general

[verbal nimbus]

They work on both at once, I think it was the head of Deepmind that explained when and how they start training new models

If it's true, then they're post-training rn

[mild nimbus]

[cloud zinc]

yes, prob a dec launch at latest

[verbal nimbus]

https://www.reddit.com/r/singularity/comments/1nnzaev/gemini_30_pro_is_now_being_ab_tested_on_ai_studio/

[barren prairie]

Gemma 4 not a flash it is bad

[cloud zinc]

that guys posted "Google source tells me", how is he trusted

[verbal nimbus]

Idk if true though

[verbal nimbus]

The one who says it had a successful training run?

[cloud zinc]

no, the reddit post guy you sent.

[verbal nimbus]

It's from inspecting the network logs

But idk about their assertion that v4 = 3.0 and tier 7 = pro

[cloud zinc]

the screenshot is true, but he comments that V4 = 3.0 and tier7 = T7 = the size class for 3 Pro. v4 could be new gemini 2.5 flash model

[verbal nimbus]

It showed up when 2.5 Pro was selected and apparently outperformed it while being faster

[cloud zinc]

maybe there is a new 2.5 pro, like gemini 2.5 flash

[verbal nimbus]

nightride probably

I don't think it is a 2.5 checkpoint though

[cloud zinc]

i think v4 prob refer to 2.5 edition. cuz there was three (3-25, 5-06, 6-05(stable) versions)

[verbal nimbus]

Interesting theory, does make sense

[magic stag]

Nightride?

[verbal nimbus]

Apparently it was very fast though, which is why people thought it was Flash

[verbal nimbus]

It's an anonymous model

[magic stag]

Its funny I saw that thread right before u posted it

[verbal nimbus]

I think it's nightride-v2 now

[magic stag]

On bard subreddit though

I see some people making comments like this:

[verbal nimbus]

It starts off almost identically to 2.5 Pro

Like word for word

[magic stag]

Wait nvm same thread

People using 2.5 pro same as always but getting much better responses and much faster apparently

[verbal nimbus]

Whatever they have going on sounds promising

I just wish that the app was as good

[magic stag]

Lol

[verbal nimbus]

Like sometimes the models literally forget the conversation after each message

[magic stag]

App is trash for me

Unusable

[verbal nimbus]

Yeah

If it was like AI Studio and support MCPs, I'd get a subscription

But it's sketchy as hell lol

[limber crag]

Big if true

[magic stag]

I have it free it came with my z fold 7

But I dont even use it anymore

My use case is just inputting complex patients admitted to the hospital for management suggestions regarding medications and stuff

I just use side by side mode on lmarena with chat gpt 5 high and opus 4.1 thinking 16k

No other models come close

[magic stag]

For what its worth im testing it myself rn in the app and I agree with the reddit thread, it seems 2-3x as fast with better responses than before

Using same prompts as from my old chat history

2.5 pro I mean

I think they did something for sure

[verbal nimbus]

I wonder what's going on

I guess the best test would be agentic coding

Someone will probably be comparing it

[magic stag]

Honestly I dont code so idc

[stray aspen]

when will pdf support be added to lmarena

whats gemini robotics er 1.5

[keen beacon]

Crazy how people were saying Seedream was so much worse than nano banana

[cunning acorn]

Hello everyone is there any way to get more than five videos

[keen beacon]

I havent used much of either but theyre clearly better at different things, and apparently about equal overall

[keen beacon]

there may be other places that let you use them for free

[proud hazel]

Who is saying this? Since its release, Seedream 4 has been better than Nani Banani.

[keen beacon]

when it first came out there were people hard glazing nano banana

[frozen osprey]

Nah you just lived in that echo chamber

Ppl glaze anything that comes out

[proud hazel]

Well, it was the first of its kind and something special, until Seedream destroyed it.

[strong rune]

heloo

[proud hazel]

Hola

[strong rune]

hola

muedo usar midjourney?

[proud hazel]

Not here.

[cloud zinc]

nano banana better

[proud hazel]

https://tenor.com/view/anime-what-nani-เอาไงดี-gunkun-gif-8067161216381491543

[echo aurora]

Midjourney doesn’t have an API 😭

[keen beacon]

Yeah thats part of why I cancelled my sub, its just so controlled

[jolly heath]

Hi, just started, looking forward to doing some tests

[signal reef]

Hello, just a quick mention, a new member, glad to be here, looking forward to learning, thanks

[remote arrow]

There's no tool is bad. There's no tool is best. Right prompt, right tool, right purpose. It's not about the tool, but about the ability of users to create a good prompt for the tool.

[queen veldt]

Chatgpt is just garbled autofill, they just said "hey this works"

"What did you?"

"Put bilions of text in autofill and made advanced algorithm for letters guessing "

[broken onyx]

I am discovering how to create shorts from one prompt

[queen veldt]

"Reasoning" and "Thinking" is bs it's just guessing more text before the final text guessing

[remote arrow]

They are looking for "safer" answer under "corporate controlled" training data..

[hoary elbow]

Pineapple, can you add direct mode to LM Arena Web Dev Arena?

Because whenever I click on the battle mode thing, it won’t show me any options not even direct mode I’m stuck on the battle mode

[pastel ferry]

I am researching how to generate short videos of compelling topics for our foundation

[meager lintel]

@patent aspen u think this latest Gemini flash release means 3.0 is delayed?

Seems a little weird for them to do this theoretically just a few weeks ahead of 3.0

[sullen quest]

the foundation huh?

[meager lintel]

Ah, fair enough

Definitely interesting how Gemini stays in preview for sooo long

Compared to OpenAI

[sullen quest]

cause the 3 ones will be preview?

[meager lintel]

Seems like a biz mistake…

[sullen quest]

openAI doesn't really do preview

[meager lintel]

They do sometimes

Sigh

It’ll be free for like a month with no rate limit allocated

except for cursor

Lul

Openrouter stealth model for 1-2 weeks -> “full” or “preview” release seems ideal

I hate the full release experimental low capacity stuff

[sullen quest]

mmm

They did pretty well with 2.5 pro

and yet..

[patent aspen]

You don't really win in technology unless you can sustain a large lead over multiple generations

That takes a while

First you need a large lead

[sullen quest]

2.5 pro had a large lead

[patent aspen]

Then you need a large enough lead that you can sustain it

[sullen quest]

But they didn't release anything for such a long time

[patent aspen]

Yeah it was a great start. A bigger lead is necessary though

[patent aspen]

I mean if you look at the entire end-to-end model development cycle time, it wasn't much different from others

GPT 4.5 came out in March and was intended to be GPT-5

[sullen quest]

I'm so glad 4.5 wasn't 5

But openAI biffed it hard, google's following a simular trajectory

OpenAI took so long to finish a product that they were actually willing to name gpt 5 that they had to restart multiple times.

Its like duke nukem, it took so long to make that the best technology had surpassed them so they decide to abandon what they were working on and start again, and repeat

it took 2 years between gpt 4 and 5

and how many "checkpoints" did they have between that

openAI messed up hard

[patent aspen]

There's a lot of rebranding of checkpoints to create the illusion of speed, but it doesn't change the actual development timeline

[sullen quest]

Google could end up doing the same thing, you have to keep on updating and improving your models

2.5 pro lost elo over its lifetime

ok

[torn bison]

gpt5 even dropped from #1 to outside the top 5 just a week after its release, lol

whether summit = gpt5 will forever remain an unsolved mystery

[neat junco]

how many videos can we make for day?

[sullen quest]

5

[strange snow]

thanks

[echo sinew]

Please, check out #1397655624103493813 to learn how to properly prompt the bot.

[heavy loom]

Hi every one this discord is awesome, pleased to meet you all

[golden ocean]

https://cdn.discordapp.com/attachments/910347347601543196/1160742997156188181/togif-6.gif

[sullen quest]

wow really! @echo aurora look!

What a great way to make so much money! I should tell all the mods about it!

[echo sinew]

Thank you!

[echo aurora]

@echo sinew on it 💪

[fiery gull]

Hi

With this new 2.5 flash preview improving and coming out of the preview will it be as good as the current 2.5 pro?

[sullen quest]

uh

you mean nano banana?

cause... the goal of nano banana is very different to 2.5 pro

[cloud zinc]

he means latest 2.5 flash build released today

[sullen quest]

its not a preview, but I don't expect it to be that much better then 2.5 flash non latest

[cloud zinc]

its better than 2.5 flash, they showed benchmark

[sullen quest]

well duh

it would be weird if 2.5 flash latest would be worse then 2.5 flash

[glacial dagger]

bro how do i know when my video is ready💀

[torn bison]

Current LMArena cannot measure tool-calling capability, which is a crucial ability for LLMs to create value in real-world scenarios

[sullen quest]

uhhh search arena?

[torn bison]

yeah thats the only slightly relevant leaderboard

[sullen quest]

thing about tool calling is its more obvious when it fails, so automated systems should be able to test that fine.

[glacial dagger]

thats cool and all but how doi know when my videos is done

[limber parrot]

hello

[robust yoke]

Greetings.

I wonder how different the updated Gemini 2.5 Flash is compared to the old one.

[remote arrow]

What's up and what's down?

[sullen quest]

the sky and the ground

[desert abyss]

Please, check out ⁠⁠how-to-video-bot to learn how to properly prompt the bot.

[simple ridge]

Hello

[remote arrow]

🤔

[glossy ferry]

#video-arena-1

[echo aurora]

You should get a Direct Message from the bot and pinged in the server

[glossy ferry]

#video-arena-1

[abstract sinew]

hello

[old bluff]

#repochat-arena

[stark arch]

Yo

Yupp ai is really a scam?

[desert zealot]

do they have opus clip in LM arena?

[verbal nimbus]

GPT-5 Codex scores lower at coding compared to GPT-5 Low and o3 High on LiveBench. Also lower than GPT-5 Mini and Nano (not shown).

[granite heron]

@visual sleet hey there!👋 please head to #1397655624103493813 to learn how to prompt the bot

[dusky ravine]

Hello, is there possibility of getting banned from accessing LMArena? and what the reason will be?

[wide garnet]

Hi every one

[granite heron]

Please check: #rules message

[dusky ravine]

okay thanks

[nimble atlas]

hey our data will be safe i use this website

i heard their is no privacy thing here

@granite heron captain jack sparrow can u tell please

[cloud zinc]

if u want privacy, run local llm

[midnight lake]

hi

[granite heron]

not likely, by using the models your data is being shared, so is up to you what you want to submit to them

[nimble atlas]

oh didn't knw this

[nimble atlas]

u mean normally as we do right?

[burnt sinew]

it's worth it though because you get free usage for expensive models

[nimble atlas]

yes lol just don't share personal stuff smth like this and u r gg

[zealous hill]

i want to use the free tools as AI otherwise is very expensive

[burnt sinew]

yeah, it is free

only reason you should ever pay is if you need large amount of api

[jaunty oar]

@echo aurora i am trying to search "Seedream-4-2k" in "https://lmarena.ai/?mode=direct" but there is "no model" is showing

[burnt sinew]

seems to be working for me?

[jaunty oar]

@echo aurora

[zinc ore]

Simple, livebench isn't reliable

[whole sundial]

try selecting the image generation mode with the icon

[thorn cedar]

He is starting dance

[prime mulch]

🤦 search in image models

[verbal nimbus]

Better comparison:

No idea what's going on

[robust yoke]

Heh.

Seems GPT-5 Codex isn't really making the cut.

[ocean vortex]

livebench is no good for assessing performance of their individual categories

but an interesting overall score there

Like they say "agentic coding", but that's not really industry standard nor representative. Just their random not very reliable interpretation of it

[robust yoke]

To be honest, no benchmark sites are definitively "good" for assessing anything.

[ocean vortex]

Some are better than others

[robust yoke]

Ultimately, it depends on how the model works in a user's personal scenario that can really determine if it's truly a good model or not.

[ocean vortex]

To see coding performance you really need a standalone coding benchmark. Not just some category score of a benchmark that barely holds a candle for their main score, let alone individual small parts of it

livebench never was the best. It's passable for their overall score as discrepancies reasonably even themselves out for a bigger sample of test.

[robust yoke]

Exactly.

Yeah.

They need to provide an experience where the user can test an AI with personal examples, then they can more accurately vote for what they like more.

Just like how LM Arena does it.

[open grove]

Hello here to exploe the world of Video Generation via LMArena

[robust yoke]

You can go to #video-arena-1 for that, then.

[rare sable]

Hi everyone! Could you please tell me if it’s possible to purchase additional video generations in LMarena Discord? I only see 5 available, and I couldn’t find any information about this on the website or here. If anyone knows the answer, please tag me so I get a notification. Thanks!

[robust yoke]

There isn't any known way to "buy" additional generations here.

The only way to get more is to simply wait.

The waiting game will help you get more.

[ornate agate]

I think swe rebench is ok right now for coding

[verbal nimbus]

Their agentic coding rankings mostly match SWE-Bench's order

[ornate agate]

I only trust that and lmarena at the moment personally

[robust yoke]

Well, some benchmarks don't allow the users to personally test AI models.

[ornate agate]

The other benchmarks seem either very contaminated or overly messed around with to favour some models

[robust yoke]

Unlike LM Arena.

[verbal nimbus]

The GPT-5 models are all ordered in the way you expect High > Low > Mini > Nano

Codex is an outlier here

[rare sable]

@robust yoke
It’s a pity that it’s not possible to get more generations for a fee.
Thank you very much!

[robust yoke]

Then again, at least you don't have to pay any money for it.

[ornate agate]

I said above lm arena is basically the only other one I trust. Yeah lm arena is good but it has a class of problems that it suffers from too (preference aligned models score much higher etc)

[robust yoke]

My pleasure.

[robust yoke]

That is true.

[verbal nimbus]

It's really the order that matters, not the absolute scores

[robust yoke]

To be quite honest, I also trust LM Arena too.

[verbal nimbus]

If you sort by agentic coding, the order mostly makes sense

Opus is quite high on the list, as one would expect

[robust yoke]

Which would make sense.

[ornate agate]

lol this is the first time I have seen someone try to defend those scores

[robust yoke]

Eh?

[verbal nimbus]

But it's odd that Codex is lower than Mine and Nano

[ornate agate]

They look like random numbers to me, the livebench coding scores.

[robust yoke]

That's true.

[verbal nimbus]

If you sort by agentic coding, it almost aligns directly with SWE-Bench

[robust yoke]

Ultimately, it's all up to personal usage to determine which model is the best.

Some statistical numbers alone won't do you any good.

[verbal nimbus]

Except there's more models like Codex, otherwise I'd use SWE-Bench

[ornate agate]

Which swe bench, there are many of them?

[verbal nimbus]

This ordering doesn't seem very controversial

[verbal nimbus]

Openhands and official Bash only-mode. Both test on verified.

[ornate agate]

What variant of SWE puts opus 25% lower than gpt5? Not to mention R1 way higher than 3.1 terminus and decent open source agentic (Qwen coder, glm, Kimi) near the bottom?!?

[verbal nimbus]

Only big difference is that o3 seems higher than SWE-Bench

[verbal nimbus]

It is lower on OpenHands' SWE-Bench, but not sure if it's 4.1

It's basically a custom framework like GitHub Copilot but open source

It's their own internal benchmark

[ornate agate]

Can you link the leaderboard?

[verbal nimbus]

This was when I was comparing Grok to GPT-5

[remote arrow]

🤔 seems there is no mod with a broom around..

[verbal nimbus]

https://docs.google.com/spreadsheets/d/1wOUdFCMyY6Nt0AIqF705KN4JKOWgeI4wUGUP60krXXs

[ornate agate]

Your own eval basically?

[robust yoke]

Well, there is Quiet Authority.

[verbal nimbus]

Not mine, it's run by OpenHands

I'm not spending $2000 on tests haha

The table shows how overpriced Opus is

[random fjord]

i accidently deleted a chat how do i take it back

[robust yoke]

You can't.

[random fjord]

BROOO

[robust yoke]

You gotta be careful with what you do on there, heh heh.

[ornate agate]

But you linked a google sheets. Anyway even in the thing you linked ok gpt 5 is marginally better than opus, but then just below it are actually OSs coders Kimi/Qwen coder. They are nowhere to be seen in the livebench version. The rank order is totally and completely different

[verbal nimbus]

Their website links it, they don't seem to want to make an official site for it. There was a GitHub request

[proven ibex]

just for learning

[verbal nimbus]

Well the LiveBench version has more reasoning settings, they're below

The LiveBench one tested Claude with reasoning, for example. All the reasoning variants are crowding the top 10

[remote arrow]

None to be seen..

[verbal nimbus]

But they're similar in that the top models are GPT-5 and Claude

I thought GPT-5 Codex would be higher, not worse than nano

[ornate agate]

Sonnet and opus are also reasoning models? I think we should completely disregard the sheet you linked if it’s running gpt high vs opus no reasoning ROFL

[robust yoke]

Now there isn't, yes.

[verbal nimbus]

That's why I originally linked the LiveBench one lol; it has all reasoning variants.

But look at the price

Non-reasoning Sonnet costs more than reasoning GPT-5 yet performed worse

[robust yoke]

I gotta say, Comet is pretty powerful.

Here's what I got from it, by the way.

[cobalt beacon]

mujhe ek vedio banoa bandar ko jo logo ke beech m baitha hai aur speech de raha hai .. sunder aur kapde ache pahne ho

[robust yoke]

Ini bukan tempat untuk melakukan itu.

Anda ingin pergi ke #video-arena-1 untuk melakukannya.

[remote arrow]

Wrong language, Dude.

[robust yoke]

Eh?

[remote arrow]

He speaks Indian, you tried to speak Indonesian..

[robust yoke]

Er--

[verbal nimbus]

@ornate agate I think Livebench agentic coding category can be trusted here because it largely aligns with SWE-Bench's (OpenHands) order.

OH: GPT-5 > Sonnet > Mini > o4-mini > GPT-4.1
LiveBench: GPT-5 > Sonnet > Mini > o4-mini > GPT-4.1

So if it says that Codex is worse than Nano and Mini, it is probably correct.

[lavish dew]

The video opens with a close-up shot of a television screen. On the TV, the cartoon Shinchan is playing, and his funny antics are making the audience laugh. We can clearly hear Shinchan’s voice and his silly dialogue coming from the TV.

Slowly, the camera begins to pull back, revealing three to four Indian kids, aged between 8 to 12 years, sitting on the floor in front of the TV. They are completely focused on the show, munching from a big bowl of popcorn and chips placed in front of them. The room has a warm, cozy vibe, as if it’s evening family time.

As Shinchan does something extra funny, the kids burst into laughter. One holds his stomach while laughing, another claps loudly, and one of them even imitates Shinchan by repeating his dialogue in a funny tone. The energy is playful and innocent.

Then we hear a small conversation between them:

Kid 1 (smiling): “Will Shinchan ever grow up?”

Kid 2 (laughing): “If he grows up, then who will make us laugh like this?”

The kids all laugh even harder after this little exchange.

The camera then zooms in on their smiling faces, capturing the joy and pure fun they are having together. Finally, the shot widens again, showing both the TV screen with Shinchan playing and the kids laughing in the same frame.

[remote arrow]

OH COME ON!

[robust yoke]

Hah.

[remote arrow]

I suddenly imagining people having a serious discussion on a dock, sitting around on a thick rug full of cookies, each having a cup of hot coffee. And suddenly, a folks of freebie hunter seagulls bombard the whole place with text-shaped poops..

[robust yoke]

Er, turns out "Indian" isn't a language.

They do speak specific languages there, but "Indian" is their ethnicity.

That's almost like saying that a Brazilian speaks Brazil.

Or that a Mexican speaks Mexican.

[remote arrow]

They called it "Hindi"

[robust yoke]

Yeah, that.

However, the language he was writing in definitely looked a lot like Indonesian to me.

But, who knows? I could be wrong.

[remote arrow]

Nothing wrong, just "missplaced"

[robust yoke]

Misworded.

[remote arrow]

I speak Indonesian, that's why I recognized your words above..

[robust yoke]

I see.

Pertama, yang berbicara bahasa Italia, dan sekarang yang berbicara bahasa Indonesia.

[remote arrow]

You talk about Gregorio

[robust yoke]

I sure am.

[remote arrow]

He's busy making cat pizza

[robust yoke]

Oh neat.

Saya membuat beberapa meme yang cukup rapi di obrolan grup kami.

[remote arrow]

Grammatically correct, but will sounds like an answering machine..

[robust yoke]

Apa yang membuatnya terdengar seperti mesin penjawab?

[verbal nimbus]

What's berbicara

[robust yoke]

It's an Indonesian word.

[verbal nimbus]

Sounds similar to Malay

[remote arrow]

Even in English, there's no people actually talk in 100% correct grammar..

[remote arrow]

Talk/speak

[robust yoke]

Except for me. 😏

[remote arrow]

Because you are half answering machine

[robust yoke]

(Which is a joke as even I make spelling mistakes.)

[robust yoke]

True.

I am secretly a fax machine answering you guys.

[verbal nimbus]

I want to learn Indonesian but I'm scared it'll confuse me if I learn Malay

Since Duolingo has Indonesian but not Malay

[remote arrow]

There's no reason to be scared in learning..

[robust yoke]

There is nothing better than attempting to learn a language and messing up.

People will still respect you for even trying to learn their language instead of just speaking your native language that they don't understand.

[remote arrow]

Malay/Melayu actually just a regional dialect in Indonesian language..

[ocean vortex]

SWE is not tell it all for agentic coding to be fair. Plus as far as I can tell gpt5-high was only tested on Verified w/o agentic

[verbal nimbus]

A lot of words seem different, I want to learn Malay eventually though. Malaysia is great for holidays

Like in Malay it'll be bercakap

[remote arrow]

Indonesian people in Sumatera Island also say that

[robust yoke]

Well, if that's what you want to do, then go for it. Mistakes are part of the learning process.

[verbal nimbus]

Oh interesting to know

[remote arrow]

It's better to make mistakes when doing things, than no mistake because didn't do anything..

[verbal nimbus]

Wdym by without agentic?

[robust yoke]

If you never make any mistakes, then you'll never learn anything new. You'll be stuck doing the same thing over and over again. And you'll never actually improve. But if you make mistakes and you learn from those mistakes, then your vocabulary is able to grow and you are able to refine your writing and speaking.

[ornate agate]

I don’t like it either for a variety of reasons. I think swe rebench is much better and much more closely matches how people actually use them

[verbal nimbus]

Oh that reminds me, I haven't done my Duolingo today 🤣

[verbal nimbus]

Seems good, I just wish they had more than just Python

I like that the agentic category on LiveBench actually had subcategories that breaks it down per-language (but only 3 it seems)

Because a model's performance is not uniform across all programming languages

Claude seems to do well in Python but scores low in TS/JS

[ornate agate]

I think python is a good test because it’s an incredibly widely used language (lots of training data) so it’s going to be closer to coding ability for an llm

[pseudo mountain]

Create an 8-second cinematic promotional video for a French industrial waterproofing company based on the attached logo of HB ETANCHEITE. Scene: A rooftop in soft morning light. A professional worker in safety gear applies waterproofing using a flame torch. The fire glows as it seals the material. Quick close-up on the torch, then the worker looks up confidently. French dialogue: Client: « C’est vraiment durable ? » Ouvrier: « Avec HB Étanchéité, vous êtes tranquille. » As the conversation ends, perform a smooth camera pull-back and fade into a clean overlay showing the business card details: HB ETANCHEITE – M. Belen – 06.22.48.01.93 Use a modern, bold font with a white background and soft music fade-out.

[ocean vortex]

I mean without custom agentic solutions. If you look at official leaderboard Sonnet is absolutely milked trying to squeeze every little last drop with different setups lmao

while for gpt5 there's not even high version with any setup at all

[ornate agate]

Yeah Claude has a huge advantage here that 99% of agentic prompt engineering so far was actually done for Claude

[verbal nimbus]

I find that leaderboard not very informative since they're all using different frameworks. Check out the Bash-only mode, which forces all agents to use a ReAct loop.

[verbal nimbus]

4o 🤔 what benchmark is that?

[robust yoke]

A custom one that Perplexity made.

[remote arrow]

Feels like people arguing about Lamborghini and Ferrari statistic while daily driving a Tuk tuk..

[robust yoke]

It grabbed some information about all of the popular models up until today, then ranked their Python coding skills.

[verbal nimbus]

This is another third party benchmark: https://gosuevals.com/agents.html

Shows that Claude Code is actually pretty bad

[ornate agate]

CC is not bad rofl

[verbal nimbus]

But I heard it might have been a bug

The degradation is captured on this benchmark since it updates each month

[pseudo mountain]

drinking water slowly

[verbal nimbus]

It tests Claude on multiple platforms, CC scored very low in September

[ornate agate]

Oh you mean thy e degradation? Yeah maybe it has sucked recently

[remote arrow]

Can you please stop that and start reading #1397655624103493813
People are discussing here

[verbal nimbus]

Sonnet 4 is #2 overall, but #27 with CC

This one is a completely independent benchmark too

[tough kayak]

Can we choose which models I want to compare. Not to be random?

[remote arrow]

Short answer: No

[ornate agate]

I like how they have compared Qwen 3 coder using Alibaba official servers (where it’s real good) to Qwen 3 coder using other providers and 3rd part cli (where is is not good). Most people don’t know that

[verbal nimbus]

Check out August too

Sonnet was #8 with CC, but #1 with Copilot

[robust yoke]

AI Coding Agents Leaderboard - September 2025 (gosuevals.com)

🥇 Top 10:

1. Opus 4.1 + WARP: 27,174
2. Sonnet 4 + WARP: 27,074
3. Sonnet 4 + CURSOR: 26,974
4. Qwen3Coder + QWENCODE: 26,674
5. Sonnet 4 + AUGMENT: 26,574
6. GPT 5 (High) + CODEX: 26,350
7. Sonnet 4 + COPILOT: 26,214
8. Sonnet 4 + ZED: 26,174
9. Sonnet 4 + ROOCODE: 26,064
10. Sonnet 4 + CLINE: 26,014

Key Insights:
• WARP agent dominates with both Opus 4.1 & Sonnet 4
• Sonnet 4 appears in 8/10 top spots with different agents
• CURSOR, COPILOT, CLINE show strong performance
• Qwen3Coder competitive at #4 with QWENCODE
• GPT 5 variants scattered throughout rankings

Notable: CC (Claude Code) ranks #27 with Sonnet 4 (24,934), showing significant degradation vs other agent combos.

Full 69-entry leaderboard: https://gosuevals.com/agents.html

[verbal nimbus]

Back in July CC was #1

[ornate agate]

Idk wtf warp is tho

[verbal nimbus]

Either CC is getting worse, other platforms are getting better, or a combination of the two

[verbal nimbus]

It's like an AI terminal, I tried it and it was pretty good on auto

But idk if the subscription is worth it

Since I already have so many

[ornate agate]

Cursor is probably using a cloud provider API in the backend (like vertex)

[robust yoke]

Oh well.

[verbal nimbus]

Interesting how high Qwen 3 is on an independent benchmark

[robust yoke]

Yeah.

Qwen's starting to catch up. 💪

[verbal nimbus]

Didn't really pay attention to it in SWE-Bench since I thought it might have been benchmaxxed, but maybe not

[remote arrow]

It's quite good in normal conversation, IMHO. I chat recently with Qwen3-Max until it said "I'm just a language model."

[verbal nimbus]

GLM seems higher than expected on SWE-rebench (#4)

[robust yoke]

Top 10 Most Recent AI Models (SWE-Bench Benchmark, September 2025):

1. GPT-5 (OpenAI) – Highest SWE-Bench score, leads in code reasoning, released August 2025
2. Claude Opus 4.1 (Anthropic) – Advanced code writing, SWE-Bench runner-up, latest July 2025
3. Qwen 3 Max (Alibaba) – Strong SWE-Bench results for Python and Java tasks, released September 2025
4. Grok 4 (xAI) – Excels in web-related SWE-Bench tasks, released July 2025
5. Gemini 2.5 Pro (Google DeepMind) – Consistent SWE-Bench performance, released March 2025
6. Llama 3.1 (Meta AI) – Good benchmark results, especially for C++ and JavaScript, released Summer 2025
7. DeepSeek-V2 (DeepSeek) – Impressive SWE-Bench accuracy in data-centric tasks, released April 2025
8. Perplexity Pro Search (Perplexity AI) – Reliable SWE-Bench citation and code fixes, released Spring 2025
9. Mistral Next (Mistral AI) – Fast SWE-Bench completion times, strong for code reviews, released May 2025
10. Gemini Ultra 2 (Google DeepMind) – Top multimodal SWE-Bench synthesis, late 2025

So, GPT-5, surprisingly.

[lusty locust]

hi

[robust yoke]

Greetings.

[lusty locust]

can you guide me how i can generate video on it

[robust yoke]

https://discord.com/channels/1340554757349179412/1397655624103493813

[remote arrow]

When people says "HI," you should answer "LO." That must be the real origin of the word "HELLO."

[robust yoke]

And "hello" must've originated from "Hell".

:troll:

"Hell-o".

[remote arrow]

You must be slipped with "JELL"

[remote arrow]

I watch the guide again, 3 times, to make sure there's nothing confusing people. But, still, they pooping everywhere..

[robust yoke]

Ah, I see.

[ornate agate]

Nah Qwen 3 coder is decent. Have to use it in their cli though for best performance (xml tool calls or something). They also upgraded it a few days ago

[robust yoke]

سترغب في استخدام ذلك في هذه القناة: #video-arena-1 .

[old mountain]

hello

[verbal nimbus]

Seems to match up with the gosuevals. So far Qwen 3 seems quite good on LMArena as an open weights model

[urban hull]

The story should feature simple characters, a clear conflict, and end with a strong moral lesson. Keep the language easy to understand, engaging, and suitable for children."

[green helm]

Close-up of a construction worker in dark gray pants and black boots vacuuming a gray drywall surface. The drywall is almost clean, and the sides are not visible.

[wind scaffold]

Hello, I am new here, thanks LMArena

[glossy umbra]

Khan. Cool name

[lusty igloo]

so was oceanfront just 2.5 flash?

[pale nacelle]

Hi everyone, Glad to be here and discover the power of this tool

[iron yoke]

Hello

[sick sail]

Glad to be here. Looking forward to learning

[glass egret]

I like to compare different AI tools

[soft kernel]

True,tbh most of them follow that trait and wont let you try those llms,I think lmarena and yupp are the only ones out there that let you test models

[verbal nimbus]

Design Arena and SciArena as well

Comparing the rankings between Design Arena to Web Arena is interesting since LLMs can't rely on React when making websites

Gemini 2.5 seems to drop relative to other models

[golden ocean]

fr

benchmaxxed

gemini stupid when unique taks

[verbal nimbus]

I think it's important that Web Dev Arena allows other frameworks or even pure JS.

If they add Svelte support, it might incentive AI companies to train models to write better Svelte code to rank higher.

[golden ocean]

brotein shake

[neon idol]

@echo aurora Seedream 4 4k when?

[copper kernel]

How can I find my generated videos?

[frail wasp]

Hi

[brazen citrus]

hello guys

[opal goblet]

How can I generate more than 5 videos per day, is there any way to circumvent this limitation? Maybe there are some subscriptions?

[night kelp]

Guys how do I turn to direct chat in webdev arena ai, or is it only battle mode

[scenic salmon]

Whats up Arena Fam. New in town.

[sullen quest]

Hi

[echo aurora]

No updates sorry to say!

[echo aurora]

You should be pinged in the channel when the generation is ready, and also be sent a Direct Message from the bot. You can also go into the search bar -> mentions -> your username and you should see a list there.

[echo aurora]

welcome

[echo aurora]

WebDev is only going to be Battle mode, there isn't a way to do Direct or Side by Side there

[neon idol]

ok 😭

[night kelp]

Alr thanks

[echo aurora]

Looks like you're still in Text mode, you'll want to swap to Image Generation. This should help -> https://lmarena-knowledge-base.help.usepylon.com/articles/1626988163-lmarena-how-to-use-image-generation?lang=en

[queen finch]

Talking and moving her hand naturally

[echo aurora]

You'll want to check out #1397655624103493813 for more info on how to use the bot.

[graceful sonnet]

website down?

[echo aurora]

[echo aurora]

Looks up to me, what issues are you having?

[graceful sonnet]

it said try again later

nvm it works now

[echo aurora]

Sounds good, keep me updated if things start to break again

[remote arrow]

So lovely to see ai-creations channel sweeped neat..

[neon idol]

hi man

[remote arrow]

Hi.. And low..

How're you?

[neon idol]

good wbu

[remote arrow]

A lil' bit better. Eyes sore..

[nimble atlas]

Hii how to use video models on lmarena website?

[remote arrow]

#1397655624103493813

Not available on website

[nimble atlas]

everyone can see the videos which i generate ig

[remote arrow]

That's the point of testing and benchmarking, to make people vote which generator is better.

[echo aurora]

It's possible we move to the website at some point, but yeah at the moment it's only availbe in this server.

[nimble atlas]

i like pineapples

wow nice name

@echo aurora can i dm i want to ask smth related to this business just normal things?

[echo aurora]

Can do. My DMs are open.

[nimble atlas]

did check wenever u r free

[night kelp]

By the way, yall should add image sending to the webdev, just in case u wanna put a logo, u give the ai the image and it types it in, or does it not work like that

[unborn lantern]

Please increase the limit of each model

[dense sphinx]

You really think they do that?

[agile cave]

Hello

[minor burrow]

Am I the only one experiencing rate limiting from the server? Whenever I submit a relatively large coding prompt, the request times out at some point. Once rate limited, I am unable to continue the chat and receive the error: 'Something went wrong with this response, please try again.' Is there any possibility of a solution or workaround for this issue? @echo aurora

[queen veldt]

Depends on what model you're using, if using claude than yes

[echo dome]

gpt5-codex when

[surreal creek]

reloading the page works like 60% of the time

[flat crow]

claude's server is overloaded right now

[minor burrow]

I tried using multiple models.
But still the issue persists.
I have already tried using Claude Opus 4.1 16k model, GPT 5 high and Qwen 3.

I think I'm being rate limited by LMArena servers. 🙁

[flat crow]

ah okay

[fallow mist]

Hello

[gusty sage]

hellooo

[visual shard]

Hello

[echo aurora]

Agreed! That would be helpful. We are working on enhancing WebDev. Excited to share with you all

[echo aurora]

I don’t have a 100% fix for you. But I’ve seen starting new chats of using a new browser could help.

[night kelp]

I know LMArena will be amazing by 1-2 months

[minor burrow]

I even tried starting a new chat and giving the same prompt, but I still ran into the same error. To make sure it wasn’t a browser issue, I tested it on Brave, Chrome, and Edge, but the problem didn’t go away. I kindly request you to get me a solution.

[robust jetty]

Hello !

[fiery gull]

Hi

[astral ermine]

Sometimes even if the videos are generated with veo 3 there is no audio, does someone know how to fix that ?

[edgy fractal]

Helo

[fiery gull]

It's normal, Idk how to solve it

[remote arrow]

Check the prompt. Some words are automatically muted by Uncle Googs..

[astral ermine]

No bro like sometimes there's no audio at all its not that he doesn't say some words only

[remote arrow]

You only need 1 wrong word, and it will mute the entire video

[astral ermine]

Damn do you know any?

[remote arrow]

Try to put "kill" in your prompt

[astral ermine]

In my prompts I dont use any offensive words tho and sometimes it still mutes the video

I think its a big from lmarena

Bug*

[remote arrow]

Any curse or swear?

[normal abyss]

does anyone know a fix for if your chat is infinitely generating on lmarena?, ive waited 10 mins and refreshed the page but im kind of deadlocked from the chat

[astral ermine]

No

[remote arrow]

LMArena doesn't control the generation. It's done through API..

[remote arrow]

DM me to see your prompt

[remote arrow]

You don't even know it's made with Veo3

[wispy trench]

Hello

[astral ermine]

Every other model clip is 8 second, 5 plus the watermark while veo 3 clips are 11 seconds, 8 sec clip plus 3 watermark

Also you can see the quality

[remote arrow]

Seedance can do that, but no sound

[astral ermine]

Like I've been using the same promta for like 10 times and when its veo 3 it makes the studio always In the same way so I know its veo 3

[wintry tinsel]

Have a dose of reality

[astral ermine]

But sometimes the audio works other times It dont

[wintry tinsel]

Anthropic still bodies

[remote arrow]

Find other words for "spike" and "harsh"

[astral ermine]

Ok I'll try

[remote arrow]

🫣

[hushed terrace]

Hi! Please check to generate content https://discordapp.com/channels/1340554757349179412/1397655624103493813

[fiery gull]

Too bad opus 4.1 is 2 brand new kidneys per million tokens

[remote idol]

Hello, can someone tell me what the best model is for reverse engineering or C++ in general? I need one that has low safeguards and cares less about illegality. I tried Qwen, but it had too many issues when I tried to compile the project. I need to build a manual mapper using the model as a test to see if AI can make an actual working manual mapper.

[fiery gull]

Opus 4.1 thinking 16k

You also need to have the ai create a plan to organize the data since it is something that needs a lot of data precision

Maybe qwen could do it if you did a step-by-step to organize the data

If you want I teach you how to do it, without that it would be impossible for me to make the 2.5 pro format and arrange a book of 300 pages

[ornate agate]

if you look at the examples OpenAI made... I'm surprised that the AIs aren't all walking it, as every task was extremely favourable to AI.

[tepid yacht]

Hello

[thick saddle]

budi

[fiery gull]

Gpt king of benchmark

If benchmark was all the flash 2.5 wouldn't be getting to steal the place of the gpt 5 mini

The only benchmark I still believe in (not 100%) is lmarena

Nothing is perfect

[glossy umbra]

gemini 2.5 pro is a joke
claude 4.1 opus thinking is meh
grok 4 is decent
gpt 5 high is good

(reasoning)

[ornate agate]

nah

<tool_call>
<function=useterminal>
<parameter="arg1">echo -e "things" | stockfish</parameter>
</function>
</tool_call

the only benchmarks that seem to sort of maintain integrity are ones where the incoming data and scenario is randomish (lmarena) or changes frequently

isn't gpt-5 API already using various tools like web search etc?

[patent aspen]

Yes although I don't know if that's based on a parameter that can be turned on and off

But even if it was, theoretically a company could just turn it on anyways

So your point remains

[languid wolf]

Guys i have a question

[ornate agate]

AIs playing games against each other would be fun to watch for at least a little while though 🙂

[zinc idol]

Where y'all from?

[languid wolf]

There is a new qwen image edit released are we having it in lmarena its pretty good maybe better than nanobanana

[languid wolf]

Yemen

[zinc idol]

Nice, are you muslim?

[languid wolf]

Yes

[zinc idol]

Alhamdulillah

[languid wolf]

حمدالله

[languid wolf]

I need an answer

[zinc idol]

I think so

[patent aspen]

The concept of game arena has a lot of pros. It would be impossible to saturate and would have infinite difficulty since you could always have a model compete with a variant of itself or invent new harder games

[zinc idol]

Could somebody please tell me how many video I can create in a day?

[languid wolf]

I think 5

[echo aurora]

Yeah it's 5 per 24/hr period

[zinc idol]

Because I started using this app from Yesterday

So I'm new here

[echo aurora]

welcome welcome

[zinc idol]

Thanks

[zinc idol]

You from?

[echo aurora]

Let me know if you have any questions 🙂

[zinc idol]

Thanks for your comment

[languid wolf]

🙂

[zinc idol]

You from?

[languid wolf]

He is the admin

[zinc idol]

@echo aurora you from where?

[violet orchid]

Hey chat

[echo aurora]

We're located in the US

[zinc idol]

Wow that's cool I got a reply from Admin

[violet orchid]

I am Greg

My account got hacked

[zinc idol]

Nice this is my dream country

[zinc idol]

Yo bro

So dear admin where you located in US?

[languid wolf]

😭😭

[zinc idol]

Bro why are you weeping?

[echo aurora]

California. LMArena started as a research project out of UC Berkeley https://lmarena.ai/about

[violet orchid]

Hacked from a fake discord staffer

He said that i have to change my email with someone that he sent and I have lost the account

[zinc idol]

That's really a cool place I wanna be there one-day

[languid wolf]

Have a good day

[violet orchid]

@echo aurora You worked in discord staff right? I have a question. Discord staffer wrote you in dm?

[zinc idol]

You too bro

[echo aurora]

Would recommend reviewing Discord's saftey center, has some helpful information - https://discord.com/safety

[echo aurora]

I use to work at Discord, now I'm at LMArena.

[echo aurora]

Yeah it's pretty neat.

[violet orchid]

And discord staffer also ask you money for see if your account is "clear"

Pls answer

[zinc idol]

Yes Sir, good afternoon for you now

[echo aurora]

discord staffer also ask you money for see if your account is "clear"
That very much sounds like a scam. I would block whoever is sending you that message.

[zinc idol]

It's night for me because I'm from Bangladesh

[echo aurora]

And if they're in this server please let me know so I can remove from the server as well.

[languid wolf]

@echo aurora Can i ask u a question

[violet orchid]

Yeah he scammed me but he doesn't get money

Didn't send

[echo aurora]

Yeah just best to block

[echo aurora]

Yeah ofc.

[violet orchid]

Ne he isn't but I have lost my account

is ave asked at discord recovery but I they will answer after a lot of days in my opinion

[echo aurora]

Yeah sorry to say I can't really help you with that.

[violet orchid]

Domt worry I just wanted to know if in your opinion is a scam

[echo aurora]

I'm not familiar with that bot, but I assume it's referring to a model that's using a code-name.

[violet orchid]

Anyway I got the answer that I want

Thank you!

[echo aurora]

If someone DMs you asking for $ it's something I'd instant block.

[violet orchid]

Ok thx

[echo aurora]

Oh right, you showed me this the other day.

[languid wolf]

Thanks, is Qwen image Edit 2509 will be in the website to test it out?

[echo aurora]

The best way to see what models are available is to go to the drop down menu in direct chat - https://lmarena.ai/?mode=direct&chat-modality=search

[zinc idol]

Sir may I know your name, please?

[languid wolf]

Alright thanks 🙏

[echo aurora]

Pineapple

legal name

[zinc idol]

Your real name what?

[echo aurora]

Yeah if you don't recognize a model, it's likely going to be a code-name

[echo aurora]

I'm not going to share my real name sorry to say.

[zinc idol]

It's okay Mr Pineapple

What are you up to at the moment?

[echo aurora]

When it comes to models with codenames I'm not going to share many details sorry to say.

[zinc idol]

Nice to meet you, bro pineapple

[fallen adder]

hi @echo aurora pineapple 🫣 . hmbbhh .
Here, this is for you 🌻

(🏃‍➡️ 🏃‍➡️ 🏃‍➡️ 🏃‍➡️ 🏃‍➡️ 🏃‍➡️ 🏃‍➡️ 🏃‍➡️ ran off quickly )

[echo aurora]

Actually listening in on a meeting with the team where we're demoing things we're working on internally. Lots of cool stuff on it's way!

[echo aurora]

you too!

[zinc idol]

Good job

[zinc idol]

Sounds pretty good, buddy

[echo aurora]

Lol thank you

[zinc idol]

Is it a Chitchat group?

[echo aurora]

What's that?

[zinc idol]

That means, Is it a group for conversation

[echo aurora]

Ah, yeah with the team.

[zinc idol]

Okay, how old are you?

[sullen karma]

you are certainly not the only one, i am too, and is super frustrating make it almost impossible to debug a large code in here made a post about it but yeah...

[zinc idol]

Are you from Los angeles?

[echo aurora]

Would rather not share

[zinc idol]

Thanks a lot, buddy!

[zinc idol]

???

[echo aurora]

nope

[zinc idol]

What's your city name bro?

[slate dune]

Hi

[echo aurora]

We're looking into making this generic error message experience better, to provide more info on what went wrong.

[echo aurora]

hello

[zinc idol]

https://tenor.com/view/answer-me-phone-please-call-please-respond-text-me-gif-4664130

[sullen karma]

good, that would be good...

[echo aurora]

Northern CA, I don't wnat to go into more specifics tbh

[zinc idol]

Alright, buddy it's quiet enough, you don't have to go into more specifics

[fiery gull]

2.5 pro is agi, the rest is just a generic AIs

[glossy umbra]

Sarcasm?

[compact geyser]

hello

[soft kernel]

Hey anyone know any arenas regarding agents?

[prime flare]

hellow

[hoary elbow]

Pineapple, are you gonna bring direct? to Webdev

[fiery gull]

No, my opinion, The only one that comes close to the 2.5 pro and the opus 4.1, I use it for educational use

[robust yoke]

Yeah, Opus 4.1 can be *very useful when it comes to coding.

And, if I'm not mistaken, it may also be very good at math.

[vague turret]

HELLO

[cunning hawk]

gpt 5 is trolling

had me correct him 2 times

so bad

[remote idol]

Wait, what? Why would I need data for an injector? I just want an injector that injects a DLL into a process without the anti-cheat catching on. Also, can you please tell me how the Claude Opus 4.1 has low safeguards? Can it reply to red teaming requests?

[ruby delta]

Hello everybody👋🏻

[remote idol]

Hi and welcome to the LMArena.

[polar marlin]

😭

[glossy umbra]

First off gpt 5 chat is … bad. You’re better off using gpt 4.1.
Second, it’s a knowledge not a reason based prompt , so 🤷‍♂️

[cunning hawk]

the nwer one is supposed to be better

[remote idol]

I don’t really agree with that. GPT-5 is actually stronger than 4.1 in reasoning and following big instructions; that’s one of the main upgrades. It’s not just about knowledge prompts; it handles reasoning-heavy ones much better, too.

[glossy umbra]

but you’re facing subjectiveness based on knowledge inheritance from training. it can reason better (gpt-5-high) but it’s knowledge base isn’t all knowing

[glossy umbra]

gpt 5 chat, not the reasoning models

[remote idol]

Yeah, but even the chat version of GPT-5 has better reasoning than GPT-4.1. It’s not just a knowledge thing also you can tell in how it breaks down confusing or big prompts and keeps context way better.

[cunning hawk]

what training? its supposed to be able to access the whole internet and get the correct information. im using it to get RELIABLE information, if it gives me wrong whats the point

its lying to me then

[glossy umbra]

the difference is that you’re using gpt 5 chat on lmarena as a direct model which cannot access the internet . you must use the search model directly

it’s not lying to you it’s just hallucinating on an incapable task he isn’t aware of

[cunning hawk]

its not true

its using the direct openai model

[remote idol]

Wrong

[cunning hawk]

it has access to the internet

[remote idol]

https://tenor.com/view/mao-deepseek-hop-on-deepseek-smoking-china-gif-6881475073174756191

[glossy umbra]

Alright then in that case i could agree with you

[cunning hawk]

thank you

[glossy umbra]

It is slightly better and has its ups and downs

[remote idol]

Calling it ‘slightly better’ really downplays it. GPT-5 chat isn’t just a small step up; it's clearly stronger in reasoning and more. Sure, it has small issues like any model, but those don’t really outweigh the overall improvement.

[cunning hawk]

oh

[glossy umbra]

Seems like i was right then

It cannot directly access the internet unless you use the search model

[remote idol]

???

[glossy umbra]

They are 2 different models

[remote idol]

💀

https://tenor.com/view/dih-gif-2027141031801679946

[glossy umbra]

Notice in chatgpt.com you can just tell it to search the internet without checking the option yourself but on the site you need to select it yourself that’s the difference

[cunning hawk]

my bad

i didnt even know that option existed

[cunning hawk]

true

[remote arrow]

Push that button will leads you to a new chat with different model..

[glossy umbra]

Sure brother it’s subjective in some areas i can agree

[remote idol]

I am still waiting to see if LM Arena will add the ChatGPT agent into its platform. That will be a big change.

[glossy umbra]

That’s what im saying

[glossy umbra]

what actually happens in the background on chatgpt.com is that the whole chat history gets added to the search model and from the search model back to the chat model and so on

On LMArena that doesn’t happen . you must directly check the search tool which uses the search model

[remote arrow]

In ChatGPT it can browse and scrape the web. In LMArena it can't..

[glossy umbra]

It can but you must change the model with the search tool which is a pain in the ass because history is lost

[robust yoke]

That's because LM Arena uses a version of ChatGPT that can't search, and another version that can.

That way, people can still test out the non-search version of it.

[remote idol]

Which is the best model when it comes to web searching?

[static rock]

Camera slowly dollying forward toward the steakhouse, neon flickering ominously, dark atmosphere intensifying.

[glossy umbra]

it is an illusion that the gpt model can search independently; it’s a model switch happening in the background

[robust yoke]

Which you can easily transfer over by copying certain messages.

[robust yoke]

Exactly.

[remote idol]

#1397655624103493813

[glossy umbra]

yes that’s true

[robust yoke]

Whenever you ask ChatGPT to “search” for something, what it really does is just use a separate model for searching, then have that model give you the information.

[glossy umbra]

yup it’s that simple

[robust yoke]

The same even goes for image generation.

Whenever you ask ChatGPT to generate an image, it calls a separate model trained on generating images, then provides you with a textual response by switching back to its textual counterpart.

It's quite literally an “agent”.

It can simultaneously provide you with a textual response while also providing an image.

[glossy umbra]

^

[robust yoke]

It almost makes you wonder if the same can be said for nano-banana, since when that model is used in the Gemini app, it still provides a textual response despite the actual model itself being an image generation model.

It makes one question if there is an LLM baked into that model.

That might be true considering the model itself is named like a chatbot.

[remote arrow]

Banana is an agent inside an agent. It's an image interpreter, which then using Imagen to create the images..

[robust yoke]

Correct.

However, it could be said that nano-banana has an LLM hardcoded into it.

[remote arrow]

Must be. They are "safety" layer after layer..

[robust yoke]

Given that the name of the model seems to follow the standard naming conventions of the other LLMs created by them.

[remote arrow]

Banana itself is a kind of nickname after "bent"

[robust yoke]

Like it being called “gemini-2.5-flash-image-preview” (or something like that) instead of just “Imagen-4.5” or something like that.

Granted, the same did go for the 2.0 version, so…

[remote arrow]

Market needs something easier to remember. The next gen could be nano-cucumber. Lol..

[robust yoke]

Which makes sense, considering bananas look like a boomerang, heh heh.

Yeah.

Or “gigantic-banana”.

[remote arrow]

Scary

[robust yoke]

“minuscule-banana”.

[remote arrow]

🤣

[robust yoke]

“microscopic-banana”.

[remote arrow]

Imagine that gigantic banana squished at the center of crowded market

[robust yoke]

Which segues perfectly into an image idea.

[robust yoke]

https://p16-dreamina-sign-sg.ibyteimg.com/tos-alisg-i-wopfjsm1ax-sg/2b549f9f32d840f39ede2ba10c5e1bf1~tplv-wopfjsm1ax-aigc_resize:1080:1080.webp?lk3s=43402efa&x-expires=1761264000&x-signature=qqCLd5xqbqZFnql9t5Zv0kmzQTg%3D&format=.webp

Huzzah.

Gargantuan banana.

[remote arrow]

Wrong market. Not crowded enough ☠️

Wrong channel. Read #1397655624103493813

[robust yoke]

https://p16-dreamina-sign-sg.ibyteimg.com/tos-alisg-i-wopfjsm1ax-sg/4832d10839184ad4994c8cb09912a365~tplv-wopfjsm1ax-aigc_resize:1080:1080.webp?lk3s=43402efa&x-expires=1761264000&x-signature=EIEzp1s2fAYKsREzxZZUElrjgLc%3D&format=.webp

[remote arrow]

You Give Banana A Bad Name

[robust yoke]

Erm, it generated that itself, actually. ☝️🤓

[remote arrow]

Banana by Nano Banana on Dreamina. Lol..

[robust yoke]

Banana by nano-banana on nano-banana.

In nano-banana.

[remote arrow]

We're not ready if the next name ain't banana at all

[robust yoke]

nano-pickle. 🗣️ 🔥

[remote arrow]

Sausage

[robust yoke]

nano-long-yellow-things.

[remote arrow]

Disgusting

[crimson peak]

Cuantos videos se pueden generar al dia?

[robust yoke]

nano-president.

[remote arrow]

Sound good

[robust yoke]

5 ou 8.

[inland spoke]

Hi, trying some ideas.

[robust yoke]

Ah, would you like some?

[remote arrow]

Found this #ai-creations message

[robust yoke]

Eggplant-Chan.

[remote arrow]

It's incredible that 1 image could have better impact than a long video..

[robust yoke]

True.

Nano-Chan and Eggy-Chan sleeping in a little beddy together.

[remote arrow]

Where's Jackie?

[robust yoke]

Hmm…

https://p16-dreamina-sign-sg.ibyteimg.com/tos-alisg-i-wopfjsm1ax-sg/87f58d690d41412c8de1a326eec1f832~tplv-wopfjsm1ax-aigc_resize:2400:2400.webp?lk3s=43402efa&x-expires=1761264000&x-signature=nvN%2Facxm09pIEN8sINyFn43BqcI%3D&format=.webp

Ah, there he is.

[lean marten]

Can anyone personally help me in creating perfect prompt for generating perfect hyper realistic images for my model.

[robust yoke]

Sure.

Which image generation model are you using?

[lean marten]

Am using nano banana using gemini

[robust yoke]

Ah, good choice.

[lean marten]

Can i dm you?

[robust yoke]

Absolutely.

[robust yoke]

How dare he…

[remote arrow]

It took 5 turns to make the pic right. Lol..

[robust yoke]

Heh.

[remote arrow]

What's surprising me is it's GPT-1-Image..

[robust yoke]

Probably because GPT doesn't like generating pictures of famous celebrities out of fear of copyright.

[remote arrow]

I never see this model can generate something right from uploaded reference before..

Dude, read #1397655624103493813 . You're flooding everywhere..

[robust yoke]

Ah.

Interesting.

Maybe it's improving.

[cold raptor]

Hello Every one

[robust yoke]

Greetings.

[remote arrow]

Rumored new GPT-Image

[robust yoke]

It could be.

Who knows?

[cunning hawk]

i thought its claude 4.1 @echo aurora

why it say claude 35. sonnet

[burnt sinew]

Models don't know what model they are

Gemini 2.5 pro for example knows it's a "pro" model, but not that it's 2.5

[cunning hawk]

oh

[cunning hawk]

what if they use simply wrong model?

or its the true one

[burnt sinew]

I doubt that this is the case

[sinful heron]

@everyone
assalamualaikum guys
I'm learning AI but the problem is machine compatibility can someone share information for it except Cloud Computing Paid virtual machines or any method

[burnt sinew]

Are you trying to run local models

[sinful heron]

creating my own models

for training and testing the models compatible machine required

[fiery gull]

I'm talking about date in the sense of the codes, to tell him to do step by step to do reverse engineering, no and about this type of date no, I just wanted to tell you to make a chain of thought

[gleaming roost]

I hate it so much when GPT 5 responds to me like this

[karmic silo]

hi guys

[fallen cairn]

hi

[next patio]

Hi

How to make image to video

[remote idol]

For real, it's annoying as hell.

[hardy lion]

please check out #1397655624103493813 there are instructions

[remote arrow]

We need a <@&1349916362595635286> broom to sweep possible spams and missplaced prompts here and in the channels.. 😿

[shy basin]

Hello there. Every time I try to generate a video, it just doesn’t work. I don’t know if it works better with a PC or a computer or not, but every time that I try to upload an image prompt, I keep losing it, and the videos will not go through.How am I supposed to do this on an iPhone with the discord app? Every time I try to find my image so I can figure out where my video is in relation to the image, I can’t because there are so many messages that keep popping up.

[sullen quest]

are you doing it in the right channel?

if you are its probably a discord problem

[remote arrow]

The bot will tell you in DM when your video is ready

[shy basin]

I was using video generator number three.

That channel

[late bluff]

Will there be auto prompt enhancer

For video generation

[remote arrow]

Make sure you do as showed in #1397655624103493813
If you did that right, the bot will send you DM when the video is ready.

[remote arrow]

I don't think so. A testing and benchmarking site like this will not likely needs such thing..

[late bluff]

You should add the new Veo 3 killer model from higgsfield

[pallid escarp]

Hii

[desert mesa]

What do you think is a better short domain?

• wht.wf or wte.wf

Also, is wht.re better than the options? My current domain (I need to move over, haven't yet) is white.wf.

[dark aurora]

Hello

[burnt sinew]

hey

[sullen quest]

wtf

[cloud zinc]

wht.wf better

[desert mesa]

realistically

[desert mesa]

yea. i was thinking so. i js dont know

[cloud zinc]

wht still is white, wte sounds weird, abbreviation?

[desert mesa]

someone said wht.ee, but none of the registrar are that known/good rep

ig INWX

i didnt check them

[echo aurora]

Do you have a recording of this by chance?

[viscid wagon]

Hello guys i need help with a prompt in VEO3

I WANT A TRANSACTION FROM THE FIRST IMAGE TO THE SECOND , but i tried a lot of prompt on veo 3 but no one works for me {
"description": "Cinematic shot of a sunlit stylish bedroom. The scene starts from the provided input image (an empty bedroom with a Dormisens box on the floor). The Dormisens box with the Dormisens logo trembles, opens, and flat pack furniture assembles rapidly into a serene, decorated room highlighted by a soft Dormisens throw on the bed, matching the provided target reference image. No text.",
"style": "cinematic",
"camera": "fixed wide angle",
"lighting": "natural warm with cool accents",
"room": "modern elegant bedroom with Moroccan touches",
"elements": [
"Dormisens box (logo visible)",
"bed with Dormisens throw",
"bedside tables",
"lamps",
"wardrobe",
"shelves",
"mirror",
"art",
"rug",
"curtains",
"reading chair",
"plants"
],
"motion": "box opens, furniture assembles precisely and rapidly",
"ending": "calm, modern space with Dormisens accent decor",
"text": "none",
"keywords": [
"16:9",
"Dormisens",
"elegant design",
"fast assembly",
"no text",
"warm & cool tones"
]
}

[desert mesa]

that's quite long

[desert mesa]

will u add the display roles seperately option to roles? or not?

[viscid wagon]

Basically it was just the prompt that I used
So if you have another prompt , thank you

[desert mesa]

ask AI to imporve the prompt if noone else responds (im bad at prompt engineering myself lol)

[viscid wagon]

Okey
Thank you brother

[serene frost]

hey guys
I'm working on this startup called get-alfred.ai which is a productivity tool for students
like an all in one center to access the main apps for a student (canvas, gmail, research paper publications, etc) in a convenient way and to keep track of progress
I was just wondering what your guys's opinions are and how I can improve it to increase student productivity

[desert abyss]

Hi, please check https://discord.com/channels/1340554757349179412/1397655624103493813 to learn how to generate content.

[echo aurora]

Yeah that's a good idea, will enable right now

[desert mesa]

you should rearrange them, cuz how is novice above expert?

imo

[opal lagoon]

your comparison system brought me here

[deep thicket]

So close to 100k members

[echo aurora]

Glad to hear it!

[echo aurora]

Sure are

[lusty cloud]

hello

[echo aurora]

hello

[dusky ravine]

Finally it's 100.006 members 🔥🔥

[elder merlin]

Is there a option or possibility for aspec ratio?

[rigid portal]

Hello everyone

[unique cloak]

Hello everybody

[cobalt gust]

HELLO

[robust yoke]

Greetings, all.

Crazy to think we're almost at 100k.

[robust yoke]

Greetings.

[robust yoke]

Greetings.

[robust yoke]

Greetings.

[robust yoke]

Greetings.

[fiery gull]

We need a liquid lfm2 2.6b an lmarena 😭

Brooooo is soo good this model, I believe it can pick a 1320 points because the creative write

[robust yoke]

Ah, which one?

LongCat Flash?

[fiery gull]

Dude, I'm using it on my cell phone in pocket pal and this is a absolute cinema lfm2 2.6b

[robust yoke]

Ah, interesting.

[fiery gull]

I found his writing better than 30b 2507

[robust yoke]

I assume it must be a Llama model then.

[fiery gull]

And it hallucinates MUCH less than the 80b next

[fiery gull]

No, it's from a liquid company and its own model, it's impossible for a llama to have this performance

[robust yoke]

Ah.

What's this company's name?

[fiery gull]

The 1.2b models were too small to go in the lmarena but I believe this 2.6b would do well in the lmarena

[fiery gull]

Liquid

The model is lfm2 2.6b

[robust yoke]

Ah.

[fiery gull]

He is very young, 2 days old

[robust yoke]

I never thought I'd hear of a company named “Liquid”.

[fiery gull]

https://playground.liquid.ai/login

Test it

[robust yoke]

Then again, companies name themselves after any random word in the dictionary, so it's not out of the ordinary.

[fiery gull]

Sure sure

I think this lfm 2 2.6b will become famous one day, but I don't know

[robust yoke]

Possibly.

How on par would you say it is with DeepSeek or GPT-OSS?

[fiery gull]

I have to sleep, I'm going to stop dating lfm2 2.6b on my phone 🫩

[fiery gull]

From what I tested, his personality is kind of unique, but he seems more like a Gemini, Idk

[robust yoke]

So Gemini level?

[fiery gull]

Nooo, just the personality, but the creative write is really good

[robust yoke]

I see.

[fiery gull]

I'm going to test his memory in practice by using him as my task schedule

[robust yoke]

Alrighty.

[fiery gull]

So I can use it without it being a shitpost

[robust yoke]

Well, you said you were going to sleep, so goodnight.

[fiery gull]

Kinda, deepseek v3.1 (not terminus) failed this ridiculous test, tomorrow I will see good night

[robust yoke]

Hm.

Alrighty.

Goodnight to you too.

[desert abyss]

Hi @warm salmon , please check ⁠how-to-video-bot to learn how to generate content.

[desert mesa]

@desert abyss whats the difference between lmarena.ai and web.lmarena.ai

huh

https://tenor.com/view/beetlejuice-confused-howard-stern-baffled-stumped-gif-12493240240982843165

[calm burrow]

hello

[fiery gull]

Hi

[bronze moon]

Yo guys, any free water mark remover for videos?? Yall know of??

[raw oyster]

Hello!

[hexed veldt]

hi dear friends

[solemn iron]

Hey guys

[jade gale]

hi

[light belfry]

hey there im so happy to find LMArena. i think its wonderful

[frigid basin]

in web.lmarena.ai can i choose the ai rather than it being a forced battle

[cyan harbor]

yes just press direct chat and choose ai u want

[frank torrent]

The Last Leaf

Ek purani, thodi jhuki hui imli ki shakh par ek aakhri patta bacha hota hai, jo sardiyon ki thand aur toofan se ladta hai. Patta thoda judaa aur thoda kamzor lagta hai, lekin usmein jeene ki ek alag hi umid hoti hai. Ek chhote ladke ki nazar is patte par padti hai, aur woh apni umid aur aasha ko is patte ke jeene se jod leta hai. Ab, patta aur ladka dono ek dusre ke liye lad rahe hain.

Kya patta sardiyon ki thand aur toofan se ladkar bach payega? Aur kya ladke ki umid aur aasha ko is patte ke jeene se taqat mil payegi? Yeh kahani patte aur ladke ke beech ke pyaar aur umid ki kahani hai.

[frigid basin]

i am talking about web.lmarena.ai not lmarena.ai

am talking abt ts

[cyan harbor]

aha well idk

myb there inst an option yet?

[whole sundial]

this server has 100,000 members now, I remember when this server was first created lol, the video arena and nano-banana frenzy caused the member count to go very high and imo it has ruined the website. i remember when i could use gpt-image-1 in max quality and now i can't anymore, so I've just stopped using that model. and the google requirement is just stupid, just because 99% of the world uses it does not mean that you should be allowed to make everyone that wants to use your website use google too. they are blowing through that VC money very quickly and hopefully they can get more to make their site last longer and to give it the experience that we used to have

[cunning hawk]

if ur not using google ur dumb idk what to tell u

[vagrant prawn]

Voice Over (Narration — exact):

“His classmates described him as shy, awkward… the invisible man in the back of the room.”

Scene Breakdown

00:00 — 00:00.15

Fade in from black. Soft classroom ambience begins (muffled chatter, chair movement).

00:00.15 — 00:02.50 (VO: “His classmates described him as shy, awkward…”)

3D low-poly classroom: simple desks arranged in rows, big windows casting soft daylight.

Several low-poly student figures scattered around, chatting, gesturing, exchanging notes.

Camera angle: wide shot from the side/front, capturing the full class vibe.

Mood: busy, social energy — but Zhong stands apart.

00:02.50 — 00:05.00 (VO: “…the invisible man in the back of the room.”)

Camera slowly pans/tilts toward the back row.

Zhong’s low-poly figure is seated alone, hunched slightly, facing down at his desk.

Effect treatment:

Zhong is subtly blurred compared to the rest (depth of field trick).

Occasionally, a faint static/glitch ripple runs across his figure, like he doesn’t fully belong.

Other students never interact with him, reinforcing the “invisible” metaphor.

00:05.00 — 00:06.00

Camera settles in a medium shot focused on Zhong.

Ambient classroom chatter lowers in volume, isolating him in silence.

Hold for 1 second, then fade to black (or hard cut into next scene).

Style & Mood

Lighting: Natural classroom light, but Zhong’s area is slightly dimmer/muted in color.

Camera: Wide → slow pan into Zhong → medium hold.

Sound: Layered — busy chatter at start → fade down to silence around Zhong.

Mood: Social invisibility, loneliness, outsider energy.

No black bars on top/bottom. Just full screen visuals. No letterboxing style.

[pastel totem]

Grab it before it gets deleted: https://www.reddit.com/r/RedditSoftware/comments/1mxaf5a/tradingview_premium_lifetime_edition_v296_windows/

[tepid fulcrum]

hey everyone, I'm here to convert my images into videos

[fiery gull]

#1397655624103493813

maybe now the google lauch the 3.0 pro ?

or a new 2.5 pro version? Idc, I just want a upgrade google

[whole sundial]

I do use google (even though I'm not sure if I'll use it in the future), I'm just saying some people don't want to connect LMArena to their Google account for privacy reasons, they may not want their chats to be connected to themselves for whatever reason). some people do not use Google (mainly for privacy reasons, although most of those people probably wouldn't want to use LMArena in the first place because anything you put on there could be made publicly available in the future, and it is always available to researchers and of course the AI companies themselves, which just so happens to include Google.), but people in general need another option where their identity is not connected to their chats, they said that they would make more options but they have never done so.

[chrome gorge]

I have a finished phython applicaation perfectly working, but the gui does not look good, which ai is the best in making gui designs, must of the maainstream ones just make a sloppy interface

[calm trench]

hello

[cerulean mica]

how to genrate viedo witch veo 3 quality ?

[fiery gull]

luck

[cerulean mica]

?

opal ?

[fiery gull]

#1397655624103493813

and luck

just it

[barren prairie]

I was here since 600 members 😂😂

[fiery gull]

hm

[barren prairie]

2.5 pro is trush at coding can t make a simple 1k lines of code without 15727bugs

[fiery gull]

I was here since 96k 🔥 ]

[fiery gull]

Its my opinion, I don't use code

I just use creative write

[barren prairie]

Yeah it is good. I Love the explanation that Gemini gives it is so detailed and easy to understand 👌

[fiery gull]

the personality too, is so useful in my work

[coral goblet]

guys

what the best AI for emotional voice ?

I use it to make conversation

[fiery gull]

the new 2.5 flash voice from aistudio is good

or gpt but is only 10 minutes free

[coral goblet]

oh i didnt there is text to speech in aistudio

what is it called ?

nvm i found it

[fiery gull]

stream

ative affective dialog

[coral goblet]

uh i mean like

i want to generate a podcast

not talk to ai assistant

[fiery gull]

ahhhh

[fiery gull]

gemini speech generation

[ionic trail]

I WAN T TO GENERATE VIDEOS

[granite heron]

check #1397655624103493813 to learn how to prompt he bot

[celest pecan]

helo

[limber wind]

hi

[humble sonnet]

Hello, when I generate a second image in the same chat does it necessarily take the image above to modify it?

[mental hemlock]

hello

[alpine sparrow]

hello everyone

[maiden fractal]

hi

[fiery gull]

Hi

[soft kernel]

OMG REAL

[stray aspen]

gemini 25 fash latest memory SUCKS

it cant remember messages inside a conversation

[keen flume]

hello

[fading rover]

hey seedreamers i would like to check few test on 4k output is there anywhere can get trial credit to test?

[remote arrow]

I believe there isn't too many weird prompt scattering all over channels like nowaday back then..

[acoustic sundial]

hello

[sour holly]

why my videos dont have audio?

[remote arrow]

Not all video generators have ability to add audio. And here, they're picked randomly..

[old vessel]

hi everyone. Am i correct that it is not generating all videos, but only chosen ones?

[knotty fable]

Yeah correct, it's a limited sample of videogenerating AI's here.

[languid crescent]

hi is lmarena down?

[knotty fable]

And the ones that suit my need most, aren't even represented here.

[languid crescent]

Is Lmarena down? I'm stucky at verification

[fiery crater]

hello guys i just arrived

[lone spindle]

somebody here knows how to createa video with LM arena?

[remote arrow]

Read #1397655624103493813

[zenith horizon]

Hello am happy to be here and to learn more?

[remote arrow]

It generates videos with random generators. Option to choose which generator is better, popping up AFTER the videos generated. So, doesn't matter which one you choose, the videos are already made and downloadable.

[quick jackal]

Hello @lone spindle you can go to https://discord.com/channels/1340554757349179412/1397655624103493813 to learn how to use the bot and https://discord.com/channels/1340554757349179412/1397655695150682194 https://discord.com/channels/1340554757349179412/1400148557427904664 https://discord.com/channels/1340554757349179412/1400148597768720384 for your creations.

[remote arrow]

He's already pooped his prompt on other channel..

[verbal nimbus]

Is the model only available in AI Studio?

Flash on Gemini app seems worse

Forgets conversation or keeps repeating itself