general

[ocean vortex]

have you tried this? https://prism.openai.com/

[magic imp]

I know I am saying that g3 pro did mistake

[ocean vortex]

literally entire product made just for this lol

[modern wedge]

Bruh i just found out MAX AI is kinda good ngl is almost like Opus

[magic imp]

Seriously

[golden ocean]

Claude Opus 4.6

[magic imp]

Yaah like little but seriously that iv2 didn't even feel

Haha grok still have to work on it

But if you notice the user name is same user name in chat

[half yew]

Another image I gen using iv2

[hollow mulch]

Hello guy

[golden ocean]

what is image v2's codename

[half yew]

But like a mini version

[half yew]

Packing tape, gaffer tape, masking tape

[golden ocean]

thank you sir

[hollow mulch]

Fix gork multiagent beta is reach limit after one message

[golden ocean]

gork

[hollow mulch]

Please 🥺 @echo aurora

[echo aurora]

Context limit?

[magic imp]

Bro add grok multi image edit modal we still can upload only single image

[inland quest]

"Maximum tokens limit reached in that conversation"

[gray isle]

i use overleaf and that

[inland quest]

Grok fr has insanely low size context in arena

Especially in search

[misty harness]

https://tenor.com/view/tiktok-shaquille-shaq-timeout-ethical-gif-9311067668442735802

thanks to arena funders and other team for giving us access to top models for free

[hallow siren]

So the top models are getting away !

[ocean venture]

ye

[cursive creek]

Gemini is causing that for me rn-
it lets me do one thing
then it reaches token limit
i leave and return, use another model, it works
go back to gemini, token limit before the prompt is even processed-

[hallow siren]

Well time to interact other models too , I think it's good time to do that atleast we can learn how each models behave atleast. It's tough to maintain the services runing for free for long time

[arctic warren]

eh what happened to gemini 3.1 pro?

[hallow siren]

Still believe arena will comeback

[arctic warren]

saw it got removed

[hybrid ridge]

no

[hallow siren]

They are removing heavy models from the platform to maintain others

[arctic warren]

i see

[hallow siren]

Is the community dead ?

[severe swift]

Hey guys.

I'm working on an autonomous agent, the kind where you give it a goal and walk away. Browser, terminal, full OS, real internet access etc.

I'm curious, what's something you've wanted AI to just handle end-to-end, not help with, actually handle, where every tool you've tried still left you doing half the work yourself?

I'm open to fun ideas too.

[pseudo hemlock]

hi

[sterile tartan]

Build AGI

[severe swift]

ok well, let's say it's an AGI with autonomous capabilities. What'd you want it to do?

[sterile tartan]

Build ASI

[hollow willow]

They released all the modules like Opus 4.6 and the most powerful ones, right?

[pseudo hemlock]

wdym

[severe swift]

ASI, nor AGI, is a software. I explicitly said when the agent has autonomous capabilities. Never once said it has access to a full billion dollar data center lmao

[echo aurora]

Unfortunately, if you're running into this error message there currently isn't the ability to expand that context limitation to resume that chat session. There is more information that can be found in this article: https://help.arena.ai/articles/3975292349-arena-troubleshooting-session-token-limits. Unfortunately, this means starting a new chat session is your best next step.

[pseudo hemlock]

hi pineapple

[sterile tartan]

In that case Build and Handle a Multi Billion Dollars Business

[pseudo hemlock]

how u doin

[hollow willow]

A few days ago, I could select "Direct" and choose the Claude Opus 4.6 module, but now it's gone, only Sonnet is available. The same goes for GPT 5.4-high; now you can only choose 5.2.

[echo aurora]

Little sleepy, hbu?

[pseudo hemlock]

Yea, the really expensive models are temporarily disabled while smart people try to figure out how to make it more sustainable.

[severe swift]

again, realistically, let's say it has terminal and os access, never said it has credential or accounting access, or investment

[echo aurora]

There is more information about this removal in this announcement: #announcements message

[pseudo hemlock]

bored, i'm in class and it is sooooo boring

[sterile tartan]

Alright give me a Absolutely Through Analysis of The Stocks and Crypto

[severe swift]

see that's a great idea

[hollow willow]

Are you saying they're coming back? I regret not taking better advantage of it :/

looked for other similar platforms, but they require API credits to use them.

[pseudo hemlock]

Arena's people is looking into what they can change in the future to make everyone happy

If you want you can look into #announcements message and #1491461236448170134

[echo aurora]

We do want to bring the models back to Direct and Side by Side when it can be done in a sustainable way.

[sterile tartan]

When Money is on the Line the ideas are always great

[echo aurora]

Pay attention in class! Get off Discord!

[sterile tartan]

Ik is just very expensive

[pseudo hemlock]

Discord > class

[severe swift]

Yeah but perplexity ai is capable of that. It's not really anything special. I'm talking about independent tasks, like a human employee could do.. with the constraints I mentioned of course

[pseudo hemlock]

I don't pay for tuition so this class is free

[sterile tartan]

At all cost arena.ai must not shut down like Yupp Aj

[hollow willow]

thxx @pseudo hemlock @echo aurora hope there's an alternative. I think the only option is to install a GitHub repository locally, but I don't know much about that and I don't want to risk it jajajaj

[sterile tartan]

Probably would be like Leads Handling

[pseudo hemlock]

Don't worry, I have my best people working on it

[severe swift]

Great idea again dude ❗

[pseudo hemlock]

Make me something that cold emails the hiring manager of a company and role i put in

(I need a job)

[golden ocean]

cold call instead

[pseudo hemlock]

have it copy my voice and then cold call them

yes

[hollow mulch]

Lm arena is peak but im know they still imporiving it

[golden ocean]

that technology just exists

do it

[pseudo hemlock]

can you do it for me

im lazy

[golden ocean]

sure
send me 100 hours of ur voice for training purposes

[sterile tartan]

Cold Calls/Emails another Great Idea indeed

[pseudo hemlock]

ok

do you want my SSN while you're at it?

[golden ocean]

yes

[severe swift]

I see the direction ngl. but, this has edge cases, since the hiring manager of the company, would first off all have to be available on the internet, with their contact info specifically. secondly, your identity also can't be confidential with this approach, which also means that if, your resume or history isn't a great first-impression, the ai would try but ultimately make a bad impression, with your actual credentials too..

[pseudo hemlock]

i mean 99% of them will be on linkedin

with that info public

[severe swift]

yeah but what are the chances they reply to YOUR ai agent anyways

even if the ai agent is capable, us HUMANS get ignored..

[pseudo hemlock]

because my ai agent will be so amazing and get 100% response rate

[severe swift]

I mean sure if you're willing to wait months before the AI agent creates a fake identity/persona, creates social media recognition, gets great reputation, and so, 100% response rate..?

considering the fact that, running the agent for a month, comes with its own costs anyways

[obsidian cargo]

watching the triple i initiative premiere while waiting for gpt image 2 to drop

[golden ocean]

do u get leads manually from linkedin or u scrape them

[pseudo hemlock]

idk

[golden ocean]

https://tenor.com/view/kitten-police-chase-on-the-run-escape-off-road-gif-15122162029294723139

[still musk]

GPT image 2 may be released in a few minutes

[golden ocean]

so true

[obsidian cargo]

20 is more than a few

[golden ocean]

source

[severe swift]

guys.

I'm working on an autonomous agent, the kind where you give it a goal and walk away. Browser, terminal, full OS, real internet access etc.

I'm curious, what's something you've wanted AI to just handle end-to-end, not help with, actually handle, where every tool you've tried still left you doing half the work yourself?

I'm open to fun ideas too. {yeah copy pasted it again}

[obsidian cargo]

Start a company, make money ethically, pay for its own existence, use the leftovers to lobby for ubi

[golden ocean]

Start a company, make money unethically, pay for its own existence, use the leftovers to lobby for ubi

[severe swift]

I already answered this previously. realistically, let's say it has terminal and os access, never said it has credential or accounting access, or investment

[obsidian cargo]

ends justify the means lol

[pseudo hemlock]

My professor put on the wrong video for 45 mins

🥲

[golden ocean]

https://tenor.com/view/dragon-warrior-kung-fu-panda-rtx-gif-19233196

[severe swift]

Its, lets say, like Jarvis, but without a human-like passport or legal docs it needs to be able to do tax-related things.. yeah

[pseudo hemlock]

Make me Jarvis

[severe swift]

ok. what do you want it to do

[pseudo hemlock]

I want it to download me free ram

Find me free ram online

[severe swift]

great idea

[pseudo hemlock]

Then download it

[severe swift]

it can also do tax evasion for you

[pseudo hemlock]

LETS GOOO

unfortunately I already submitted mine for this year

[golden ocean]

https://tenor.com/view/cat-smile-happy-happy-cat-smile-cat-gif-18259408098481512759

[severe swift]

it can make you iron man

[pseudo hemlock]

Cute

[severe swift]

🙂

[pseudo hemlock]

Can it make me a palladium electromagnet heart

[severe swift]

no but it can make you a heart out of ASCII art 🙂

[pseudo hemlock]

😮

Do we not have anything better to do than this

[severe swift]

idk man im asking you for what you want it to do

[pseudo hemlock]

Uhhhh

[severe swift]

realistically

[pseudo hemlock]

Idk man

[severe swift]

lmao

[pseudo hemlock]

Claude solves all my problems

I have $300/month

[severe swift]

any problem it CAN'T solve?

[pseudo hemlock]

Which isn’t enough but like

[pseudo hemlock]

Do YOU have any ideas

Like what are you wanting to do

[severe swift]

Well, I do have some

and I did realize them already, in reality

[pseudo hemlock]

What ideas

[severe swift]

well, technically, claude can do some employee-level work, including scheduled tasks.

but you can't really trust it, by telling it 'get feedback from my email, customers, manage my site's database every day, 24/7, ensure you fully eliminate and analyze all competitors that ever come and go against my platform'

etc. smth like that.

[clear copper]

Wait arena removed Claude opus models from the list

[severe swift]

so, I did smth like that. cool?

[pseudo hemlock]

Yea, hopefully only temporarily

[clear copper]

Ok

[pseudo hemlock]

They’re looking for a more sustainable solution

Also removed gpt 5.4 5.4 high and Gemini 3.1 pro

[pseudo hemlock]

What can Claude do in terms of this? I’ve never scheduled tasks

[severe swift]

best it could do is, scrape your site's data every morning with 'scheduled tasks'.. not much further than that. probably could give you a report, telling you what could improve..

what I proposed tho is basically an autonomous business employee, full time essentially

[pseudo hemlock]

how do you do that

local model?

with basically cron jobs?

[severe swift]

no.

[pseudo hemlock]

oh

[severe swift]

Model routing, orchestration layer, it's a bit technical

[pseudo hemlock]

lets hear it

[severe swift]

I did achieve it. I'm just trying to understand the market right now

[pseudo hemlock]

open source?

[severe swift]

proprietary, autonomous orchestration layer

could provide you a demo vid if you want

[pseudo hemlock]

Sure

[severe swift]

wait actually I do have a new recording but its too long.

[pseudo hemlock]

basically when X happens, model does Y

[severe swift]

but this vid shows a much older version of the agent. I've re-engineered several parts of it, including the ability to do tasks without human interpretation.

[pseudo hemlock]

What model are you using

[severe swift]

it has model routing system. also, kind of technical... It uses the best possible released model, from top providers (e.g. anthropic, openai, google), uses the fastest latency model possible for which task fits the category of efficiency, specific tools that require brain-storming, etc.

[pseudo hemlock]

got it, that makes sense

so its basically max

[severe swift]

yeah no, you don't gotta make sense of it much

just think of it as, you get the best quality, without paying to just a single provider, or depending on a single provider's model.

[pseudo hemlock]

So you're a middleman

[severe swift]

but also, the fact that, its an agent, not just one-shot LLM, so, its an autonomous business employee

[severe swift]

no... no.. not merely middleman.. its basically an autonomous business employee, full time. not a wrapper

I think its better if you ask claude what autonomous orchestration layer could mean, in AI

[pseudo hemlock]

No I think I get it

it routes to which model, it calls tools when needed

it breaks up tasks into subtasks

etc.

[severe swift]

and, is able to understand general issues, fix them by itself, without asking you

[thick blade]

Jarvis model incoming

[severe swift]

so, if it sees, multiple users say a feature isn't working, it doesn't need you to remind it to fix it

btw has any of you heard of the claude mythos model?

[thick blade]

Whats even claude for

[severe swift]

for, being good boy 🙂

[thick blade]

.

[hollow mulch]

Beside code in arena is good is help me make web so much 10/10, execpt kat pro coding

[severe swift]

your english is giving me a stroke bro

[still musk]

It's 7:04 PM and GPT image 2 hasn't been released yet

[obsidian cargo]

arghghghhhh tears out hair

[pseudo hemlock]

claude mythos is crazy

[severe swift]

wanna know a secret on claude-mythos?

[sullen creek]

guys, lets just run local llms

[pseudo hemlock]

sure

[severe swift]

claude mythos is, 80% hype

[pseudo hemlock]

how do you know

[severe swift]

because I work with AI output patterns constantly

and I've seen their documentation of mythos

[pseudo hemlock]

you think finding problems in linux that has existed for years is hype?

[severe swift]

the benchmarks are genuinely impressive, but every single thing stated as a 'shocking factor' is genuinely not

[pseudo hemlock]

or an OpenBSD vulnerability that existed for i think 27 years

[severe swift]

it sounds impressive right?

[pseudo hemlock]

that could crash a server

[severe swift]

yeah I know it sounds very impressive

but they key here is, claude mythos, was not in a base environment during this

they frame this like they gave birth to an Alien they never predicted

its not

the agent, was within a specific 'Anti-Sycophancy' agent environment

during testing, the setup was coordinated in the direction they wanted the agent to be able to constantly find unique strategies, to existing environmental patterns

Ik my words sound kind of unbelieveable here

[vital mantle]

Where is the meta model?

Why I can’t see it or test it

[severe swift]

but wait till you understand the fact, how much it costed Mythos to run it constantly, constantly letting it iterate, constantly running a loop ensuring it finds 'shock factors' in the execution environment

[pseudo hemlock]

Meta.ai

Website

[severe swift]

btw whats the pricing of the new meta spark model?

[vital mantle]

Yes but I want to see the stats on arena

[pseudo hemlock]

Not available via API yet

[vital mantle]

It’s free

[pseudo hemlock]

Only meta.ai

Or Facebook or instragram

[vital mantle]

Meta made a whole website that’s it the best it’s true or no

[pseudo hemlock]

Hey I mean us humans haven’t figured out we’re in a simulation yet

[pseudo hemlock]

It’s definitely up there

HUGE step up from llama 4 😂

I messed around with it yesterday and it’s solid

Haven’t tried anything crazy

[vital mantle]

I’m paying 200$ on Claude opus 1M should I go to meta AI

[vivid coral]

is the new Meta open-source or no

[pseudo hemlock]

I mean try it out, it’s free for now so 🤷‍♂️

[pseudo hemlock]

No

[vivid coral]

whoa

[severe swift]

again, they framed it very interestingly. They key is, its a Language model, NOT a human employee. It didn't find Consciousness, it was just asked about it and it gave a answer based on its statistics

[pseudo hemlock]

It’s supposed to be SOTA, so not open source

[pseudo hemlock]

But that’s all LLM right now

That’s all ML right now

[severe swift]

yeah, but see the anthropic's model card. they're framing it as alien model

[pseudo hemlock]

It’s a step up from what we have, not an alien

Especially in the cybersecurity space

[pseudo hemlock]

Why would Palo Alto be partnered for something that isn’t amazing?

They’re THE cybersecurity people

[severe swift]

its a step up in benchmarks. In specific software engineering cases (again, the word Specific is crucial here). but again. Alright look this sounds confusing, but wait till Mythos is released, lmao you'll see what I mean in a few months

[vital mantle]

So do I need to cancel my Claude? I’m on the 200$ plan

[pseudo hemlock]

You can call it hype all you want, but these people wouldnt be partnered if it wasnt as good as they say

[pseudo hemlock]

Give muse spark a shot

and then decide

I only talked with it, no coding or real hard problems

[vital mantle]

Yes it’s good and free but the dispatch Claude option is also good

[pseudo hemlock]

also if you use claude code or anything, it can't be used like that

like it only has web search + google calendar + gmail + outlook

also it has a vm

[vital mantle]

So I can’t use it like Claude on desktop

[pseudo hemlock]

I don't think so

[vital mantle]

So it controls my pc

[pseudo hemlock]

Only available on meta.ai, facebook, or instagram

[vital mantle]

Yes I asked it months ago what’s the new iPhone it answered wrong now it knows

[pseudo hemlock]

I asked it yesterday, it has:

Web search
Social search (Search facebook and instagram posts)
Image generation
Web artifacts (HTML websites, actually hosted and you can interact with them)
Code sandbox (python environment)
Subagents (up to 24 in parallel)
Third-party linking (google calendar, gmail, outlook, they're read-only)

[vital mantle]

So the free version has 24 subagents?

[pseudo hemlock]

I tried it yesterday and it only ran 3 but I didn't ask anything crazy

I just said prove to me you can run multiple subagents

[vital mantle]

How can I make it run subagents

[queen veldt]

In what did you ask? On Facebook or?

[severe swift]

Ok, I want you to stop viewing it as like they partnered because they thought this beats every other thing on the planet.

Anthropic didn't secure those partnerships because Apple and Microsoft think they've birthed an alien god; they got those logos because Anthropic handed out $100 million in free usage credits for these companies to fuzz their own infrastructure for zero-days. Not to mention, being included in this project is also a huge credit in the AI market, showing that these companies are apparantly 'contributing' to the advancement of AI research.

If a vendor walks into Microsoft, Apple, or JPMorgan and says, 'Our new model is exceptionally good at finding 15-year-old CVEs, and we will pay you $100M in compute to let your red teams test it on your own code before hackers get it,' every single CISO on earth signs that paper. That isn't a testament to its capability; it's basic corporate risk management and liability shielding. They are using it as an offensive cyber-tool, not bowing to its sentience.

[pseudo hemlock]

You can say "run subagents to ...", or if you ask something difficult enough it'll run them automatically

[pseudo hemlock]

Meta.ai

[cunning heron]

when gemini 3.1 pro will comeback?

i need that

[pseudo hemlock]

I agree, but if the model SUCKED, these companies wouldn't want to put their name next to it

[severe swift]

honestly You are confusing the deployment harness with the model weights.

The reason it operates in a VM with a restricted toolset isn't because it's some incomprehensible new lifeform—it's because it's a standard text-prediction engine wrapped in a highly restrictive execution substrate. They have to sandbox it in a VM because they are running a loop where it generates and tests live cyber-exploits (like the 181 Firefox exploits it wrote). If they gave it unconstrained apply_patch or terminal access on a live host, it would nuke the system.

[pseudo hemlock]

TBD

[pseudo hemlock]

bro what, i never said anything about it running in a VM

[vital mantle]

Only 4

There is no premium

[pseudo hemlock]

Try something harder maybe? Not sure if we're limited to 4 rn and they advertise 24 but thats not available yet

[obsidian cargo]

STILL no gpt-image 2? guess that 1:00 PM prediction was fake then :(

[pseudo hemlock]

who said its coming out today?

[vital mantle]

I have gpt codex also so what you recommend me to do

[obsidian cargo]

someone shared a few tweets saying so, including one that it'd drop at 1:00 PM

[vital mantle]

Use meta or Claude

[pseudo hemlock]

Do you have ClaudeCode, or just Claude.ai?

If you're doing coding the harness (codex or claudecode) is very helpful

meta is only available online

[severe swift]

You did say vm dude. Nobody is saying the model 'sucks.' It’s obviously the current State of the Art (SOTA). But you’re confusing a Defensive Liability Shield with a Product Endorsement.

Look at the specific nature of Project Glasswing: Mythos found a 27-year-old bug in OpenBSD and 181 Firefox exploits. If you are the CISO of Microsoft, Apple, or Google, and a lab tells you, 'We have a model that can autonomously find vulnerabilities in your OS that have been hidden for three decades,' you don't partner with them because you think the model is AGI.

You partner with them so you aren't the only one left outside the bunker when the disclosure hits.

If Apple wasn't on that list, and Mythos dropped a zero-day exploit for macOS tomorrow, the board would fire the executive team for negligence. Joining the consortium is a defensive PR requirement, not a testimonial of 'alien intelligence.'

[pseudo hemlock]

This was about meta spark

[cunning heron]

is there anywhere else I can use that model for free without limits like here?

[severe swift]

oh my bad

[vital mantle]

I have Claude max 20

[pseudo hemlock]

I'd be very surprised considering it is an expensive model

[vital mantle]

and codex the 20$ one

There is a 200$ codex one but not sure if worth it

[pseudo hemlock]

Personally I only have ClaudeCode and I love it

But I also have an enterprise plan through my university

so I don't pay for anything

so

If you have a decent computer I'd say check out OpenCode (or claude code w/ ollama) + a local model

[vital mantle]

I had grok also the 30$ one now that meta is out not sure which one to go for I also tried Gemma 4

[cunning heron]

do u know some another model thats good for coding, html css js, for fivem scripts?

[pseudo hemlock]

Claude Sonnet 4.6 is pretty awesome

Sonnet 4.6 is better than gemini 3.1 pro according to arena, https://arena.ai/leaderboard/code

[vital mantle]

I use the 1M one

[pseudo hemlock]

Thats the expensive one

If you don't NEED 1M context, don't use it

[wicked talon]

Meta ai copying Gemini lol

[vital mantle]

on medium not sure what the effort one do

[cunning heron]

i try that but this doing me bad, i use the same prompt like on gemini 3.1 pro, and claude doing me so much bugs and worst things : D

[pseudo hemlock]

You can try qwen 3.6 plus?

[vital mantle]

It’s good

[cunning heron]

same haha

[vital mantle]

I have 2 more Claude acc on 20$ plan

[pseudo hemlock]

but expensive

[vital mantle]

Yes 200$

[pseudo hemlock]

try glm 5.1, or gemini 3 flash is pretty good too imo

[vital mantle]

I like the Claude dispatch option and it has channels also

so no open claw? I guess

[pseudo hemlock]

I have no idea what dispatch or channels are

and have never used openclaw

or whatever it is called now lmao

[vital mantle]

Dispatch on Claude is you use phone tell it to do something and it message you back

[pseudo hemlock]

Ohhh

Claude has that? Like natively?

or you mean OpenClaw

[vital mantle]

Yes on phone and Mac

Claude

[pseudo hemlock]

interesting

wish I had a real plan instead of enterprise lol

everything is managed and turned off for me 🙁

[ocean venture]

anybody had been trying this? give a simple review on it

[pseudo hemlock]

Solid

[thorny schooner]

😭 of course I got to verification lool why the hell is it on even retries

[thorny schooner]

I haven't what is it?

[pseudo hemlock]

Meta's new model

Muse Spark

Benchmarks are pretty crazy

[ocean venture]

Better than GLM and Qwen?

[pseudo hemlock]

better than everything lmfao

[ocean venture]

Meta ai newest ai

[vital mantle]

Better then mythos

[pseudo hemlock]

besides mythos

[ocean venture]

mythos was unbelievable OP dawh

[pseudo hemlock]

not sure how much of mythos is just hype

if my brain was able to output $100M in tokens id probably find some bugs too

[vital mantle]

Claude is telling me that opus 1M is the best but that gemin 3.1 pro preview is the winner

[ocean venture]

i need that site, did you have it? or is it from SWE Bench?

[pseudo hemlock]

https://ai.meta.com/blog/introducing-muse-spark-msl/

[ocean venture]

Dang

[pseudo hemlock]

Dayumn its more censored than claude

[hollow mulch]

Meta ai?

[pseudo hemlock]

thats impressive

[pseudo hemlock]

yep

[hollow mulch]

Link please 🙂

[ocean venture]

https://www.meta.ai/

[spring oar]

this is not possible ?

[pseudo hemlock]

Wdym

[spring oar]

Gemini is better at vision

[pseudo hemlock]

I guess people disagree

[vital mantle]

Why it’s not on the list

The meta?

[urban trench]

Totally not benchmaxxing.

[covert totem]

Why isn't Claude Opus not on the website?

[pseudo hemlock]

It is new and not available via the API

[flint zenith]

he is very expensive

[pseudo hemlock]

I mean people are the ones voting, idk how you can "benchmax" that

[pseudo hemlock]

Expensive, people are using a lot of it, Arena people are trying to figure out new usage limits

[flint zenith]

However, they removed all the bad models, and now the best one is the Claude Sonnet 4.6

[pseudo hemlock]

For the time being Claude Opus 4.6, Gemini 3.1 Pro, and GPT 5.4 / 5.4 Pro aren't available on the website

[hollow mulch]

Did meta have any model?

[spring oar]

Good benchmark how you see this ?

[pseudo hemlock]

https://ai.meta.com/blog/introducing-muse-spark-msl/

[pseudo hemlock]

Meta has a new model yes

[novel plover]

Damn it.. Opus 4.6 has been removed..?

[pseudo hemlock]

for now, yea

[spring oar]

What ?

[spring oar]

I dont understand ?

[hollow mulch]

What button i can changing to new model

[pseudo hemlock]

Opus 4.6 isn't availalbe on Arena.ai right now

[pseudo hemlock]

go to meta.ai, it isn't available on arena.ai

[novel plover]

Claude model man..

[covert totem]

its over

[pseudo hemlock]

I dont think so lol

[novel plover]

Sheet.. Only Pain now ..

[vital mantle]

It’s on chat

[spring oar]

Why opus is removed i can use it

[covert totem]

where

[vital mantle]

Before I asked it last iOS version it didn’t knew

[covert totem]

also i didn't know there was a limit

I thought we could chat as much as we want to

[pseudo hemlock]

Nope, there is a limit

[novel plover]

Nope.

[covert totem]

how many messages per day

[pseudo hemlock]

It costs Arena money every time you send a prompt

[pseudo hemlock]

🤷‍♂️

[ocean venture]

That was sonnet bro

[hollow mulch]

So ummm did muse spark you guy mention is good at roleplaying?

[covert totem]

yea but they are using Our data to train Ai or whatever

[pseudo hemlock]

Go try lol

[hollow mulch]

Okay lol 😄

[pseudo hemlock]

When you use the website you agree for your chats to be saved but not sure what they do with them

[ocean venture]

https://tenor.com/view/my-mom-is-kinda-homeless-ishowspeed-speeding-please-speed-i-need-this-ishowspeed-trying-not-to-laugh-gif-16620227105127147208

[vital mantle]

so without api it can’t track or

[covert totem]

😭😭🥀

[hollow mulch]

Theu still find way bring opus back don't worry

If they doing something and become rich 🙂

[spring oar]

you know why ?

[hollow mulch]

█░░ █▄░▄█
█░░ █░█░█
▀▀▀ ▀░░░▀
▄▀▄ █▀▀▄ █▀ █▄░█ ▄▀▄
█▀█ █▐█▀ █▀ █░▀█ █▀█
▀░▀ ▀░▀▀ ▀▀ ▀░░▀ ▀░▀

[pseudo hemlock]

Money, expensive

[spring oar]

Ok thanks but you know why ?

[hollow mulch]

They can't afford more

[split hamlet]

I can't see gpt 5.4 and claude 4.6

[spring oar]

is why opus is on the top in llm arena

[novel plover]

https://media.discordapp.net/stickers/1431919271386026087.gif?size=160&name=11satta_harne_ka_sadma

[hollow mulch]

Is removed bro

[vital mantle]

24?

[pseudo hemlock]

yep, 24.

supposedly

[spring oar]

Llm arena is reliabel benchmark or is not reliabel ?

because some people can cheat

[pseudo hemlock]

Arena is working through new usage limits, while that is happening Claude 4.6, GPT 5.4 / 5.4 High, and Gemini 3.1 Pro are disabled

[vital mantle]

Thank you

[pseudo hemlock]

Opus is really big and expensive

[vital mantle]

😂

How much Claude has

[split hamlet]

Okay Thanks for your kind

[novel plover]

damn it.. Yuppai is also winding down..
No opus there too..

[pseudo hemlock]

Damn what did you ask it to do hahahah

[vital mantle]

can you use all 24 agents to research x

[novel plover]

Cheat Activation got Real

[split hamlet]

wow

[vital mantle]

damn meta is advanced

[normal abyss]

nah we need 50 research agents 💪

[pseudo hemlock]

Does it give you sources?

[ocean venture]

Use A Jailbreak ai script from reddit, i wanna see if new Meta ai was affected by the script 😂

[vital mantle]

Yes but I don’t have the speech option or that it read me the text

[inner gate]

Ive always wondered. What even are agents? In this contexr

[pseudo hemlock]

It’s basically x instances of the LLM working in parallel

All doing subtasks of the main task

[inner gate]

Does it increase the speed of it?

[pseudo hemlock]

Yes because 1 task is broken into 10 smaller tasks and instead of running 1st sub task, then 2nd sub task, then 3rd sub task, etc, all 10 subtasks are running at the same time

Makes sense?

[obsidian cargo]

so this leak was fake

[vital mantle]

@pseudo hemlock so which ai you recommend

[pseudo hemlock]

uhh

i like claude opus the most but that is also because i have really high limits

but gemini is also great, but my limits are really low

[vital mantle]

gemin 3.1 pro preview ?

[pseudo hemlock]

yea

[covert totem]

claude sonnet always cooking

[golden ocean]

{"error":"Selected model is not available for user selection"} 💔

opus 4.6

[golden ocean]

true

[covert totem]

peak pfp

[vital mantle]

@pseudo hemlock I guess gemin is right but 1M one is also good and the new meta one

[covert totem]

they removed it i heard

too expensive someone said

[pseudo hemlock]

What website is this?

[pseudo hemlock]

Opus 4.6, Gemini 3.1 pro, and gpt 5.4/5.4 high are temporarily disabled

[vital mantle]

Claude 1M made the graph

[pseudo hemlock]

Oh lol

Yes that’s just wrong

Muse spark doesn’t have an elo

Don’t trust anything in that

[vital mantle]

Okay so I’ll use gemin for questions since it has highest intelligence and Claude for coding

[pseudo hemlock]

Looks like artificialanalysis already benchmarked Muse Spark

Pretty damn good

[vital mantle]

Gemma 4 3b thinking has also 52 intelligence

[pseudo hemlock]

what?

Gemma 4 3b isn't a thing, it doesn't exist

[vital mantle]

Yes it does on Ai studio

[pseudo hemlock]

31b?

[quaint ocean]

does the site down ?

[pseudo hemlock]

No, works for me

[vital mantle]

@pseudo hemlock

[echo aurora]

What are you seeing?

[pseudo hemlock]

Yea, there is Gemma 4 26B A4B, and Gemma 4 31B

a MoE and a dense model

But those aren't 52 intelligence score on artificialanalysis

[vital mantle]

They updated the site now

It looks different

[ocean venture]

I had try to make a simple website on Muse Spark(Meta), i think this was quite decent (Simple test :v) https://embed.fbsbx.com/playables/view/765147019862658/?ext=1783535657&hash=Q92gDAFhSzkiyyvkVnYkb95fe8II

[pseudo hemlock]

Looks good

also i love that they give you actual working websites

not just the html code to run locally

[vital mantle]

Your picture looks different now

[vast fern]

where to try it

[pseudo hemlock]

Meta.ai

[vital mantle]

Claude told me meta ai has 1b users now lol

[pseudo hemlock]

You’ve gotta try this crazy new thing

It’s called Google

[vital mantle]

Google says the same 1b maybe yesterday everyone went on it

[pseudo hemlock]

Oh yea meta meaning Facebook + Instagram maybe

Not meta.ai

[ocean venture]

I want to try Deepseek V4 but idk, it's not out yet...

[vital mantle]

Yes meta ai 1b users

[pseudo hemlock]

That’s just wrong

[vital mantle]

Yes it’s in the news

[vital mantle]

Maybe becouse people can use it via instagram Facebook WhatsApp and so on

[obsidian cargo]

ughghgh when is gpt image 2 dropping!? :C

[meager harbor]

Will claude mythos be tested on arena ?

[echo aurora]

I won't be able to share details about what models could be landing on the platform.

[vast fern]

its on pc deepseek website

[vast fern]

peak ragebait

??

bruhh

[inner relic]

I think there's deepseek v4 secret model on lmarena

go find it on #codename-discussion

Iguess

[ivory latch]

i have a doubt like thorught claude and all can we vibe code and make a app and publish it on apple store any guide and
is any limit to talk in claude in areana

[prime karma]

@echo aurora hello

[ivory latch]

hey

[echo aurora]

[next trout]

meta's ai is so dumb

[polar horizon]

g**n

[prime karma]

As I was using Gemma 4, it will work just like the new Google AI Edge Gallery. Based on the same model

[polar horizon]

catastrophic typo

[prime karma]

@echo aurora

[prime karma]

@echo aurora please reply

[echo aurora]

Sorry what's the question?

[prime karma]

As I was using Gemma 4, it will work just like the new Google AI Edge Gallery. Based on the same model @echo aurora

[ivory latch]

@echo aurora Hi, I’m trying to build an app and I’m concerned about hitting usage/API limits during development.

For example, if I’m halfway through building the app and I reach the limit, what are the best ways to handle it?
Should I upgrade the plan, optimize usage, or are there other recommended approaches?

It would be helpful if you could guide me on how developers usually manage or avoid these limits while building apps. Thanks!

[echo aurora]

Sorry to say I'm not familar with Google AI Edge Gallery so I couldn't say for use. The models you see on Arena are going to be what is provided via the model provider's API.

[ivory latch]

@echo aurora ?

[echo aurora]

Hey @ivory latch the session context limit is going to be in place, and there isn't the ability to upgrade to a plan that'd expand this. That being the case, I'd recommend adjusting your prompting keeping this context limit in mind.

[ivory latch]

means be optimize usage,

[echo arch]

Any news?

[spring oar]

Better for study gemini 3.1 pro or opus 4.6 ?

More reliabel

[ocean vortex]

Yeah this is kind of unexpected

People were still waiting for Behemoth 🗿

[twin zinc]

i think gemini

[abstract hinge]

Simple: optimize usage.

The best way to do this is by using an IDE—like VS Code with Codex, Antigravity, Cursor, OpenCode, or other Claude Code-style tools. Using an IDE helps optimize the model’s context and can reduce token usage, since it won’t need to read thousands of unnecessary lines of code.

[ocean vortex]

My guess is that it's going to be smth like Grok. Resources allocated to make it look good on the charts. Not a ton of substance but still solid/decent thanks to sheer amount of tasks tested in those benchmarks. Yet to test it in-depth though

[abstract hinge]

And depending on the task, you don’t always need the top model on the market. GLM 5.1 is a great model—maybe a bit below Sonnet 4.6—but it’s still very solid. If you use it properly, you can build a full project with it, and it’s much more cost-effective than Claude.

[brazen thunder]

Are claude 4.6 and gemini pro permanently gone in arena?

[vital mantle]

For intelligence?

[abstract hinge]

Both models can fail sometimes. I mean, Claude Opus tends to fail less, but it’s more expensive to use. One possible approach is to use Gemini and always ask it to validate the output with another model.

[ocean vortex]

Study in general --> 3.1 Pro is the best general purpose model out there overall tbh

Only a small part of that would be an actual coding, where Opus is probably the best

[echo aurora]

In the announcement (#announcements message) we mention it's our intent and hopes to bring them back to Direct & Side by Side when it can be done in a sustainable way.

[abstract hinge]

The problem with Gemini is that it tends to hallucinate, and sometimes it’s so confident that it’s right when it’s actually completely wrong.

[vital mantle]

@echo aurora Why there is no meta stats

[ocean vortex]

Pro can do same amount of work with near 2 times less tokens, better fundamental understanding of the world and relating logical principles as well. Hallucinations are something to keep in mind but those are manageable

[vital mantle]

I want to know which ai to use for intelligence

[weak dagger]

@echo aurora new pfp?

[vital mantle]

also the 1M what score does it have

[brazen thunder]

Oh yeah I saw now, thank you

[weak dagger]

What's this ai again

[upper remnant]

I’d say Opus also outshines it in terms of creative writing.

[ocean vortex]

Well you do need to be aware of it's limitations. But this applies to all models so it's hard to draw the line somewhere.

[vital mantle]

Claude the 200$ plan

[abstract hinge]

Could you implement a system to attach .txt or .py files?

I actually built a simple console script that does this, and it helps me a lot. What I did is very simple and could be used on the site, since the file itself isn’t sent to the backend—the text is captured on the frontend in a pretty clean way.

[ocean vortex]

Nearly all models are gonna be confidently incorrect if you ask for specific output length with certain specific number of words etc

[vital mantle]

This what I paid 200$ for lol 😭

[primal orbit]

Opus didn't refuse to discuss personal stuff in direct chat here

probably better to use api than web version

[pseudo hemlock]

bro why are you using opus 4.6 1M for things like this

[echo aurora]

New pfp, the community voted, am now pineapple juice

[pseudo hemlock]

and why are you using claude code 😭

[pseudo hemlock]

why not

i like this one more

[primal orbit]

@echo aurora are there plans to add Muse Spark to direct chat?

[pseudo hemlock]

its not available via api yet, only on meta.ai, facebook, instagram, whatsapp

[echo aurora]

Was very tempted to change the server pfp to that on 4/1

[primal orbit]

ah ok

[echo aurora]

Sorry to say I won't be able to give an early heads up when new models may be added to the platform

[vital mantle]

So I use the opus 4.6 extended on chat?

[pseudo hemlock]

but this is something that could be answered by sonnet, or even haiku

no need for opus LET ALONE opus 1M

[echo aurora]

Yeah we absolutely have plans to expanding these file upload options. Something to consider though is when we expand these abilities, there is a lot more that goes into it than just enabling the ability. With voting data/votes/etc. we have to implement these in a thoughtful way to ensure things run smoothly.

[pseudo hemlock]

pineapple do you know when youre given access to new models before the public

or do you just get an api and are like

"add this to arena"

[vital mantle]

Yes the chat is good but I think gpt or gemin is better not sure

[primal orbit]

Opus 4.6 is much better for personal advice than gpt or gemini.

[brazen thunder]

yes but gemini is cheaper

[primal orbit]

gpt is too sterile and gemini is better but tends to hallucinate

[weak dagger]

I'ma drink you

[pseudo hemlock]

HUH

[weak dagger]

I was jokeing

[ocean vortex]

I still can't get used to this server icon resembling building architecture without AI or tech lol

[vital mantle]

How much is it?

I have gpt and Claude right now

[abstract hinge]

But you could implement something temporary, just on the frontend.

I built a system where it fetches the file and converts it into text when sending it to the chat. Just having this already would be a great option—it wouldn’t require any backend changes or affect the voting system, since it would be a provisional frontend-only implementation.

[ocean vortex]

it's way too general in it's meaning, imho 🤷‍♂️

on that note can do an icon of Earth, this would apply to every single startup/company/project lol

[abstract hinge]

Another useful feature would be a download button for files. Sometimes having to create a .py file, copy, and paste everything is annoying—so having a direct download option would be really helpful.

It could also include the ability to download the full account history, which would be useful for fine-tuning a model.

[spring oar]

How you can validate with other model

[spring oar]

I have 200 plan i dont have this why you trolling ?

[vital mantle]

Im Not

[golden ocean]

claude

[spring oar]

I dont have

[meager harbor]

I mean it could well be at least anonymously

[vital mantle]

Someone said this is the correct one on X

[dusk hill]

Do you know why we can't access Claude 4.6 thinking mode anymore ?

Or the 3.1 pro

It's like the model vanished what happend

[obsidian cargo]

the models died. they got overworked and fried their neurons

[dusk hill]

Sorry are you joking or what's actually going on

[obsidian cargo]

but nah arena can't afford to give them for free so they're only in battle mode for now while they work on giving us daily free credits instead

[dusk hill]

I mean why isn't there a pro version

Id happily pay

[obsidian cargo]

then you might as well get an account at claude.ai

[dusk hill]

Nah way more restricted for what I want it for

[obsidian cargo]

then use openrouter

[dusk hill]

I know it sounds dumb but I'm to low iq for that

Tried couldn't get it to work

[obsidian cargo]

you just buy some credits then start a new chat, no need to use an API

[dusk hill]

Plus I don't want to make it pay per use

Id prefer monthly

[unreal hatch]

Still Anthropic got Mythos and it will mog chatgpt

[echo arch]

Is there anything new? Are they putting the models back?

[unreal hatch]

👀

[echo aurora]

We don't have an update to share regarding this.

[echo arch]

But why was it so easy to remove them and so difficult to put them back? 😢

[obsidian cargo]

thats not image-2, it's been there a while and it only edits images

[rough surge]

gpt 5.4 high is gone

[next trout]

https://klipy.com/gifs/crying-cat-sad-cat

glm 5.1 is actually good

[echo arch]

Don't talk about which models are good, otherwise they will remove those too. 😅 😢

[unreal hatch]

NEW MODEL?

[obsidian cargo]

yeah, but it's not that great. it's definitely not gpt-image-2

[unreal hatch]

dayum fr

[obsidian cargo]

yeah packingtape-alpha etc was way better

[unreal hatch]

that is gpt-image-2

[obsidian cargo]

yeah packingtape and the others were gpt-image-2

flashbrown is not as good as those three were

[unreal hatch]

Let me get gpt-image-2

[obsidian cargo]

mood -_-

[unreal hatch]

🇼 OR 🇱

[obsidian cargo]

gpt-image-1.5 is ass

[unreal hatch]

its not bad sometimes its better then Nano Banana Pro

[obsidian cargo]

its definitely not better than nano banana pro

you're right its not bad sometimes but

[unreal hatch]

ngl i think Google will release new image model after OpenAi release that model

[quartz light]

ppl should mention model speed more often to help identify size cuz it is a good method ngl

but no there will never be mythos

[quartz light]

nah, added 2 days ago

[halcyon valley]

why they do not fix the platform with this CAPTCHA every 2 sekonds

[sterile tartan]

2 isn't tho

[echo dome]

recaptcha hiding the truth

[golden ocean]

https://cdn.discordapp.com/attachments/858401518877540372/1442786831812001832/Shimmy_ahh_shimmy_ahh_shimmy_ahhh.mp4?ex=69d956b2&is=69d80532&hm=7a3bdb7b4f5facb3ec868259b81673d9ed8f28dd9fda1a3a07f4bfd26d1691e0&

[open mountain]

do you write this every time, the meaning of your answer then?

[halcyon valley]

WTF

[echo aurora]

Hmm what do you mean?

[echo dome]

they will screenshot you

because recaptcha is now problematic

i exposed them using sonnet 4.5 search

[halcyon valley]

Is there any way to take away from it

[open mountain]

Every time you write that you "don't know" or "can't answer this question," but in the end, it's the other way around. What's the point of answering the questions?

[light sleet]

U think gpt image 2 would be in direct image chat?

[echo aurora]

If we don't yet have an update to share, then that's what I'm going to tell everyone

[echo dome]

only if pineapple will know what is going on

but he won't know

[polar horizon]

tap in to mimo v2 pro

was NOT familiar with bro's game !

[polar horizon]

someone shot bro mid sentence

[echo dome]

what

[polar horizon]

tap in to mimo v2 pro

[echo dome]

wdym

[polar horizon]

mimo v2 pro is goated son

from what im seeing rn

[echo dome]

idk what is this model

[polar horizon]

brah just search it up on the site

[echo dome]

it's Xiaomi's AI assistant

[polar horizon]

say wallahi

now that im seeing it they do have the same logo

[topaz epoch]

@echo aurora i have question

[native flame]

Hii, it seems Gemini 3.1 pro y gpt 5.4 high are back, but opus is totally gone in arena and in canary 😭😭😭

[rigid pasture]

Back?

[unreal hatch]

@echo aurora HELP

[echo aurora]

I'll look up the Trace ID, give me a moment if you could

[unreal hatch]

Ok

[echo aurora]

They are...?

[echo aurora]

Would you mind asking in #ask-here first and tag me there if you don't get an answer?

[native flame]

Ok I think it was a bug 😭😭😭

[polar horizon]

ts cooking me

[native flame]

Bcuz they appear as available in the list but now they're not, so I guess it's a bug

[echo aurora]

Yeah I'm not seeing them

[topaz epoch]

See @echo aurora there

[lean stirrup]

Anyone from America?

[echo aurora]

Looks like you got the answer from the bot, no?

[echo aurora]

You're being rate limited here it looks like. You can find more information about this limit here: https://help.arena.ai/articles/8931786544-arena-how-to-rate-limit

[unreal hatch]

wait its not i can generate still

i just cant vote in Battle Mode

@echo aurora

ok nvm its fixed

[echo aurora]

It's the 429 Status Code one. This appears as a different error message (Something went wrong/Failed).

[unreal hatch]

yeah its fine the problem got fixed

[vital mantle]

Seriously someone help me

This is not what I paid 200$ for

[next trout]

cancel it

don't let it manipulate you

[vital mantle]

He is trolling me

[next trout]

is it perplexity or what?

[vital mantle]

Claude opus 4.6

[polar horizon]

the way these ai's be lying will never not be funny

[vital mantle]

He is trying to manipulate me

[vast fern]

send me this lmao

[unreal hatch]

dayum he cooked you

[vernal raft]

lmao, eureka just told me it is meta Muse Spark

[obsidian cargo]

?

[vital mantle]

He is

[unreal hatch]

..

[dim oak]

Also gpt 5.3 now? you're killing me

[obsidian cargo]

"Calibrated playful tone for light hearted interaction"

[light sleet]

When will gpt image 2 be available in direct chat?

[golden ocean]

https://cdn.discordapp.com/attachments/969651970757509151/1490437681765945585/45643.gif?ex=69d953c2&is=69d80242&hm=24c39e1b677afe59d22e4afc10187aaa5bfc35384fc279a376c0195999258f86&

[sterile tartan]

Probably when it's released

[obsidian cargo]

Doesn't look like it's releasing today :(

[topaz epoch]

Yes

[pseudo hemlock]

Gemini can use YouTube

[light sleet]

Gpt image 2 tomorrow 100%

[vernal raft]

@echo aurora this is the second time opus cames back alone

just like mythos

break out of the sandbox

aaahhahhahaha

[echo aurora]

Yeah I saw that in the thread, about to respond

[light sleet]

Yo pineapple when is Gpt Image 2 expected to drop in direct chat? What are the chances

[desert pendant]

sup chat

[native flame]

?? Opus is back??

[vernal raft]

no

its a bug

[echo aurora]

Overall I won't be sharing details about potential models that could be landing on the platform/modes.

[vernal raft]

or was, dunno

[viral notch]

gemini 3.1 pro is back? lets go

[desert pendant]

no

[echo aurora]

It's not, but if you hard refresh the site you shouldn't see it.

[desert pendant]

sup pineapple

[echo aurora]

yo

[viral notch]

i see

[vernal raft]

yo @echo aurora when did u guys release muse spark?

i checked like 1h ago

and was not showing

[echo aurora]

Sorry to say isn't a very satisfying answer, but it's the same that I mentioned earlier:

Overall I won't be sharing details about potential models that could be landing on the platform/modes.

[vernal raft]

i think we are past the could be landing

i guess is intended?

[chrome goblet]

Unprivate Gemini 3.1

[vernal raft]

what does "Entertainment, Sports, & Media" category mean?

[echo aurora]

I mainly mean if you don't see us putting out an #announcements it's unlikely we'll provide more information in the text channels.

[vital mantle]

It added

[lapis galleon]

im entering depression because all the models are gone

[spare vigil]

does this new system include text chat as well?

[mossy ocean]

im entering depression because all the

[pseudo hemlock]

Hi

[undone saffron]

Someone know Claude Mythos?

[pseudo hemlock]

I know about it

Why

[faint epoch]

why does muse spark take 30 seconds to say hi and respond to 1+1

this must be the unreleased contemplating model

[pseudo hemlock]

Right now as far as I know it isn’t available via API

So I don’t even know how they’re sending the messages to it

Maybe they’re using a browser and send it and then extracting the response from the website lol

[faint epoch]

unreleased api just for arena 🤫

[pseudo hemlock]

Potentially

Artificial analysis also has a ranking for it

But also says no API is available 🤷‍♂️

[faint epoch]

need to show that theyre competing with the top 3 asap

(maybe 4)

[pseudo hemlock]

Well it’s definitely better than llama 4 😂

[faint epoch]

yea

on style off, its directly competing with gpt 5.4 high

[pseudo hemlock]

It’s been so long since I’ve used gpt lol

And never their top model

[abstract hinge]

The model-based validation could evaluate two things:

1 - Whether the feedback is actually a valid opinion—something that clearly shows a complete and genuine response.
2 - Whether the opinion makes sense in relation to the user’s input or the model’s output.

[jovial palm]

Did anyone notice Gemini Pro 3 series disappeared?

[marsh horizon]

everyone

and opus

and gpt 5.4

[honest verge]

Why gpt 5.3 instant so trash

I sent it a screenshot of my error on the pc I thought it will say how to fix it

But instead it completely didn't understand my screenshot

And said it's broken

Even though I checked it

And it also said I have android instead of windows

[marsh horizon]

try new muse-spark

[whole sundial]

3.1 pro, opus, and gpt-5.4 were removed by arena due to high usage, 3 pro was removed due to the model being killed off by google

gemini 2.5 pro will share the same fate 2 months from now

[honest verge]

It's being deleted soon

[marsh horizon]

fr?

[honest verge]

Even though it's a great and GA model

[honest verge]

I heard in june

[marsh horizon]

why is 3 pro being removed bro

[whole sundial]

same with all other 2.0 and 2.5 models except for nano banana

[honest verge]

They removed 3 pro in march

I don't know why

[whole sundial]

google really likes to kill off models often, probably to manage their highly limited compute

[honest verge]

Then why 2.0 and 2.5 still alive

What was the point of cuting 3 pro

Just remove 2.0 already

[whole sundial]

openai models last a long time, dall-e 2 is still available (that will be killed off next month along with dall-e 3)

[whole sundial]

because a few people still use them due to being really cheap

and non-thinking, at least for 2.0 and 2.5 flash-lite and flash

also fun fact - 2023's gpt 3.5 turbo is still available in the api along with some other ancient models

[honest verge]

Dall e 1 is already removed?

[whole sundial]

yes

[whole sundial]

who is still using this lol

this is the original chatgpt from 2022 btw

it will last for nearly 4 years

sora 2, on the other hand, is being killed off a little over a year after its release

btw for those saying 4o is dead: no it's not, several versions of it are still in the api, the image gen models based on it (gpt-image-1 and gpt-image-1-mini) are still going strong, audio versions of it are also still around.

it will be another year or two before gpt-4o is gone for good

[honest verge]

My grandpa uses it

[wicked talon]

They privatized Mythos to 12 companies.

https://vm.tiktok.com/ZNR4fhLFy/

L on Claude

[mild spade]

new model of Claude?

[undone saffron]

Yes

[hybrid lava]

Hello 👋
I am new to arena and I have came across this strange thing in gemini 3 flash ground .

I am using this model since more than 2 weeks within the same chat. Today when i sent response it generated weird response filled with train emojis and when i refreshed the page it got stucked in generating loop it's been 30 min waiting for it's response. Does Anyone know how solve this kind issue?

[wicked talon]

Not publicly released

[wicked talon]

Is there a stop button where the send message button is?

[undone saffron]

If they removed Opus 4.6 because it's expensive, Mythos is much more expensive, so if by some miracle they do end up adding it to Arena, it will be in extremely limited

[hybrid lava]

Nope

[ocean venture]

https://vt.tiktok.com/ZSHx2W3MX/ what do you guys think?

[obsidian cargo]

Tiktok won't let me watch it without the app

[hybrid lava]

Idk how can I view my archived chats ?

[timber sandal]

Hi guys!

[hybrid lava]

Thanks
Is that a app ?

[undone saffron]

arena.ai

[hybrid lava]

I'm in the web
Oh maybe I was in android i guess ?

I don't see the option

[hollow mulch]

Fix the issue i can't enter chat back, when im pressing a chat is back to new chat and there have nofication say 'Session not found, redercting to home' @echo aurora

[bleak lake]

mythos is genuinely getting out of hands, I've heard in one system it deliberately created a loophole And identified the owner then solved it, just to get praised

[barren burrow]

opus

[undone saffron]

That's because you need to log in again

[velvet furnace]

how about GLM 5.1

Perhaps he can take the place of Opus.

[burnt sinew]

I dont see opus

[short solstice]

Soo im new here nd I wanna generate ai videos bt how do I do That?!!! Nd fyi I ve lil to no knowledge abt ai stuff

[fossil socket]

Any idea why this is happening rn

@echo aurora

[silk dew]

what the

[modern wedge]

FINALLY BRUH ATLEAST I CAN CONTINUE MY PEAK

[silk dew]

its gonna get patched soon

definitely

[unkempt dock]

yup

[silk dew]

might aswell use it while it last

[sullen creek]

yo guys, when is mythos going to come on lmarena

battle mode

[silk dew]

never

[sullen creek]

i think it probably already is in battle mode

[silk dew]

its not going to be release for the public

[sullen creek]

no like when it does, will it be on arena

[silk dew]

idk

the model is probably expensive to run

so most likely no

[sullen creek]

it might be cheaper

because as the models get more advanced they get cheaper

opus 4.1 was 70$

but opus 4.6 is only 25%

25$

it might be cheaper and better

[feral bloom]

Mythos is so good, it can execute code and identify vulnerabilities

[thick blade]

Rip to chatgpt 5.4

[sullen creek]

i dont think its that good too tbh

they might be just exaggerating, the capabilities

[thick blade]

Whats the best rn

[sullen creek]

sonett 4.6

[thick blade]

How

[sullen creek]

wdym how

[feral bloom]

The reason its not released is that its early access is given to companies products that it identified loopholes to patch them up, so as they dont get attacked themselves

[thick blade]

Is it for coding

[sullen creek]

opus is better, but sonnet is also really good

[feral bloom]

Plus from the leaked source code its there, there is support to it but its in a very restrictive testing environment

[thick blade]

What abt max?

[sullen creek]

nah max isnt really good for me

it always redirects me to grok

[feral bloom]

Max uses the best model for the task at hand

[thick blade]

True

[bitter thorn]

lm areana cant make videos anymore ?

[sullen creek]

battle mode

[bitter thorn]

yeah where to get that?

[sullen creek]

click on the generate video icon in battle mode

[bitter thorn]

like how to use it

where brother ?

[sullen creek]

go to arena.ai

and then click on battle mode

u will video icon, and click on

[bitter thorn]

oh so now its directly on website and not on discord, thanks brother

[sullen creek]

yes

[latent steppe]

/create

[calm lagoon]

Thanks for adding Muse Spark ❤️ @echo aurora

[latent steppe]

can i always create video to image plz? i forget the /

image to video sorry

[golden horizon]

lol

[ocean venture]

Umm guys... how this thing happend?

[pearl cargo]

yoo guys did anyone know why they remove claude opus models gemini3.1 and gpt 5,4

[zealous berry]

bro wtf is this

[burnt sinew]

Pricey

[static yew]

Gemini 3.1 Pro was removed here. That was my favorite AI. I'm really sad. I've tried using Gemini 3.1 Pro elsewhere, but it doesn't feel as smart as the Gemini 3.1 Pro here. Even when I use Gemini 3.1 Pro on the official website, it's just not as good as the one here. Does anyone know why? Why is there a difference between the Gemini 3.1 Pro here and the one elsewhere? Where can I use a Gemini 3.1 Pro like the one here? I'm even willing to pay for it.

[lofty frigate]

Am i the o ly one getting a captcha every single time

[zealous berry]

does anyone know what is this

[subtle rose]

ChatGPT Stealer???

[hollow comet]

Maybe they added a system prompt that made the model smarter.

[zealous berry]

but is it official chatgpt or bro installed something from a sketchy website

[static yew]

@hollow comet Really? I'm really curious.

[uncut spoke]

just use it on aistudio

[static yew]

What is AI Studio? Sorry, I don't really understand. Can you tell me more about it in detail?

[empty sky]

Bro what the hell is muse spark

Amusement park?

[golden horizon]

seems to be a new model, and it ranks quite high.

[empty sky]

Still not better than my opus

[analog bone]

It's Meta new model

[stable arch]

H gì ow do I create image tò vídeo

[empty sky]

Go to the arena.ai website and press on the video icon in the textbar

It only works on battle mode

[stable arch]

?

[empty sky]

You can put your image there

[golden horizon]

yep，but don't know opus4.6 what time will be back,keep waiting😭

[empty sky]

Its gone bro

I cant move on

Baby opus

Waiting on announcements

[chrome goblet]

can they unprivate Gemini 3.1?

[empty sky]

Flipping through models

We got sonnet at home

And that outrageous

[velvet furnace]

when opus back? anyone know?

[golden horizon]

metoo, miss opus4.6 so much 🥹

[ocean venture]

Actually i use Brave, i had not install anything Suspicious extension or anything else, but last night i've been through sum kinda suspicious ai website platform...

[static yew]

Where else can I use Gemini 3.1 Pro?

[uncut spoke]

its googles official website but for developers so it offers higher less restrictive usage limits and better model performance and better control over the settings over the ai but dont stress about that js search it up and try it

[north scroll]

make her giving pose to camra hd resolution beautiful realistic

[sterile tartan]

Ai Studio, Woozlit

[jaunty mist]

AI Studio will give you both a better interface and better ratelimits

[short sluice]

Okay they just patched the opus glitch

[primal orbit]

Is it me or muse spark is extremely slow to generate a reply in direct chat?

[short sluice]

same

[slow elk]

how to upload large text in Arena.ai

[strange prawn]

[weak dew]

hello

[vast fern]

tf

[vague quarry]

idk how i did that

[sullen creek]

yo HOW

[vernal raft]

It's a bug

[sullen creek]

yo why cant these bugs happen to me 😭

[vague quarry]

its a visual bug

[sullen creek]

oh so its not actually gemini 3.1 pro?

[ocean vortex]

Meanwhile:

[vague quarry]

idk either, probably gemini 2.5 pro or something

[ocean vortex]

600+ lines, this is actually not bad at all

[frosty lava]

while qwen 3.6 was free for a week to try i used it and there was a bug it couldn't solve and i just went to gpt and legitimately in 3 minute it solved it

qwen 3.6 was literally unable

[vague quarry]

damn i like that button style

[frosty lava]

make me think its only benchmaxxing

[molten cargo]

no its actually gemini 3.1 pro

at least from what i observed

the output was exactly the same as normal gemini 3.1 pro

but yes it's just a bug, it'll probably be patched soon

[dire gulch]

aistudio is free

[molten cargo]

doesnt ai studio have some horrendous rate limits? Last i used it, it hit the rate limit in 5 responses

[dire gulch]

yeah they do now, but they used to be infinite ngl

still, better than nothing especially for free

[molten cargo]

yeah it did, but the rate limits now are horrendous. and on top of that, they introduced a bs content filter which deletes gemini's entire response

[frosty lava]

i think google is cooking a very good ai honestly the next gemini might actually be really great

[dire gulch]

i thought i was schizophrenic bc of this bro i couldnt find so many stuff

[frosty lava]

their gemma 4 is already very good for running locally

[dire gulch]

i feel like gemini models are good but not for agentic use, in my experience

[molten cargo]

its so ass bruh, they love introducing updates to ai studio that everybody dislikes

[dire gulch]

google never gonna embrace transparency and honesty

[molten cargo]

well as long as the knowledge cutoff is better than january 2025 i'll be happy. gemini 2.5 pro and gemini 3 both have a knowledge cutoff of jan 2025, it's crazy at this point

[frosty lava]

i think its hard cause of many ai content in 2026

[molten cargo]

its funny because gemini is the least censored model out of the big ai models, from what i've seen. it's definitely less censored than gpt and claude

[frosty lava]

mostly cause of this i guess

[molten cargo]

yeah but they can't just keep a knowledge cutoff of jan 2025 forever. the models are going to be outdated af then

[frosty lava]

yeah they just need to innovate a really good anti ai filter for their training i guess or something like that if it do not already exist

but imo it already exist probably

[molten cargo]

hopefully it does, but i don't think we're getting a knowledge cutoff upgrade until at least gemini 3.5, whenever we get it

if the next model is gemini 3.2 then the knowledge cutoff is for sure staying jan 2025

[dire gulch]

genuinely hoping deepseek makes a coding plan with their v4 release, their api pricing is so little

[frosty lava]

if you just want front end work or thing like that then glm 5.1 is so good actually and for his price its worth it

but for actually fixing bug and doing hard work nothing beat the frontier

[dire gulch]

yeah mostly backend stuff is what i need

are u using glm's plan by chance?

i was considering it but i havent asked anyone yet how it is for them and the limits

[frosty lava]

I tried it with opencode but i didn't bought the plan

i just wanted to try

what it could do

its very good at front end for sure

[fickle forge]

you can still found Gemini 3.1 in battle mode btw

[vague quarry]

yes

[ocean vortex]

They have some room to go for just with fine-tuning tbh