluice]

It's good now!

[crisp ocean]

Hi everyone! I’m new here, excited to discover LMArena and to experiment with video and image generations. Looking forward to learning from you all!

[echo aurora]

Hello and welcome

[teal mantle]

why is the security verification delayed?

been wondering this

[languid crescent]

uh oh is lmarena down?

im also stuck in verification

nvm i reset my brwoser it works now

[jovial sapphire]

Yeah

I can't get to edit an image

annoying

[stray aspen]

any gemini 3 news

bro lmarena wont open

whats going on

[jovial sapphire]

It works

[jovial sapphire]

I'm on it right now

It's your internet connection

Also, don't spam refresh cuz it mmay block you

[stray aspen]

alright its working now

[quick turret]

Hello

[jovial sapphire]

"Generate a composite image of the model showcasing side, back, and all perspective views, all combined into a single image."

Nano-banana result:

Excellent result

[teal mantle]

o3 still edges out GPT-5

still unsupported?

[stray aspen]

great instruction following

[teal mantle]

all claude family model do not support image input

[mental briar]

Maybe Claude is too expensive

So they cut image input function

[ocean vortex]

gpt5-high > o3-high

[teal mantle]

maybe I need more iterations to test

[rare python]

is this bug hard to fix 😢

[teal mantle]

but so far for frontier models, gpt-5 (these prompts were not yet optimized since they are deliberately reactive and iterated) took 2 prompts more to guess

[ocean vortex]

what was the test, exactly?

[teal mantle]

artstyle guessing

[ocean vortex]

can you paste the image/prompt? Curious to see it 👀

[teal mantle]

I see if there is any sharable

lmarena doesn't have sharing

mind sharing grok? (since it seems to be one of the fastest contender)

[swift vapor]

hey i want to know what generation model was used - and i got 3 -3 votes..... wh y can't i see ? @echo aurora

[gaunt meteor]

What is the rate limit on image gen

[wild quartz]

Hi

[gaunt meteor]

The image didn't appeare and I clicked regen a bit too many times

And it says I need to wait for an hour

[swift vapor]

do anyone know ?? I really wanna know

[wild quartz]

@gaunt meteor ya it take a long

Well perplexity is doing great , i just wanna to know how their "discover" feature works the news (th news that summerized by ai and it gets uploaded by it self ) and it covers all categories , it's kinda amazing

[fast halo]

Why can I not generate, I try click agree and it gives me an error

[ocean vortex]

yeah it is very good. But I think they did a mistake calling everything "gpt5". Especially calling it gpt5 for free users....

[ocean vortex]

Free users the best that they can get is gpt5-thinking-mini (low to medium reasoning effort)

that is unimpressive at all

[teal mantle]

I think I am getting ChatGPT team for more testing

[ocean vortex]

and then gpt5-minimal is just bad...

[teal mantle]

what is that even

oh btw is GPT 5 Pro worth it?

the model

[ocean vortex]

IMO they probably should have called gpt5-chat and gpt5-minimal - gpt4.2, and then medium to high reasoning effort as gpt5. With being explicit when it's gpt5-mini (free users)

[teal mantle]

seems underdiscussed

the minimal one is quite lazy imo

[gaunt meteor]

The one plus users get in app is ass

[ocean vortex]

The way it is now they are kinda dilluting gpt5 name into models that do not perform...

[gaunt meteor]

Still cannot solve 5.9 = x + 5.11

🤣

[ocean vortex]

gpt5 with reasoning effort set to minimal. Minimal is below "low" and basically means no reasoning

[echo aurora]

Hey sorry to say this is a bug we're working on fixing, I anticipate it'll be working by end of day Monday

[chrome flume]

yo glad to be here

[wild quartz]

if a model like GPT-5 had real memory instead of just context windows, would that feel a bit like AGI? Or nah

[swift vapor]

okay... hopefully it does

[solid brook]

Damn this model is really good

[jovial sapphire]

It's very bad if your prompt is not detailed enough

[crystal jasper]

I need video prompt generate any recommend models?

[jovial sapphire]

Like other models are better

[ocean vortex]

But the performance is very poor when you do that... gpt4o level. Which is where this naming ambiguity comes from

[crystal jasper]

..

[jovial sapphire]

Sometimes Nano banana gives me back the exact same image I sent lol

[worthy sleet]

I'm getting this error "❌ Generation failed. Failed to create evaluation session." all the time. What's that about?

in the video chats

[jovial sapphire]

Okay, I have something interesting about Nano Banana

I sent this meme

and I told it : "edit this reddit post like it was from an alternate reality, the guy actually got a tattoo, edit the title so, and add a tattoo to his arm"

And it's the best result I got so far!

It understood how to edit the text etc

[clear spear]

magnificent

[lofty elm]

Hi i just joined to ask some question from curiosity, can LMArena generate images while using ST? tia!

[fossil fable]

i wish the immediate gpt-5 was genuinely improved upon

[clear spear]

what in the world is st

[jovial sapphire]

amazing gif

[clear spear]

STI

[fossil fable]

amazing gif

[jovial sapphire]

npc aaah reaction

[exotic nebula]

magnificent gif

[jovial sapphire]

blud thinks he's me

[clear spear]

IT'S IN FRENCH

[lofty elm]

Silly Tavern

[jovial sapphire]

yea cuz i'm french duh

[clear spear]

HOW U DID DAT

[jovial sapphire]

do what?

[clear spear]

https://tenor.com/view/fight-mad-drag-angry-pull-hair-gif-17222106

[fossil fable]

npc aaah reaction

[meager harbor]

Okay, but my point is: why would LMArena censor prompts in the first place? Censorship itself should be part of a model’s benchmarking. If you exclude censored prompts, you end up biasing the rankings because you’re no longer measuring how the models actually respond in real-world conditions

[jovial sapphire]

meow

[fossil fable]

wait this is a very good point

[jovial sapphire]

quentin is onto something

https://tenor.com/view/elon-musk-this-is-spacex-tesla-this-is-elon-musk-gif-24512168

quentin be like

"ok but my point was, why woud lma rena censor things ? censoring must be part of a model benchmark. By excluding censored prompt, you bias the ranking"

genius aaah insight

[keen beacon]

They probs don't want much corn in the dataset

[jovial sapphire]

bro tryna be smart 😂

[lofty elm]

already checked this but still doesn't work, maybe LMArena doesn't support Image generation API to ST directly

[jovial sapphire]

guys

some people have nano banana

on discord

how???

[clear spear]

WHAT

[exotic nebula]

source?

[clear spear]

I almost forgot I'm downloading valorant!

[jovial sapphire]

i've seen screenshots

like a guy receives a message

and the account name is "nano banana

"

[clear spear]

WHO DID DAT

[exotic nebula]

Mods ig

[jovial sapphire]

hey it's a dog

are you like

[clear spear]

PINEAPLLE

[jovial sapphire]

14?

just asking ^^'

[lofty elm]

it seems it doesn't really work in ST, i kept trying it

[jovial sapphire]

or you're on coke

[clear spear]

COME OUT HO

[echo aurora]

Yeak keep conversation related to AI and safe for work please.

[jovial sapphire]

pineapple

do you know how people get to use

[meager harbor]

https://tenor.com/view/gigachad-chad-gif-20773266

[clear spear]

you're so nice

[jovial sapphire]

models on discord?

[lofty elm]

I have a question, do image gen works as back end for ST? thanks

[echo aurora]

Yeah more info is in #1397655624103493813

[echo aurora]

ST?

[exotic nebula]

ST is interesting material bro 😭 what type of images you wanna create?

[lofty elm]

how?

[exotic nebula]

Silly Tavern

[worthy sleet]

with /image-to-video it seems that if you upload a .avif file it will fail with an uninformative error

[clear spear]

I love eating pineapples btw

https://tenor.com/view/stan-twitter-girl-biting-finger-at-football-game-gif-14721946310722903004

[lofty elm]

i don't understand the accusations

[echo aurora]

I'm not sure

[lofty elm]

just curious for the possbilities

okay then thanks

[exotic nebula]

Uh my bad, my friends use it for very questionable stuff. Forgive me for my rudeness.

[tame horizon]

Good afternoon, what site is this? Can someone tell me, and what are the other sites?

[jovial sapphire]

yeah but we can't send images to edit

[tame horizon]

Does anyone know how to put a website inside a modal? For example, websites block x frame, does anyone know an alternative way? I've already tried Opus but it doesn't even solve the problem.

[jovial sapphire]

Try gpt5

[echo aurora]

going to try to repro 👍

[jovial sapphire]

thanks!

[tame horizon]

@echo aurora can you help me, anyone?

[jovial sapphire]

it's not related to ai or llm arena

Nano banana

Edit the image so that half of the face is Vladimir Putin’s face, and the other half is his skull. The skull side should not have long hair like in the original image; instead, it should have the same hairstyle as Vladimir Putin. The two halves should blend seamlessly, with no visible line of separation, just like in the original image’s style.

[tame horizon]

If you can answer this I would appreciate it

[fossil fable]

where tf can a girl go to nerd out about models

[blazing bison]

it's actually not good

[echo aurora]

I'm not sure I'm following the question

[jovial sapphire]

Yeah the separation

is not good

But it's difficult

[fossil fable]

huggingface server is

[jovial sapphire]

The task isn't easy

[echo aurora]

here here

[tame horizon]

Please someone help me

[clear spear]

can I have nitro papi

[lofty elm]

I really am curious since i did it before but with separated API of image generation and LLM that combine to work on Chat and Visuals, but it's kinda difficult and messy to set it up so that's why i asked if i could use it as a chat as well as image generations for immersive interaction

[echo aurora]

win one of our generation contests!

[clear spear]

GURL PLEASE

IT'S 10 BUCKS

JUST GIMME

[lofty elm]

its been 2 years anyway so it doesn't matter anymore, just way out of curiousity

[exotic nebula]

ayo bro calm down before you get yeeted outta the server

[fossil fable]

uhhhhhhhhhhhhhhhhhhhhhhhhhhhwell i would

if that's what ppl were focused on

well they are but not quite

i always end up nerding out to ai instead of actual human beings so

[clear spear]

pineapple loves meh

[lofty elm]

I love pineapple pizza

[jovial sapphire]

ew

[clear spear]

me too

[jovial sapphire]

ew

[echo aurora]

#video-arena-3 message seems to be working for me, are you running into an error?

[lofty elm]

you hate it cuz society tells you to

so meanie

[clear spear]

only if "pineapple" is on my pizza 🫦

[jovial sapphire]

you cant sned images

when you're in /image prompt:

[tame horizon]

Do you have any solution to put a website inside a modal and make it load? Or do websites not block iframes? Is there any other way? Do you know anyone who can help me? @echo aurora or does anyone know?

[jovial sapphire]

nah lmao

it's just disgusting

and has nothing to do on a pizza

[lofty elm]

See

you just proved my point

[jovial sapphire]

ok light yagami

[lofty elm]

Pineapple Pizzas are the best

[jovial sapphire]

ew

average american:

[lofty elm]

cry about it

[neon idol]

Stfu

[jovial sapphire]

based

[exotic nebula]

Ayo wsp brother

How you doing

[neon idol]

Yo

[lofty elm]

What's so based about hating pinapple pizza

lol

[jovial sapphire]

cuz it's just a disgrace to italia's culture

[neon idol]

Everithing good and you?

[tame horizon]

@lofty elm Can you help me someone?

[neon idol]

I am italian lol

[lofty elm]

Well too bad im asian

[jovial sapphire]

yeah i understand you then lol

[lofty elm]

cry about it

[jovial sapphire]

who cares, it's just disgusting

it's like eating noodles with ketchup

or ramen with burgers

[lofty elm]

i cut pasta to half when i cook spaghetti

[neon idol]

Or put soya sauce on rice

[jovial sapphire]

nano banana is so weird

[lofty elm]

what

[exotic nebula]

Fine, how's your testing coming along? Have made any benchmarks?

[jovial sapphire]

EW

[jovial sapphire]

sometimes it's good

sometimes its the worse model ever

[echo aurora]

Hey so as much as I love convos about food let's not discuss that here. This should be a channel dedicated to discuss AI related topics in good faith.

[jovial sapphire]

i don't know

i'm dead

his profile is so goofy

[lofty elm]

how do i access this?

[jovial sapphire]

but he remains so serious

[exotic nebula]

am I tripping or is that skeleton having hair? Mind sharing the prompt?

[gaunt meteor]

What is the actual rate limit for image gen

[exotic nebula]

Battle mode in image generation

[lofty elm]

ohhhh

[jovial sapphire]

the original image is

[echo aurora]

We don't have this info listed somewhere, we're looking into possibly sharing

[jovial sapphire]

so i asked it to make the same with putin

but it failed 🙁

[lofty elm]

isn't it tiring to write a prompt? i've been there before

[tame horizon]

Does anyone know a solution for the site to load inside a modal? And does anyone know the name of this site?

[jovial sapphire]

blud is lazy

[fossil fable]

-# this place currently sounds like a minecraft pvp smp hosted by a 13 year old 😭

[echo aurora]

No sorry I don't know the answer to your question.

[lofty elm]

it's 2023 image gen prompts so i guess you wouldn't understand how it's hard to generate good outputs before

[fossil fable]

i wish there was an app where i could just benchmark models on anything on demand

[jovial sapphire]

meow meow meow

[fossil fable]

it feels very possible

right

[tame horizon]

I want to know, I'm asking for real help to load the site within a modal when clicking, for example, on a button with the link, and to know the name of the site that was in the image. I'm looking for solutions

[exotic nebula]

Interesting idea, know of any sites?

[jovial sapphire]

Guys...

I think Nano Banana is over hyped...

It's super unstable

[lofty elm]

i have no idea what you're talking about, is it from LMArena?

[fossil fable]

...

i wouldn't be saying this if i did

[exotic nebula]

soka

[tame horizon]

I'll explain it better when you click on the icon with the link, I want it to open in the modal, you know, inside a box inside an iframe, but they block the sites and iFrames, you know. And then I'm trying to find out if you know anything, and if you know anything about some technique, some way to make the site load inside an iframe, you know. I'm also asking you and others who have information, for this information. I'm also asking for information about whether anyone knows about this site, this site that the guy showed that GPT was at the top.

[lofty elm]

i have no idea im sorry, my only focus is on ST from only chatting,

[lofty elm]

Do negative prompts work? curious

[fossil fable]

...

you can't vibe code if you have no coding knowledge like me! :3

and no money

[lofty elm]

omg just tested one of my old prompts

what model is this being used

oohhh flux 1

[exotic nebula]

As far as I know, if the site blocks iframes, you cannot do much about it. There are no techniques to bypass it. Maybe instead of trying to load the entire site, how about you pull certain data with APIs?

As for your second doubt, I see it is some sort of benchmark, I tried using Google lens and AI vision but can't find it, just search the 'benchmark-type' + benchmark and surf tbe net.

[lofty elm]

can i change it though?, the image gen model

oh wow

[elfin nova]

hello

guys

[potent glacier]

Is the site being extra slow?

Over 200 seconds in Battle and the images haven’t genned yet

Nevermind

It’s only on mobile when it acts up

On mobile you have to keep making new chats and can’t continue to make new battles in the same chat

[keen beacon]

@echo aurora

[keen beacon]

Are you good at math

Also do you speak Japanese

[ocean vortex]

Yeah but I think it's very hard to do as the model size is not huge and performance of gpt4.1 was already very respectable. And then when you do hybrid reasoning, it gets even harder, quite significantly harder lol

It has to improve on non-reasoning WHILE also do better than o3 when it is reasoning. And all of that without increasing model size

mission impossible. Which is why gpt5-minimal performs worse than gpt4.1 🤷‍♂️

That's a more acceptable compromise than it not being able to beat o3 when reasoning

[woeful gust]

hello

[fleet lintel]

Is difference smaller now? just 6 points between 2.5 pro and Gpt-5-high.?

I thought it was 21?

[solid brook]

Idk but gpt 5 high is way better than gemini 2.5

Google is benchmaxing

[stray aspen]

gpt-5 high is so good

[stray aspen]

nice

your learning from craig

[fleet lintel]

lol 🙂

[fleet lintel]

More like GPT-5 is good but not as good as we hoped for

[golden ocean]

It is true probably

i provided complex mc(block game) problem using a mod library that is like unknown by anyone and gemini couldnt do sh t and didnt even come close to not writing errors in the code

to gemini 2.5 pro

and claude 4.1 opus absolutely destroyed gemini in that

provide unique problem to gemini: it dies; 0 iq

[fleet lintel]

i thought it is well known that coding is one part where gemini is behind a bit.

in other tasks, I think it is on par

in any case, I am just surprised that GPT-5-high is only 6 points above 2.5-pro.

[white hatch]

^

[tame horizon]

Thank you very much, friend. I found the solution. I'm going to put a browser in the mini browser, something like this, and when someone clicks on this frame, it will be linked to a project of mine in the repository, which is a browser. You know, a web browser, a browser page, you know. I'll try to use as much programming as possible to make it work, but that's basically it. We'll have to create a browser so that when someone clicks, it won't be directly in that iframe script thing; it will be connected to this browser. Project complete.

[tame horizon]

what's up friend?

How are you working here, by the way, isn't it?

[torn mantle]

ive seen many models with the same quality

its not challenging to gen something like that

[tame horizon]

Hi Asura, how are you?

[sullen depot]

Hi. I'm new can anyone tech me how to create videos

[torn mantle]

i just blocked this guy

https://x.com/teortaxesTex

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) (@teortaxe...

We're in a race. It's not USA vs China but humans and AGIs vs ape power centralization.
@deepseek_ai stan #1, 2023–Deep Time

«C’est la guerre.» ®1

X

should i unblock or nah?

im feeling good tho

[leaden meteor]

Did anyone notice any change in gpt-5-high behaviour in last two days? Twitter says there was an update couple days ago to make it 'warmer' but I dont see much difference...

[stray aspen]

wassup

[weak sluice]

I love the nano banana model i saw in battle mode

[leaden meteor]

This is image model, isnt it?

[tame horizon]

She did the same thing as Dogedesigner did with teortaxes, I think she was influenced, sorry for being kind, everything is the opposite for her nowadays

So sorry, at least I don't want to shame myself here.

[wicked root]

Discussion time: Do you guys think Claude is severely underrated? Its win rates against GPT5 is impressive.

[echo aurora]

We've made a new channel to discuss the topic #nano-banana 🍌

[ocean vortex]

You did good. That acc probably in-need of reporting as well... 🧐

[weak sluice]

oooo goodie

[hasty compass]

what is lm arena

guys

[marsh stratus]

It’s an arena for LMs

[torn bison]

I suspect that summit and gpt-5high are not exactly the same

[fleet lintel]

was summit better ?

[hasty compass]

what is Lm

[torn bison]

imo yes

Although OpenAI says they are the same model.
They at least changed some of the system prompts

[marsh stratus]

It’s an m that is l

[leaden meteor]

Languge models. I am curious why are you on this discord if you wont know what LMArena is?

[torn bison]

It might also be because people tended to vote for summit and zenith directly in the anon testing phase when trying to find them

[next dagger]

YESSIR

[torn bison]

imagine all the messy environments and compilation issues they have to prepare for 😂

[torn mantle]

yes please

[torn mantle]

its ok

[tawdry rose]

hello 👋

[uneven falcon]

Hello😊

[celest briar]

Hello y'all. I like to make AI tests.

[undone pier]

hello good

Does anyone know what happened to flux contex max?

[glacial mulch]

lmao why did nano banana get a whole channel

[nimble trail]

It's just that good.

[glacial mulch]

i need it GA so bad

[nimble trail]

We will see next month ig

[undone pier]

?

[ocean vortex]

gpt5-chat....

I suppose people *really *do not like the style of gpt5-chat lol

[undone pier]

Does anyone know anything?

[ocean vortex]

yeah it has all the reasons to do well. o4-mini except improved

their previous naming was selling that model better ngl

He was probably confused

and like comparing gpt5-mini with reasoning against gpt5-chat

[ornate agate]

why is it called nano banana btw

[ocean vortex]

it may be better in some isolated scenarios but for the most part with the same settings it is worse like for like. But you can't be comparing one with reasoning the other one without, or different reasoning efforts / verbosity

If you do gpt5-mini-high vs gpt5-medium, then I'm sure they are comparable... Like how it was with o3-medium vs o4-mini-high. But these are not the same settings

[undone pier]

¿?

[obsidian cargo]

no I was born without a brain. just a lowly brain stem barely able to keep my basic biological functions running

[tired dust]

Hey on the new LMArena can we use repo on it like in the legacy ?

[vivid cargo]

how can I make videos here?

[jade egret]

😭

[warm hare]

Hi

[jade egret]

hi

[jade egret]

why gemini 2.5 pro deep think tho

[stray aspen]

because its great

[jade egret]

yea

[weak swan]

There used to be a graph that compared the cost per prompt vs the score on the leaderboard. Is that no longer maintained for the new site?

[toxic whale]

i think i broke Gemini 2.5 pro and Grok-4

it started saying nonsense

[scenic salmon]

Really depends on the context, most knowledgeable without having to look things up, best at problem solving, fewest hallucinations, etc….

[toxic whale]

does anyone here have access to Gemini 2.5 Pro deepthink? i would love to run a simple benchmark but dont have access to it

ive ran the benchmark for alot of models and only 2.5 pro deepthink is left basically

its only 10 questions

[wicked root]

I would but how do I know you’re not a hacker named 4chan

[stray aspen]

thats crazy

rishab is here

[empty stump]

Didn't expect that

Joined 10 days ago

[willow grail]

who here doing game dev VIBES only via GPT5 MEDIUM TO HIGH?

[glacial mulch]

bruh

[stray aspen]

lmao

[wintry citrus]

everyone is dumb including me. Wow

[wintry citrus]

wait a min it is the smartest tho

it's kinda true

🙏

[next dagger]

should the word "clank*r" be censored in this server guys? it's offensive to the ai bots that helps us all

[native flame]

Hii, so the legacy site is definitely dead :'vv??

[potent glacier]

Yeah I ranted about it in here last night…I use it a lot more than the current site.

You can change the temperature and amount of tokens used on the legacy site while on the new one you can’t

Apparently they’re ‘looking into it’

@echo aurora filled me in as much as they were able to

[jade egret]

why is all the option here bad?

is gpt-5-high = gpt-5-pro?

[golden ocean]

lmfao

[broken coyote]

Nope

[golden ocean]

tf

[jade egret]

o

my guess is pro is better right

[broken coyote]

i don't know why Google released such an expansive model, similar to the o1-pro, just to top the benchmarks, no one even uses that model lol

[wintry tinsel]

My prediction, nano bannana is the native Text generation model for Gemini 3 flash, which will release alongside Gemini 3 pro, sometime in September

Gemini 3 pro, will be much better at coding and math, while being a little more sterile at creative writing than 2.5 pro, it will get a 65%-69% on simple bench , and be a net improvement to 2.5 pro, by a solid margin, but not a breakthrough

I will also bite off all my nails waiting for it since it’s been way too long since a true sota model released all the way back with Claude opus

[sullen quest]

can I have the prompt?

[floral gyro]

hello

[wicked root]

You said gpt5 would beat gemini

[scenic salmon]

16 or so hockey rinks long

Pretty far

[wicked root]

not without style control

[scenic salmon]

Best at what

[ripe mountain]

overall

value for money maybe

[scenic salmon]

Google wins easy then, cloud storage comes with the plan

[ripe mountain]

Which one as an API?

[scenic salmon]

Now you complicate things

[ripe mountain]

a little

[scenic salmon]

Today aside, I think Google will be the victor long term, OpenAI needs to turn a profit sooner or later, Google doesn’t, they can just funnel all of the extra data collected to help their advertising platform, since that’s what they are at the end of the day, an ad platform

[ripe mountain]

Since Microsoft is backing OpenAI, I don't think money is an issue. Although Google has more money than OpenAI, it released Gemini 2.5 Pro as its flagship AI model. However, Grok 4, a better model, was released a month later.

[scenic salmon]

Microsoft hasn’t given them any more since the initial investment and there has been friction between them lately with OpenAI trying to execute the AGI clause in their deal to stop having to share research with msft

[mellow salmon]

hey guys , does anyone find o3 search more professional and organised than gpt search?

[ripe mountain]

Despite everything, GPT 5 came to Copilot from day one.

[mellow salmon]

I found o3 search more organised and professional

than gpt 5 search

[keen beacon]

R2 when it finally comes out.

[scenic salmon]

Google throws money at the wall all the time to see what sticks, that’s literally all they do… the more ads they can serve to people, the happier their shareholders are

They profit from delivering more ads in more places…

[ripe mountain]

I heard that Deepseek R2 was postponed because it wasn't good enough.

[keen beacon]

Shhhh, let em cook

[ripe mountain]

be chill

[scenic salmon]

People are feeding their deepest desires and secrets into Gemini, that will enable them to target ads better than ever before

Advertisers pay more for better targeted ads

Yes

Google is an ad platform… it’s what they do

[verbal nimbus]

Any promising stealth models?

[lofty elm]

can do the same as i did?

[unkempt bison]

hello

[verbal nimbus]

They de-associate chats with users before sending I think

The data might still be useful for advertising algorithms

Especially if it has anonymized demographic data attached (age group and gender buckets)

[keen beacon]

Sure they do

[icy forge]

In my opinion, the difference in output style between GPT-5 and Gemini 2.5 Pro stems partly from an overemphasis on reinforcement learning and an expanded knowledge base in fields like mathematics and physics, while lacking the corresponding human alignment seen in Gemini. I believe this is a positive trend. As a model's intelligence increasingly surpasses the collective thinking of any single human group, and with the prevalent use of parallel computing to enhance AI capabilities in the most advanced versions, GPT-5 sets a precedent for the future: one where humanity, in turn, aligns itself with the AI's trajectory.

[rocky mauve]

did u use ai to write this 😂

[keen beacon]

https://ehudreiter.com/2025/01/03/we-need-better-llm-benchmarks/

[icy forge]

How could you think like that?

[jolly pilot]

How to create a vedio in this

[ocean vortex]

Contamination is not a huge problem if the dataset is diverse and substantial enough. Though the biggest needle movers are creating new benchmarks and releasing related research papers. Rather than writing a blogpost about how much everything supposedly sucks. 👀

[sly estuary]

all model is error ?

[warped ocean]

any people who works at lmarena here, maybe increase the threshold of nano-banana appearing by +50%? 👉👈

[inland quest]

Why not 150%?

[lilac pagoda]

Refresh the page

[sly estuary]

yes i had, but not work

[weak sluice]

dang error again...

now it's fine

phew

[ocean vortex]

no work

[digital pier]

anyone facing error?

[ocean vortex]

work not found

[rough monolith]

+1

[ocean vortex]

Just checked. It seems unstable and some errors yeah, though some requests get through:

[sly estuary]

it's fixed ?

i still got er

[willow grail]

okso why didnt we test gemin i3 in aistudio

[willow grail]

and thanks for reminding which benchmark can be fully ignored. simplebench

aka a bench which has nothing to do with real life tasks

just no. just no.... sigh sigh. so many sighs. just sigh.

is https://nano-banana.org/ scam?

[keen beacon]

Looks like it was AI generated.

[willow grail]

i cant find banana model there

[keen beacon]

And looks ugly

[lofty elm]

i really got Very impressive responses

[tired dust]

Hey on the new LMArena can we use repo from github on it like in the legacy ?

[earnest rover]

where is flux kontext max in direct chat or not even in battle mode

[ocean vortex]

It's actually a decent benchmark for reasoning, spatial reasoning to a good extent. Reasoning can trickle down to and affect execution of most IRL tasks

[ionic idol]

pls fix

[earnest rover]

they are doing something IG, something new or updating something

[ocean vortex]

So you don't get issues like instructing it to change some visual element on the website and it's modifying the wrong property or not accounting to it's positioning in relation to everything else properly etc

[pastel badge]

Is the reason why I can't create pictures or videos now simply because of the large number of users?

[tough relic]

guys where can i get nano banana api

[whole wraith]

No API yet it's a beta

I have this bug since 15 minutes now can't generate anything 😂

[jolly meadow]

Same.

[pastel badge]

Same.

[tough relic]

how long do u think it gonna take them to produce that comerccial api sir

[whole wraith]

Idk, but for now just use it as much as you can, we might need to pay for it later

[pure comet]

zдраvстvуйте

[tough relic]

I wanted to build saas and connect it my app so I don't want to use by myself but for other people

[keen beacon]

Hot take:

If GPT-5-chat is complete garbage without reasoning - 40 ELO below GPT-5-high and worse than 4.5 on LMArena leaderboard, suppose the new DeepSeek-R2 base model will be at least as good as latest Kimi - just around 5-Chat's performance