general

[ocean vortex]

lol what

DId you just make this up for no reason?

GPT5 is significantly smaller than og GPT4 and their infra is much better now. They are doing just fine

They also deprecated gpt4.5 for good so probably doing better than ever on compute tbh

[stray aspen]

api or aistudio

[ocean vortex]

Define 'clearly'. If we look at all the tests and same prompt testing, gpt5 is the clear winner lol

[ripe mountain]

how is the discontinuation of gpt 4.5 related?

[blazing bison]

sama said that on twitter and on ama

what you're talking about?

even on the podcast he said that they gonna need to cut something

[ocean vortex]

gpt4.5 was their biggest model ever. Since that is dicontinued they now have much more compute to spare

[autumn cargo]

I use api with cline + gemini pro (app)

[blazing bison]

???????????

source?

[ocean vortex]

???????????????

[ocean vortex]

For what?

that 4.5 is their biggest model?

[blazing bison]

no

[ocean vortex]

Do you seriously need a source for that?

[autumn cargo]

Subjectively. But even objectively speaking, the win-rate against 2.5 pro is very low on lmarena.

[blazing bison]

that they not gonna cut things bcs they now have compute to spare

lmao

do you know that 4.5 was using training resources right?

And now with 4.5 disabled they gonna use it on gpt-6 related work

not inference

[ocean vortex]

Read again what I wrote lol

And then dispute what I wrote

[blazing bison]

bro is blind

well, just wait 2 days and when he announce sora or other shi* being cut we discuss again

[ocean vortex]

that is nothing new, they are saying the same things all the time

that's how they stay in business... smh

They are always looking for ways to cut costs

[blazing bison]

it's not cut costs bro

[ocean vortex]

and this is one of the means of justifying that

[blazing bison]

they don't have enough compute

lmao

sam is crazying traveling the world to get 500 bi and bro is saying that he has compute but don't want to use to cut costs

are you insane or what

[ocean vortex]

No one does if you read everything literally. You really shouldn't take this for face value

They have plenty of it, the only question is the price/cost

[blazing bison]

lol

[ocean vortex]

smh

[whole wagon]

It has always been the case openAI are deploying compute far too slow. The stargate project is planning to have 64k GPUs by the end of the year

That is slow compared to other companies

[wet sparrow]

Either you don't follow the news, or you're a conspiracy theorist

[whole wagon]

https://www.datacenterdynamics.com/en/news/openai-and-oracle-to-deploy-64000-gb200-gpus-at-stargate-abilene-data-center-by-2026-report/ nevermind end of 2026 is when they expected 64k

[ocean vortex]

They have their priorities on other things - like training new models. They also won't accept the higher cost than what they feel comfortable with - that's the extent of them struggling for compute

[golden ocean]

Bro is gpt 4o -

[whole wagon]

OpenAI and Oracle are expected to deploy 64,000 Nvidia GB200s at the Stargate data center in Abilene, Texas by the end of 2026.

[whole wagon]

This is basically pathetic rate lol

[ocean vortex]

yeah they are expanding all the time and scaling, obviously

[ocean vortex]

They have azure though

If they wanted, their compute is 'unlimited', but cost...

that's a different question

[whole wagon]

That's their main stargate computer. Meant for training

[wet sparrow]

It's not about priorities. It's about not having enough gpus for both

[ocean vortex]

Most of these tweets are meant for marketing. And to justify any potential ratelimits. They are not to be taken for face value

[wet sparrow]

Yes, you're a conspiracy theorist. There's no point in discussing it

[blazing bison]

Just let him yap

[ocean vortex]

I'm literally not. lmfao

I'm just not naive and know how this works

[blazing bison]

yeah bro

obviusly

[ocean vortex]

???

are you actually dumb?

[wet sparrow]

I know just enough about ML to know that he's telling the truth

[ocean vortex]

Or do you seriously think them making their models smaller and cutting models from model switcher + modernizing their infra... put them at bigger strain than before? 🤣

LOL

[wet sparrow]

The strain remains the same; they can simply serve more people without removing resources from research

ChatGPT is growing every week

[twilit valve]

hello

[ocean vortex]

Incorrect, the strain goes down. But obviously you can still sell it as having not enough compute, there's never gonna be enough of it technically...

[wet sparrow]

they use the free power for research

So, it's not going down

[ocean vortex]

They used it for research too since the start

[wet sparrow]

Bro, you can't have GPUs training models and doing inference at the same time

[ocean vortex]

But they also had to host bigger models and deal with older gpus

[ocean vortex]

obviously

[wet sparrow]

Even if they have optimized their infrastructure and now have free GPUs for training models, the strain remains the same. They are going to use them to train models

When they say they don't have enough GPUs, that includes the GPUs they need for research, the GPUs they need for current inference, and the GPUs they will need for future inference if ChatGPT continues to grow at its current pace

[ocean vortex]

Not the same. At the very least the cost of upkeeping chatgpt service is lower.

[ocean vortex]

Yes. Like I said this was the case since the very start. It's all relative.

[ocean vortex]

Predictable result...

[whole wagon]

The plus limit went from 200 to 3000. GPT5 Thinking basically unlimited for plus users, they will never hit the limit

It doesn't add up with the capacity tweet ngl

[blazing bison]

Yes my friend tryed it it's misleading a little

because it's like 2600 mini thinkings and not 3000 gpt-5 thinking

[whole wagon]

Yeah it has the router

[ocean vortex]

Yeah precisely. If they were struggling for compute 3000 would have been literally impossible

[whole wagon]

How much do you think the router helps. What percents are going to mini or even nano

[blazing bison]

Not true because the o3 mini, o4 mini, gpt 4.5 and others was using those resources that they are giving gpt 5 mini thinking now

[ocean vortex]

afaik nothing is going to mini/nano until you hit your usage caps

[blazing bison]

that's why it's possible 3000 per week

the gpt 5 thinking model that is basically new o3 or o4 still 200 per week

[ocean vortex]

GPT5 default option is 99% more costly than to run just gpt4.1 tbh

and gpt5-thinking is o3 equivalent, in terms of cost

[wet sparrow]

Not even my friends who work at OpenAI have that information. You're incredible!

[ocean vortex]

I mean roughly, that's just common sense...

doesn't take a genius, but obviously this is not official confirmed info, if you need for me to spell it out just for you?

[blazing bison]

It just you making assumptions so

without real data there is no point

[ocean vortex]

well then there's no point in this chat...?

lmao

[blazing bison]

When i discuss things here, i bring tweets from researchs, papers, news

always based on some data

[ocean vortex]

Then look at API pricing

that's your data

[blazing bison]

you're just talking about things in your head and treating it as the truth

[whole wagon]

From what I hear. GPT5 Thinking is quite a bit cheaper to serve than o3 lol

[ocean vortex]

#general message

[whole wagon]

I mean look at the speed of it. That should be a clue lol

[wet sparrow]

"From what I hear..."

[ocean vortex]

lmfao

[whole wagon]

Bro what

[wet sparrow]

There is public information about them using specific silicons that explains the O3 cut price and speed

[whole wagon]

How is API pricing data for the cost

That's literally so dumb

[ocean vortex]

It's a good indication of their cost. There were certain rumors suggesting same size as o3 as well. Nothing at all to suggest it is bigger model for now though

[whole wagon]

Do your friends work at openAI as the janitors or smth? They should know that info if they even touched GPT5 training pipeline

It is not that hidden

[wet sparrow]

They work with the front end and ChatGPT

Yes, it's hidden from them. It's hidden from everyone

[ocean vortex]

web development?

[whole wagon]

I know a GTM director and he knows it

It's not hidden

How would they even hide something that fundamental that easily

[wet sparrow]

Well, I'd rather believe my friends than you

[whole wagon]

Your friends work in web dev mate

How tf would they know

[ocean vortex]

yeah they would have no clue tbh

[wet sparrow]

I would rather trust information from insiders than from random people on Discord.

[ocean vortex]

trust what exactly?

[wet sparrow]

I'm leaving now. Bye-bye, conspiracy theorists!

[ocean vortex]

that they "don't know"?

that's not some secret information you can trust or not trust

LOL

[blazing bison]

Funny how everyone here has insiders friends lmao

You guys are funny

[whole wagon]

It is kinda impossible to prove. Because then you literally out them how do you want me to prove it without doing that

[glass gulch]

I feel like I know the answer but llmarena.ru is NOT you guys right?

[whole wagon]

no kek

i just visited the site

[ripe mountain]

wtf

[ocean vortex]

very clearly not lol. Seems like some illegal copycat, probably using stolen keys as well tbh

[steel blaze]

Can someone point me to research on automating the leaderboard with third LLM prompt creation and response evaluation? How close can it get to human results these days?

[keen beacon]

What

[zinc ore]

https://x.com/scaling01/status/1954546677185970271

Bruh

Lisan al Gaib (@scaling01)

Gemini 2.5 Pro has a 67% winrate against GPT-5 Thinking

X · Aug 10, 2025, 2:13 PM

[keen beacon]

So most people just abuse lmarena to get top of the line models for free?

[keen beacon]

Ofc all this data is used to train models but so is your Google search input to feed you with personalised ads right?

[cedar tide]

@echo aurora GPT 5 needs to be retested from the beginning on both arenas with the public API. Many people who had preview access say that GPT 5 was much better on that access than now via the official API.

[tidal ginkgo]

yes

[keen beacon]

dude

https://tenor.com/view/sayaka-madoka-magica-gif-10367661

[brisk helm]

bro lmarena is the sauce

[keen beacon]

wdym?

for the training data?

[brisk helm]

no for ppl who cant afford the top of the line ai models

[ocean vortex]

Just did some back to back testing gpt5-chat vs gpt5-reasoning (API) vs gpt5-router (chatgpt)

[tidal ginkgo]

agree

[keen beacon]

Actually not bad if you don't mind if your data will be used to train their models

[tidal ginkgo]

70% of people found out about it when a major ai came out and wanted to use if 4 free

[ocean vortex]

I would say that default model on chatgpt is halfway in between reasoning and non-reasoning variant on overall performance, but closer to the non-reasoning one

there are obvious gains to be had of having reasoning always on

[keen beacon]

Which one major AI?

[tidal ginkgo]

gpt-5, o4

grok 3

grok 4

[ocean vortex]

And that router does not nearly always work as it is supposed to

[tidal ginkgo]

claude opus 4

most of them

[keen beacon]

Wow

Someone's gambling the vc money away

[stray aspen]

which is the greatest on lmarena: gpt-5 nano,chat, gpt-5, or mini

[tidal ginkgo]

lol

not here man

[tidal ginkgo]

gpt-5

[stray aspen]

go to video arena bro

[tardy crown]

yo
anyone know how to make the viral videos with baby speaking?
when I try, veo3 doesnt let me because it contains minors

[tidal ginkgo]

lol

[stray aspen]

you dont lmao

[tidal ginkgo]

use another ai

[stray aspen]

thats google policy

[tardy crown]

yea but people still bypass it

[tidal ginkgo]

well idk

[tardy crown]

there are tons of videos

[tidal ginkgo]

i´m no jailbreaker

[tardy crown]

on my fyp

[tidal ginkgo]

use another ai

[jade egret]

huh

[echo aurora]

Hey sorry I’ve got cold so haven’t been following chat much. Are we saying that summit is different than what’s available now?

Many people
Where are you seeing that?

Sorry again for not following, I’m a bit out of it atm

[kindred solar]

I may be wrong, but I think the GPT-5 model needs to be handled a little differently than others, which may be why users are dissatisfied

So here, the model is always reasoning, unlike on the OpenAI platform, where it automatically selects the working style depending on the prompt?

so we test the gpt-5 model as on openai api but chatgpt model works diffrently right?

[ocean vortex]

I’m calling bs on that. Try testing gpt5-high against o3-high on some decent question set and you should see for yourself that it performs better. The issue is more of people using the router (default option on chatgpt) and expecting for that to perform like SOTA reasoning model. But this wouldn’t be applicable to API.

And it’s not like the metrics are showing night and day difference o3 to gpt5. So those improvements you CAN see kinda align with them… If it was really performing worse I don’t think you would be able to tell it’s better (than o3)

[brisk helm]

agreed

[low python]

Selling ChatGPT Plus for $10 for 3 months.

[verbal nimbus]

The leaderboard should a slider that only considers votes within the selected time period in order to generate the rankings.

[vapid zinc]

When will the leaderboards be updated?

[echo aurora]

Normally take about a week to get the votes/validate before posting

[stray aspen]

dude gpt-5 is incredible for lua

it keeps on impressing me

gemini couldnt one shot this

this is great

i love gpt-5

lmao

i mean the fact that it works on the first try

this is crazy

[wicked root]

Could you tell us if there’d be a change in the rankings?

[thorn vault]

Hy

[echo aurora]

Yeah I’ll make an announcement when we update

[wicked root]

Thank you 👍

[solid brook]

Yeah gemini benchmark is with unnerfed version of the model. If they benchmark with the nerfed model it would go A LOT down

I swear to god if they do the same with gemini 3 ......

[languid crescent]

Hey guys, I have a question

I've always wanted to customize the chat (the way I prompt). How do I boldened a text (like h1-h6)? Is it possible to do a strikethrough? Other methods like #, ###, **, ``, ?? How can I do that in LMarena? It makes my prompt much cleaner and clearly instructed.\

Is it called "Markdown Syntax"?

[wicked root]

man I'm getting nervous about google vs openai for August

[dire cosmos]

same here :<

[dusky aurora]

seems so

[grand coral]

📢 GLM-4.5 Technical Report is here!
We’re pulling back the curtain on how GLM-4.5 was built to excel at reasoning, coding, and agentic tasks — powered by a unique multi-stage training paradigm.

🔍 Highlights:
• Expert model iteration + self-distillation ➡️ unify reasoning, agentic, and general chat into one model
• Hybrid reasoning mode ➡️ knows when to think deeply, when to respond instantly
• Difficulty-based RL curriculum ➡️ break through performance plateaus
• Efficient function calling ➡️ more reliable tool use for code-heavy tasks

📄 Read the report: https://arxiv.org/abs/2508.06471

💬 Let us know your thoughts in the thread!

[worldly inlet]

S

[cedar tide]

Yes with all the models.

[hallow ridge]

I need help

[cedar tide]

"where": on Twitter, influencers who had early access to GPT 5 They tried the same prompts after the release and the result was very far from the preview.

I'm not sure if he's right or not., but I don't think we can trust any company, we don't know if summit has not been quantized or anything for production

[hallow ridge]

I need to get my crypto back

[delicate rapids]

[harsh flume]

Did openAI also discontinue the DeepResearch toggle button?

[delicate rapids]

🐼

[hallow ridge]

Hey anyone in the dark net or dark web

???

[ocean laurel]

Hello

[harsh flume]

Yea, I tried a Deep Research prompt at chatgpt just to check if it'd maybe understand that it's a deepresearch query and work accordingly, but it doesnt work

Bummer, they could have kept it as a function to use on their new gpt

If anyone here uses their API and can confirm weather or not its still avaliable as a API tool call or something id appreciate

[solid brook]

Why do you think gemini 3 releases in august

[wicked root]

I don't. I'm betting big on Gemini 2.5 retaining its title against GPT5 this month.

[solid brook]

Bruh. Dude gpt 5 high is better than gemini 2.5 pro in every way

[wicked root]

Without style control?

[solid brook]

Uhm

Idk what is stylencontrol?

[wicked root]

less emotes and stuff

you can turn it off on the lmarena website

[solid brook]

Oh

Btw the gemini 2.5 in benchmarks is the unnerfed one

[delicate rapids]

[solid brook]

Not the model we have access to

And

Gemini 2.5 benchmaxed

But gpt 5 benchmarks are real thing

[wicked root]

you mean my 2.5 is nerfed?

[solid brook]

Yes pretty obvious if you used gemini 2.5 on its first month

Go reddit

They talk a lot about it

[wicked root]

So do you think GPT will beat Gemini this month?

[solid brook]

Yes if gemini 3 does not come out

[solid brook]

Uhm so?

This is one benchmark

[cedar tide]

@echo aurora
We want to know where the chatgpt router would arrive in the ranking. And if you add it with open ai, be sure that it is exactly the same version as the one available on chatgpt.
(that of the plus users and not pro )

[wicked root]

but it's the most important benchmark for me lol

[solid brook]

Dude on the model ranking gpt 5 is number 1. Idc if you want to belive the lie okay then

[neon idol]

Is better gpt 5 or gpt 5 pro?

[solid brook]

Gpt pro lol

[wicked root]

it's not a lie. I'm talking about polymarket.

[neon idol]

Thx

[solid brook]

Polymarket expects gemini 3 to release

Dude the benchmarks are out

[wicked root]

Not in august

[solid brook]

Gpt 5 already beat gemini 2.5

[neon idol]

In septeber

[wicked root]

What happens in sept is a non-issue for me.

[solid brook]

Not any point in talking to you. Go to youtube watch a good comparison vid of the 2 models

[cedar tide]

If you interested you can Upvote my request
https://discord.com/channels/1340554757349179412/1404368537690308739

[verbal nimbus]

It didn't even beat gpt-4o here.

[solid brook]

Yeah man you're right.....

[vague bloom]

Yo guys is there any way to connect LMArena With R Studio

[solid brook]

What is r studio

[terse river]

R language related?

i.. guess?

[vague bloom]

Boblox

The R word is blocked for some reson

Idk why

[terse river]

oh

[solid brook]

No it doesn't provide any api

[vague bloom]

Thx

[lime coral]

No

[solid brook]

Ok

[tame granite]

just make point system and see what is happening 😄

[solid brook]

?

[viscid timber]

where do i generate images? in video arena?

[potent snow]

Is it somehow possible to implment midjourny?

[hollow imp]

Are y'all really sure gpt5 is better than o3?

Even if it's better it would be the higher tier version
Us free tier users gpt5 isn't better than lmarena o3

@deep adder answer me gpt5 pfp

[hardy pecan]

literally no one reads how this is scored, no due diligence, just copy paste misinformed charts

its a bit wild

[mortal coyote]

how can i retrieve the seed code of the image i generated using Lmarena

[inner gate]

What’s offline IQ? What does it mean

[tribal aspen]

Y'all agree?

#1403773627010388161

[keen beacon]

Reminds me of Perplexity for some reason or ChatGPT Ui in general

Doesn't mean it's a bad thing though

[tribal aspen]

Yeahhh

[tribal aspen]

But feels much better than LMArena's UI

If we have such a UI, I would happily use noting other than LMArena lol

[keen beacon]

I have a problem with deleted convos coming back for some reason even if I delete them

on LMArena

I need to put it on #1343291835845578853

[tribal aspen]

That can be dangerous sometimes

LoL

[keen beacon]

Even if my convos leaked to others, they would be about general tests and brainstorming

and translation

basic stuff

no sensitive data lol

[verbal nimbus]

It means that the dataset is private

[lusty narwhal]

hi

[inner gate]

Ah yes datasets

[inner gate]

Hey joel 👋

[verbal nimbus]

It's a very well known benchmark. Maybe this graphic is more familiar:

[keen beacon]

Is there a reason for the IQ being so low?

[hardy pecan]

Yes, I'm well aware. Do you know how they grade the results? The methodology?

[keen beacon]

Concrete reason?

[inner gate]

Does Gemini 2.5 pro have a higher version of it?

[verbal nimbus]

I presume they just grade it similarly to how they'd grade a human. These IQ tests are typically pretty straightforward to mark (multiple choice, unscrambling a word, etc.). The test set is private though, that's why it's considered "offline".

[inner gate]

Someone said they purposely dumbed down gpt5

And will

Increase later on

[verbal nimbus]

Sam Altman's excuse was a "router issue" (unrelated tweet). But official API status page didn't report it.

[keen beacon]

I see.

[hardy pecan]

Yes sure, but it's an average of the last 7 sets as stated, as you can see there is only a few results so far. Look at GPT5 , 116 and 70 iq, with an average of 93 hence that data, but there is 5 more data points they'll need to give a statistically meaningful result

[verbal nimbus]

But I do think it's partly caused by unreliable routing. It might have underestimated the difficulty and routed it to a weaker model.

[hardy pecan]

If it hits 116, I can assume it'll hit around there on average once all the testing is done

Have to wait for all the testing to be completed first,

[verbal nimbus]

I don't think the router is very reliable.

[verbal nimbus]

The 115 is on the MENSA (public) test, if that's what you mean by 116.

[hardy pecan]

This is what I mean.

More data is needed to smooth out the variance

But yes the offline is better

I mean mensa.no

[verbal nimbus]

Interesting, it increased for GPT-5 but the Thinking model was still worse on the second test.

[hardy pecan]

Yes I'm confused with the thinking result too , I assume it to be much better

[ocean vortex]

we know for sure that gpt5-high is better than o3-high

[hardy pecan]

I think it'll need to be tested more to smooth out variance

[verbal nimbus]

Actually, look at o3 Pro's score too. Did it get nerfed?

[hardy pecan]

I expect thinking to do far better that regular gpt5

[verbal nimbus]

IQ coincidentally drops by 45% right after GPT-5's release?

[ocean vortex]

that offline iq test is very weird. Thinking version did even worse iirc

[hardy pecan]

Certainly surprising

[ocean vortex]

and why is o3-pro lower than o3...

[keen beacon]

Oooooof

[willow grail]

that is gpt5 low

gpt5 high is the best ai ever

[fleet lintel]

When I use chatgpt app, I get GPT low/mid or high? how do I figure this out?

[willow grail]

wont tell ya.

its my computation. only for me.

and the other higher ups

we need to start gatekeep this more often MY KINGS AND QUEENS

THE COMPUTATION IS OURS ONLY

[fleet lintel]

umm... I have feeling that most of my queries are going to low or mid. I am getting better responses from gemini model 🙁
How do I force high ??

[willow grail]

ok try lmarena

just to see if the reply there is diff.

[gentle breach]

hey

[willow grail]

and?

[ocean vortex]

if you use gpt5-thinking rather than "gpt5" and get routed to thinking, the reasoning effort is gonna be higher

[inner gate]

Apparently they dumbed gpt5 down for a while idk why tho or if it’s evej true

Even

[ocean vortex]

it's like..

gpt5 = gpt5-minimal/low
gpt5-thinking = gpt5-medium

[neon idol]

Gpt 5 pro= gpt 5 high

[ocean vortex]

Pro is parallel requests though

so it would still be more capable even with matched reasoning effort

[neon idol]

So he is the best gpt 5 model?

[ocean vortex]

but yeah it could be high... For Pro sub maybe even normal gpt5-thinking is 'high', unsure

[ocean vortex]

Yes pro is. But it's very expensive

[neon idol]

Bro do you have a prompt for test ai?

[neon idol]

Bro just use it for free and unlimited on Genspark AI

[verbal nimbus]

It said "deeper reasoning" if you click on the thinking model bubble. No idea what mode that is

[solid brook]

No gpt 5 pro reasons almost double the time

[willow grail]

no it does not.

[solid brook]

I mean

.....

[willow grail]

thats fake.

gpt5 is best model if you have done ur research

[solid brook]

Can you stop and actually use the model?

[willow grail]

you are being pranked.

[verbal nimbus]

I have on ChatGPT, it's very odd there

I asked it to summarize the chat but it added in a bunch of random stuff that was never mentioned

[solid brook]

Gpt 5 pro is best if you dont have pro plan use gpt 5 on lm arena and say to it think very hard

Because that is their top model

[ocean vortex]

can we put this to bed now?

[autumn cargo]

Do we know whether the GPT-5 in the leaderboard is medium or high?

I think it should be made clear.

[willow grail]

juhm when i say think hard its worse than if i dont say how to think

[verbal nimbus]

That's closer to it, but that's not the model on ChatGPT Plus

[solid brook]

How so? It has been always better for md

Me

[willow grail]

thats why i prefer using poe.com. u can literally pick gpt5 high there

[verbal nimbus]

The model on ChatGPT is very odd

[solid brook]

We are comparing top model from each company. That is the top model from openai

[ocean vortex]

No this just proves that it's a sh'it test. o3-pro also scores less than o3.

[autumn cargo]

It doesn't matter if it is not gpt-5 high. The "think hard" thing is for the router to enforce gpt-5 high.

[ocean vortex]

gpt5-pro is not even significantly better than gpt5-thinking

[solid brook]

Bro no router in api. Also gpt 5 high is just gpt 5 on reason effort hard

[verbal nimbus]

Why would it be? All the other models are fine

[solid brook]

Wich lmarena gpt 5 is set to high

[verbal nimbus]

Not to mention that it tests each model multiple times, which shows changes over time.

[autumn cargo]

My point exactly. There is no pro, reasoning, etc in API. It's just 'minimal', 'low', 'medium', 'high'

[verbal nimbus]

Is it on the API?

[autumn cargo]

There is no such a thing as gpt5-pro. It's not technically correct.

ChatGPT 5 Pro: gpt-5 high
ChatGPT 5 Reasoning: gpt-5 medium
ChatGPT 5 No-Reasoning: gpt-5 chat
ChatGPT 5 Default -> Router stuff

[solid brook]

It is sad what a world we live in. People are trashing gpt 5 just because it does not satisfy their delulu

[solid brook]

I saw gpt pro reason for 11 minutes. Gpt 5 high 5 min

Sos they are diffrent

So

[ocean vortex]

LOL. That's not correct at all.

gpt5-high is just gpt5 with high reasoning effort

[verbal nimbus]

Ok wth is actually GPT-5 Pro

[ocean vortex]

Pro is prompting the same model several times in parallel

you can also have any reasoning effort with Pro

[willow grail]

wait. source?

parallel compute?

[verbal nimbus]

There's no endpoint for it

[solid brook]

The model in the pro plan

[ocean vortex]

There's no API for it yet I think but it is avail for Pro subs

[verbal nimbus]

Hmm, you mean they benchmarked the model manually?

[solid brook]

Yeah i mean they gad to

Had to

[verbal nimbus]

It took them so long to do that for Grok

[ocean vortex]

Look up grok4-heavy I think it was documented, the concept is the same here. It's nothing new.

[willow grail]

i am biased.

i dont wanna read nazi stuff. not even for 100 euro

[solid brook]

Oh so you saying that gpt 5 pro is just two gpt 5 high?

[ocean vortex]

Also gemini deep think uses similar system

[ocean vortex]

We don't know how many instances exactly, but it's more like 10 of them tbh

[verbal nimbus]

I mean, it's nothing remarkable. That's exactly the same score as o3 without Pro.

[solid brook]

No no way that would be expensive as hell would surpass the 200 doller pro plan price

[autumn cargo]

Based on this pro is high

[ocean vortex]

like cons@10 prompting except here rating of each individual response works differently. So it may choose a unique response even if it was very different from all the others

[jolly kite]

you hello guys, is on LMArena chatgpt 5 thinking? (i searched it but coudnt find it)

[solid brook]

No it is better

Gpt 5 thinking is gpt 5 medium

Lmarena gpt 5 high

[verbal nimbus]

The name has been changed

[jolly kite]

aaaa okey thanks ❤️

[verbal nimbus]

Maybe it was a typo?

It's just High now

[willow grail]

REMINDER!!
POE gives you 1000 to 2000 GPT5 HIGH PROMPTS

[ocean vortex]

This is written like that because you can have "gpt5 Pro (Medium)" or Low. They explicitly selected pro model and then selected high reasoning effort.

[willow grail]

for only 22 EURO

[solid brook]

Guys can you go and test models and see the reason time and not spread false info?

For gpt 5 pro go on youtube

[autumn cargo]

Yes because there is no api at the moment called gpt5-pro

[solid brook]

Gpt 5 high go on lmarena

[autumn cargo]

That's what I'm saying.

[ocean vortex]

if they just wrote gpt5-pro that would be incomplete. You can't run a request through API (they probably have early access...) with pro without selecting specific reasoning effort

[verbal nimbus]

Coding scores are a bit confusing

[solid brook]

Agentic coding matter more i think

[ocean vortex]

livebench category scores are far from reliable

[verbal nimbus]

Yeah

[verbal nimbus]

They're closer to the expected performance in Copilot

[ocean vortex]

@autumn cargo same applies for gpt5-pro

you can't run it without some specific reasoning effort

[verbal nimbus]

Anyone know the reasoning effort of GPT-5 in GitHub Copilot?

[autumn cargo]

I don't think so. Do you have a source on that? I believe they are using gpt-5 api with "high" reasoning effort.

[solid brook]

Huh?

5.1

[verbal nimbus]

Typo

Because it's the successor of 4.1 😆

[solid brook]

Idk they must have wrote it somewhere

[ocean vortex]

I just showed you screenshot of o3-pro. gpt5-high is not pro, just like o3-high isn't pro

[verbal nimbus]

It's cheaper in the API

At least in terms of input tokens, which is more significant I think

[ocean vortex]

You would call this o3-pro (high):

and this is o3-high:

[verbal nimbus]

The score discrepancy for low is even more confusing

Maybe the router just decided to send those class of problems to a dumber model

[autumn cargo]

Yes of course o3-pro and o3-high are two different models. I was aware of that. But based on the (now changed) livebench leaderboard and the fact that no api was officially released as gpt5-pro, I though gpt5-pro was the same gpt5 with high reasoning effort (ie gpt-5 high). Thanks for correcting me.

[ocean vortex]

gpt5-low actually looks like a very decent model to be fair...

[fervent jolt]

Does anyone know when the leader board is going to be updated next?

[ocean vortex]

I noticed this when I was testing chatgpt router

when it routed to reasoning the responses were nearly as good as you can realistically expect from a thinking model

even though most definitely this is low reasoning effort

[bright junco]

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

[verbal nimbus]

I see it as the successor to the GPT-4.1 coding model, but the routing seems a bit unreliable to me. The main model is probably good, but sometimes it just drops the ball.

Like how can there be such a discrepancy in results between the two coding categories. I think the router is messing things up

[ocean vortex]

gpt5-minimal is underperforming though. I think that's the main reason they don't give you choice for no reasoning at all

the gap minimal to low is insane

and they can't afford to lose out to gpt4.1 lol

[verbal nimbus]

I think it's affecting the routing strategy

I don't think the gap is just because of "less thinking"

[keen beacon]

gpt 5 base is just bad

[ocean vortex]

it is not less, it is quite literally no thinking at all with "minimal"

it outputs 2 times less than gpt4.1

[verbal nimbus]

Yikes

Very weird

[keen beacon]

horizon alpha version of gpt 5 base (juice 0) has a worse gpqa diamond score compared to gpt 4.1 nano

[verbal nimbus]

I hope that's not the version in Copilot. It'll be a downgrade.

[harsh flume]

I havent played the arena in a week. Any cool anonym model right now?

[keen beacon]

horizon beta (juice 5, likely an early version of minimal) did worse than gpt 4.1 mini on gpqa diamond

[verbal nimbus]

Or did gpt-4.1-mini have data contamination

[keen beacon]

openai didnt mention the horizon models in their announcement because everyone thought they were nano or mini models 😭

[ocean vortex]

I don't think it's bad. It's actually impressive the gains they were able to make with spatial reasoning. It's just that it's hard to make a hybrid model which would be SOTA both when maxed out and with reasoning off.

gpt5-minimal score is low because it's too concise when it doesn't get to use reasoning tokens

[keen beacon]

they svg maxxed this model among other things. cpt after cpt 🤷 the brain damage was apparent with the horizon checkpoints

simpleqa was 33%...

[verbal nimbus]

Spatial reasoning?

[keen beacon]

but i agree about the hybrid thinking thing lol

[ocean vortex]

I mean if that was true gpt5-high would suck. But it beats o3-high convincingly

[keen beacon]

no, i agree its probably the hybrid thinking thing causing the poor performance in general for the minimal/not thinking variants

but they definitely focused on svg specifically among other things

because horizon models had poor benchmarks whilst people liked the svg from those

the svg thing is a tangent of mine sorry 🤣

[tribal aspen]

what is the limit of gpt 5 on lmarena direct chat?

[ocean vortex]

but if you look at webdev arena...

[tribal aspen]

someone please answer

[ocean vortex]

gpt5 is now SOTA there as well

o3 used to do poorly

[tribal aspen]

???

[tribal aspen]

but is very fast ruins thinking

1415926535 8979323846 2643383279 5028841971 6939937510
5820974944 5923078164 0628620899 8628034825 3421170679
8214808651 3282306647 0938446095 5058223172 5359408128
4811174502 8410270193 8521105559 6446229489 5493038196
4428810975 6659334461 2847564823 3786783165 2712019091
4564856692 3460348610 4543266482 1339360726 0249141273
7245870066 0631558817 4881520920 9628292540 9171536436
7892590360 0113305305 4882046652 1384146951 9415116094
3305727036 5759591953 0921861173 8193261179 3105118548
0744623799 6274956735 1885752724 8912279381 8301194912
9833673362 4406566430 8602139494 6395224737 1907021798
6094370277 0539217176 2931767523 8467481846 7669405132
0005681271 4526356082 7785771342 7577896091 7363717872
1468440901 2249534301 4654958537 1050792279 6892589235
4201995611 2129021960 8640344181 5981362977 4771309960
5187072113 4999999837 2978049951 0597317328 1609631859
5024459455 3469083026 4252230825 3344685035 2619311881
7101000313 7838752886 5875332083 8142061717 7669147303
5982534904 2875546873 1159562863 8823537875 9375195778
1857780532 1712268066 1300192787 6611195909 2164201989

[keen beacon]

they focused on 'big model' things with the cpt/etc it seemed, svg, web dev, etc. just found it funny they focused on svg. the benefits are definitely real though in those areas. at least with the horizon checkpoints, those models were fried except in those regards

my points are about different things and i clustered them together for some reason and it's confusing sorry 🤣

[verbal nimbus]

The minimal reasoning version is #1 on Design Arena too.

[autumn cargo]

But again there is a discussion here https://community.openai.com/t/the-least-important-question-right-now-why-is-gpt-5-pro-not-available-in-api-at-exuberant-pricing/1339471/2 that hints that gpt-5 pro is just a maxed out gpt-5. So really unless OpenAI introduce a new model named gpt-5 pro, I'm inclined to think that gpt-5 pro doesn't exist! List of models here: https://platform.openai.com/docs/models

[verbal nimbus]

Gemini and GLM 4.5 seems benchmaxxed for React + Tailwind

On Design Arena they drop to #9 and #10, since they can't use React

[hollow imp]

Give me pro or high give me somethingggggg 😭

[unborn sleet]

can someone tell me why chat gpt doesnt host previews of their code?

[hollow imp]

Whats design arena

[verbal nimbus]

designarena.ai

It's like Web Dev Arena

[unborn sleet]

is it as good as aistudio builder

[hollow imp]

Is it part of lmarena?

[unborn sleet]

yes

[ocean vortex]

API is just not publicly released yet.
https://cdn.openai.com/pdf/8124a3ce-ab78-4f06-96eb-49ea29ffb52f/gpt5-system-card-aug7.pdf

[brittle tiger]

https://x.com/patloeber/status/1954887103105380361?t=5iJUGwzPogdcwyHCdAgGIw&s=19

Patrick Loeber (@patloeber)

X · Aug 11, 2025, 12:46 PM

Native image gen finally coming to better model than 2.0 flash. Crazy how long that one has been out without improvement

Logan also showed it editing palmer luckeys tweet

[willow grail]

https://x.com/i/status/1954577085520806067

Erik Kirkegaard (@erik_theviking_)

@apralky How can i short every one of his positions?

X · Aug 10, 2025, 4:14 PM

[solid brook]

What limit?

There is no limit on chatting with the model i think

[willow grail]

box before eating: 340g

pls weigh again after eating. thank you

[neon idol]

What is the name?

And how can try it

[solid brook]

Lol he is teasing gemini 3 like sama did with gpt 5

The death star

[whole sundial]

https://fixupx.com/Zai_org/status/1954898011181789431

Z.ai (@Zai_org)

Introducing GLM-4.5V: a breakthrough in open-source visual reasoning
︀︀
︀︀GLM-4.5V delivers state-of-the-art performance among open-source models in its size class, dominating across 41 benchmarks.
︀︀
︀︀Built on the GLM-4.5-Air base model, GLM-4.5V inherits proven techniques from GLM-4.1V-Thinking while achieving effective scaling through a powerful 106B-parameter MoE architecture.
︀︀
︀︀Hugging Face: huggingface.co/zai-org/GLM-4.5V
︀︀GitHub: github.com/zai-org/GLM-V
︀︀Z.ai API: docs.z.ai/guides/vlm/glm-4.5v
︀︀Try it now: chat.z.ai

**💬 1 🔁 3 ❤️ 19 👁️ 957 **

FixupX · Aug 11, 2025, 1:29 PM

[stray aspen]

why is gpt-5 api so forgetful

i cant keep track of older messages in a conversation

[tribal aspen]

Ohh

[inner gate]

What’s ur thoights on deep seek r1

Thoughts

[eternal niche]

btw guys gpt5 sucks

gemini 2.5 pro even better

[solid brook]

Guys with this i think gemini 3 will drop this week

[ornate ether]

deepseek is pretty nice but it explodes whenver i ask it about chinese political or geographical scenario

[stray aspen]

great when it released

now its decent

but the reasoning time is just nasty

[hearty pulsar]

AI going great guys

10 million input tokens to do nothing but waste compute

[stray aspen]

what

[keen beacon]

New imagen model will come out today

https://www.reddit.com/r/Bard/s/RRBTl8UlBk

[whole sundial]

at least something to make up for the lack of glm image gen

I will try it as soon as it comes out

[stray aspen]

why more image model

we want a llm

[keen beacon]

Perhaps scheduled for later

They'll give a taste of what's upcoming

https://www.reddit.com/r/Bard/s/J9hac6b0wj

[stray aspen]

holy

its the new image model

i hope it hass image editing

gemini 2.0 image editing was great

[keen beacon]

https://www.reddit.com/r/Bard/s/lR1Uv0MVdO

Quite sure it will. This looks like the gpt 5 teaser image with image editing

[stray aspen]

yeah

[keen beacon]

I already made the request on #1372229840131985540

Just in case

Lol

[barren prairie]

Today ? 😍

[keen beacon]

Yes

[stray aspen]

yo wassup

i saw you in the z.ai server

[inner gate]

I feel like Gemini 2.5 pro has boosted in intelligence these past few days idk if it’s just me

[stray aspen]

its been the opposite for me lol

[inner gate]

I must be lucky 😭

[barren prairie]

I know you are there , just clicking on your name and I saw LM arena and Z .ai

[inner gate]

Try again I feel like they’ve made some sort of update

[willow grail]

which ai product has the best lip sync?

[thorny cove]

i keep getting "Something went wrong with this response, please try again." with GPT 5 Chat

[echo aurora]

Okay let me look into.

I assume it's just that model you're running into issues with?

[thorny cove]

it also happened with main gpt 5

but the thing is the gpt 5 chat error is only for 1 chat

could it be a rate limit of some sort

[stray aspen]

yeah its kinda broken right now

i dont know if its because im giving extremely long prompts

[echo aurora]

I'm not seeing errors with either of these models, so it may be a rate limit.

[echo aurora]

You're seeing the same? Has this been an ongoing issue or something you've noticed recently?

[stray aspen]

im pretty sure im hitting a limit or something

it generates stuff and then it shows that error

[echo aurora]

although I am now noticing for gpt-5 the responses are coming in pretty slow and lag.

were both of you seeing the same or was it just the error message?

[stray aspen]

it must be a limit lol

its like 1500 lines of code

this is new

on the leaderboard

it used to be gpt-5

[keen beacon]

daaamn

[stray aspen]

yeah but its just on the leaderboard

i guess gpt-5 is using high effort on arena too

[keen beacon]

Yeah, that's probably the case

[thorny cove]

are rate limits only for single chats?

[echo aurora]

Nope

[solemn plank]

HOW TO CREATE VEDIOS ANYONE?

[cedar tide]

Bad webdeb

[primal orbit]

I have an error with opus 4.1 thinking when it repeates the same response it gave to the previous message. And if I try to put a new message, it gives you have to wait for 50 minutes. But the clock doesn't go down. It's 50 minutes each time.

[cedar tide]

2.5 flash lite still hasn't come back

[rapid merlin]

gpt-5 is really hit or miss with styling, either it comes with something actually good or something like this

[thorny cove]

strange

[cedar tide]

And very bad overall

[primal orbit]

it was released to make gpt 5 look good.

and for cheap marketing points

[echo aurora]

I'll flag again, thank you.

[bright junco]

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

[thorny cove]

two different chats at the same time btw

whats the diff between gpt 5 chat and gpt 5 high

[pure falcon]

How and Why did GPT-5 lose two votes?

[cedar tide]

Just good in math

[pure falcon]

https://x.com/lmarena_ai/status/1953504964275470442?s=46

lmarena.ai (@lmarena_ai)

GPT-5 soars in the Arena with the highest score to date.

X · Aug 7, 2025, 5:14 PM

3183 votes

But now the leaderboard says 3181 votes

[pure falcon]

@echo aurora Any ideas?

Shouldn’t we have a lot more votes since 5 whole days passed? Why did it LOSE two when all the others gained?

[keen beacon]

oof

[echo aurora]

No it's not clear to me, but will flag to the team.

[pure falcon]

Check this tweet. Google had 26.7K, now 28K

https://x.com/conorcraig_/status/1954409951226466412?s=46

Conor (@ConorCraig_)

Numbers > Viral Tweets
https://t.co/mOyqaoffLL

X · Aug 10, 2025, 5:10 AM

Gemini added 1.3K votes. And GPT-5 loses 2? Lol something is very wrong about that

[vernal meadow]

Wow Opus 4.1 improved more than I thought on the none Agentic coding task.

Impressive. Should retest it more. Will Opus 4.1 thinking be on #1? 😮

[exotic nebula]

I dont think so. Gemini 2.5 still beats it.

But it is very impressive to be honest.

I mean, the votes are too low to decide. Lets just wait around for a bit.

[thorny cove]

any way to fix "Something went wrong with this response, please try again." for one singular chat?

[exotic nebula]

Many experience it, all you can do is refresh chat or open a new one.

[thorny cove]

i remember after enough refreshes it did generating for like 10 minutes then went back

[thorny cove]

damn alright

[exotic nebula]

True. Have to agree with you there.

Btw, if you dont mind me asking, which model do you like the best?

[wicked root]

why was gemini updated but not gpt?

yeah seems like so

confidence interval points that way

?

you mean this isn't reliable?

bugged in what way?

sorry, I'm new to this

ah

[echo aurora]

Following up on this: we are looking into and will provide an update when we can.

[pure falcon]

Thank you!

[stray aspen]

pineapple do you know anastasios

[wheat onyx]

https://x.com/patloeber/status/1954887103105380361?s=19

Patrick Loeber (@patloeber)

X · Aug 11, 2025, 12:46 PM

[zinc ore]

Vxtwitter where x is and it'll show the vid

[floral comet]

May i ask, how many is the token count of gpt 5 on (LM arena)

[stray aspen]

idk but it stops generating for me after 2000 lines of code

[echo aurora]

The vote is based on pre-release GPT-5 testing. After GPT-5's public launch, we created a new model entry that points to its public endpoint and collecting more votes. These additional votes we've been collecting aren't yet added to the current leaderboard. We will be merging the votes in the next leaderboard release. cc @deep adder

[willow grail]

gemini was never good at anything

[floral comet]

Oh okay thanks

It's been 15 mins, I requested a code on gpt-5 and yes I refreshed the website still same.. Hope it didn't bugged or the code is just too long🤣

[echo aurora]

I believe yes, but will double check and update if that's not the case.

[slow grotto]

imagen 4.0 what are you doing man

[zinc ore]

Nooooo

[stray aspen]

lol disappointing

[echo aurora]

yes, confirmed.

[red tangle]

legit?

so many grifters these days that it's hard to tell what's real

someone claims they found it in "source code"

[willow grail]

who posted it

[neon idol]

GEMINI 3.0?

[nimble trail]

This is 100% cap tbh

[zinc ore]

Definitely made up

[nimble trail]

Why tf would they put this in source code anyway 😭😭

[mossy drum]

What was the prompt?

[small haven]

is gemini 3 coming in august

[wicked root]

Nein

[small haven]

what are they hyping about then?

imagen 5?

lol

[keen beacon]

image editing version probably

of the 2.0 flash image preview

Donno exactly

[small haven]

makes sense

[keen beacon]

I'm always excited for image model stuff because voting is easier on visual stuff rather than text

[small haven]

ah there we go

imagen 5

[sacred quail]

imagen series are fine but really bad at understanding prompt

Needs native model like gpt 1 image

[sacred quail]

Wrong

it was great at writing,long context, analyzing videos

Also it was fine with reasoning and coding

[velvet patrol]

Why did the amount of votes for gpt-5(-high) change from 3182 to 3181 after the update and the votes for 2.5pro increased by 2k?

is gpt-5 already gone from the arena?

[whole wagon]

https://x.com/OpenAI/status/1954969035713687975

OpenAI (@OpenAI)

We’ve scored highly enough to achieve gold at this year’s IOI online competition with a reasoning system — placing #6 when ranked with humans and #1 when ranked with other AIs.

In just a few weeks:
• 2nd at AtCoder
• Gold medal-level at IMO
• Gold medal-level at IOI

X · Aug 11, 2025, 6:11 PM

[velvet patrol]

hello viren 😛

[whole wagon]

Hi

[ornate agate]

What is eb45-turbo?

[hollow pebble]

look at my video Today!

[willow grail]

ok

[hollow pebble]

i only have one vote left! So vote for my video please!

[willow grail]

are you confused?

what video?

[eternal niche]

btw guys gpt5 sucks

[willow grail]

gpt5 is the only one who can do video games

[hollow pebble]

Here's a hint! A image to video prompt is in the video-arena-1: A Female woman in her 1950's cartoon was smiling, giggling & talking someone with beauty. Models: Veo-3-audio-fast vs Hailuo 02 Pro.

[eternal niche]

i can

[hollow pebble]

look at video-arena-1 and vote for my video!

[sacred quail]

Gpt 5 high is sota right now. Not sure if it can beat Opus 4.1 for coding but its pretty good

Claude must be doing something magical on codes because even benchmarks not looks great, everyone still using them for coding soo

[stray aspen]

inspect elemnt lmao

[stray aspen]

imagen gempix

[hollow pebble]

hey guys! Look at my video on video-arena-1.

[sacred quail]

Not in long context

[dim pine]

Gpt5 😂

[willow grail]

gpt 5 very high before gta 6 last day 2025

[jade egret]

nah

[wheat onyx]

https://x.com/CyberBoyAyush/status/1954985969255628978?s=19

Ayush Sharma (@CyberBoyAyush)

Hehe got you Gemini.

Something is cooking.

Gemini 3.0 Pro🫣

X · Aug 11, 2025, 7:19 PM

Pretty sure that was for flash

[zinc ore]

Fake

Or hallucinated AI PR I should say

If that's the cli one from a month back

[stray aspen]

gemini has a github emoji on lmarena

[echo aurora]

Where are you seeing this? Can't say I'm seeing the same.

[harsh flume]

I wonder how much of that is from some sunk cost thing

like once you get used to your current stack and workflow youd need a much higher threshold of improvement than mildly better to just send it all to air and get used to another tool, even if in practical terms it'd be just a slight annoyance to do

[stray aspen]

its on lmarena direct mode

it just put that idk why

[neon idol]

@echo aurora Sorry for this stupid question but in your opinion when Gemini 3.0 will be released?

[lime coral]

In 3 sec

[keen beacon]

Tomorrow

[hardy pecan]

GUIZ GPT 5 IS BAD, BRING BACK 4o!!!!

[stray aspen]

so o3 is smarter than o3 pro

thats crazy

[hardy pecan]

It's a flawed benchmark, 1 data point for o3 pro underperformed alot which brought down its average

[echo aurora]

Soon (I have no idea)

[candid storm]

How do you know?

The data is not published yet right?

35% chance is not the same as 'gpt5 will beat gemini in remove style control'

[neon idol]

I am the only that don't care about the message taken by lm arena for testing ai?

[misty vault]

no

I even put my home address and credit card numbers in it

It is probably on that huggingface link

[gentle plinth]

5-pro

Normal 5 scores worse then others

[hardy pecan]

its very scuffed, but no one actaully reads how the data is collected and presented, its bizzare. ME SEE CHART ME BELIEVE

[wicked root]

what's happening?

[keen beacon]

I dont want that sycophantic thing back at all

[candid storm]

Ok tnx, you seem like you know your stuff. I've redistributed my portfolio to account for the uncertainty

[sacred quail]

this is NOT long context

i'd say after 500k token long context starts

Gemini has no competitor

Only Minimax M1 tried to be close

Also yes, long context is important

Espicially when you try learn something, if you are a student, if you need summarize or analyze of some long text, you have no option besides gemini

[wicked root]

how so? I'm hearing mixed opinions across the board on gpt5, it seems highly polarized.

[sacred quail]

Could be true

As a gemini fan im just admitting gpt 5 is best model right now

Better than O3

Also i do showing big respect towards to their "less praise" model choice

After 06/05 goldmane update gemini turned to 4o like praising you for everything

im really not liking this

Yeah

I understand people had very high expectations but

There is so many unnecesary critizing towards to gpt 5

Also im not sure but probably gpt 5 is very efficent model too which is also important

Better than O3 but also cheaper than O3

Oh really

interesting. But they lowered O3 price soo maybe we should compare with first api price

If i not remember wrong gemini 2.0 flash think was before than deepseek R1 but weirdly people didnt care

it was quite good

2.0 was trash but

No

Listen

2.0 was a trash model but 2.0 flash think was 3x 4x better than 2.0 flash so they did really good on that reasoning thing

Even if base model is trash

[eternal niche]

btw

[ocean vortex]

Exactly. It's a sh'it benchmark

[wicked root]

@deep adder market’s pricing in your hypothesis

What? Ai market is a lot more liquid than the one I trade in

¯_(ツ)_/¯

[wicked root]

you seem very confident in gpt, can I ask why? I'm new to this so I don't know a lot.

[devout vault]

yes

[devout vault]

He's trolling

[heavy knoll]

Can someone Tell me Which one is Chat gpt 5 high

[wicked root]

@eternal nicheman your pfp creeps me out lmao

I just realized... I'm SO sorry if that's your selfie

[wicked root]

love the eccentricity, but you've now intensified my fear

this is the weirdest thing I've seen in my life and i mean this in the best way imaginable LOL. What's he saying?

[eternal niche]

sings 'Matushka Zemlya'

[wicked root]

https://www.youtube.com/watch?v=8mtxEbvzkHs

?

[eternal niche]

yes

[sacred quail]

"gpt 5" is high effort reasoning model

"gpt 5 chat" is not

[ocean vortex]

gpt5-chat is 4o-latest successor, no reasoning

Bluntly speaking it's probably the same as gpt5-minimal

just like 4o-latest kinda sorta was the same as gpt4.1

[sacred quail]

@ocean vortex do you what is gpt 5 thinking's base model ?

Gpt 5 chat ?

Or gpt 4.5 ?

[marsh stratus]

GPT 4.5 is too slow and expensive

[ocean vortex]

GPT5-minimal. It’s a hybrid reasoning model and there’s no routing on API. Routing only happens on chatgpt website

So like, that same model is also technically the base model

[torn mantle]

Best model is gemini

[ocean vortex]

They probably would have used gpt5-chat with no routing if it performed better… lol

But now by routing it occasionally to gpt5-low, it can comfortably beat gpt4.1

[rugged brook]

CHAT IS WIOTHOUT THINKING

[barren ermine]

what model is best for development and code?

[rapid merlin]

I'm confused about the division between gpt-5-chat and gpt-5 minimal, low and medium, can somebody enlighten me

[ocean vortex]

They all are derived from the same base model. Chat is like minimal except different fine-tuning and probably less RL training

[junior sonnet]

can i choose the model in the vid gen?

[verbal nimbus]

Different, AIStudio allows you to edit and save projects, and it has a React environment.

I think Design Arena uses pure HTML/JS instead for websites

[brisk helm]

no

[misty vault]

Large Language Model

[rapid merlin]

i see, and what about the gpt-5-thinking in chatgpt? Is that just gpt-5 chat with more think time?, something like gpt-5-chat-high if we will?

[wicked root]

someone here said grok4 is the best at coding

Is this for ALL coding?

@deep adder opinion on grok4?

[wintry tinsel]

Soooo my bros

When is Gemini 3

[cloud zinc]

october

[wintry tinsel]

Account was made today

Ban this freak

[wicked root]

what's hardcoding?

[mellow frigate]

What do you mean? Generating music?

[inner gate]

Is that bannable

!!?!?!

[misty vault]

Yes, you have a chance to be banned when your account is made today

[inner gate]

Oh wow how come

[misty vault]

because of @hollow ivy starting a gemini 2.5 pro gooning cult

They will eat you alive if you say anything negative about gemini 2.5 pro

[timber tulip]

hallo

[sick spire]

How to Delete a Generation
lf you'd like to delete the initial prompt and generation from the bot, right-click the bot'smessage and select Apps > Delete Generation . Note that deleting the originalprompt will also delete its corresponding generation, but deleting just the generationwill leave the original prompt intact

[golden ocean]

ts so wholesome

[sand bay]

you know that this site is a bit deceptive

[hardy lion]

GPT-5 via API without a system prompt telling it that is it GPT-5 is a bit deceptive

[stray aspen]

It's gpt 5 anyways

So who cares

[sharp yew]

Is it LMArena?

[echo aurora]

Yes, & it has been shared with the team.

[hardy lion]

The GPT-5 model without a system prompt does not know that it is GPT-5. This can be reproduced on OpenAI's API playground

[slim mesa]

WTF THE GPT 5 HIGH ARE ON THE CHAT?

[agile bloom]

woah way too many gpt5 which one is for what?

[solid brook]

Well the best is gpt 5 high

[agile bloom]

thanks my guy

[tidal ginkgo]

hey uhhhh

i was off of gpt-5 for a bit

why is there a gpt-5-high?

i thought gpt-5 was high already?

oh that sounds bad

[rare python]

they made it clearer

because new people can be confused

[echo aurora]

It's the same model, but added the high to make it more clear

[tidal ginkgo]

oh ok ty

lol

[patent bane]

hmm

[gusty helm]

Hey! Not sure if this was asked but i see a leaderboard update in text arena yesterday, but did the gpt5 votes remain unchanged? Score think changed a bit, but not votes? Some bug?

[echo aurora]

Hello - we did chat about this earlier but will share the response

The vote is based on pre-release GPT-5 testing. After GPT-5's public launch, we created a new model entry that points to its public endpoint and collecting more votes. These additional votes we've been collecting aren't yet added to the current leaderboard. We will be merging the votes in the next leaderboard release.

[gusty helm]

Thank you, that makes sense; so I guess the score only updated then

[lime oak]

guys what differint with vedio arena 1 and 2 and 3 and 4

[echo aurora]

No difference, but we want to spread out generations a bit or else the channel will get too spammy.

[lime oak]

ok nice idea

[quiet dust]

Hi guys, is the gpt-5-high on LMArena the same as the regular basic gpt-5 model in ChatGPT?

[hardy lion]

not quite, gpt-5-chat should be the closest to the experience in chatgpt

[quiet dust]

So, gpt-5-high approximately at the same level as GPT-5 Thinking (medium mode)?

[hardy lion]

gpt-5-high should be gpt-5-thinking (high mode)

[quiet dust]

Really? Thanks for the answer

[hardy lion]

I'm taking the word of this openai employee on twitter:
https://x.com/ericmitchellai/status/1954680194733863200

Eric (@ericmitchellai)

@benhylak OH. Guh… I did not know we used different nomenclature in API and chatgpt

What’s GPT-5 in api is GPT-5 thinking in chatgpt

What’s GPT-5 in chatgpt is GPT-5-chat in API

🫣

X · Aug 10, 2025, 11:04 PM

[elder lintel]

hey

[hardy lion]

sup

[keen beacon]

can someone help me, in lmarena its always stuck generating and idk what to do anymore, refreshing the page doesnt work for me

[wicked root]

How many of you think Gpt5 will beat Gemini with style control OFF this month?

CI band is insanely wide on gpt5

[dawn grove]

How can u get Lmarena unlimited free is there a method?

I*

[drifting thorn]

It's unlimited free because LMArena is paying for u

[dawn grove]

But it has limitation i mean this likenit has a amount of messages

[acoustic cliff]

Of course

[keen beacon]

Enjoy your data on Hugging Face later

[ornate agate]

A fair trade.

[drifting thorn]

A fair trade.

[keen beacon]

No imagen model yet...

I thought it would get released

Sigh.

[keen beacon]

No like, really, there must be A TON of data that nobody was intended to see in these datasets right now

Because people don't read and think they're gaming the system by getting gpt-5 for free

[sacred quail]

Well, they can read my conspiracy theory tests which is im testing them on every new model

Only problem is if they starts to believe

[ocean vortex]

wait... how come you don't have "reasoning effort" for this model?

oh you probably haven't verified your org since you don't have summaries either. But it's weird they aren't letting you adjust reasoning. You may be stuck on "minimal" lol

[sacred quail]

Which site is this

[ocean vortex]

Seems like he is confused himself lmaoo

This is roughly accurate, except when chatgpt decides to route your request to reasoning when you are using "GPT5". Then it is no longer gpt5-chat.

[ocean vortex]

https://platform.openai.com/chat

[hardy lion]

the model on the leaderboard is named gpt-5-high, it has reasoning effort high

[ocean vortex]

Yes but I'm referring to your screenshot. It's odd that you don't have that option but have access to gpt5.

[hardy lion]

oh, I'm not sure I didn't actually take that screenshot, someone else reprod and sent to me

The api calls are using reasoning_effort="high"

[gentle plinth]

💀
https://x.com/sama/status/1954705761805938734

Sam Altman (@sama)

we are considering giving a (very) small number of GPT-5 pro queries each month to plus subscribers so they can try it out! i like it too.

but yeah if you wanna pay us $1k a month for 2x the input tokens feels like we should find a way to make that happen...

X · Aug 11, 2025, 12:45 AM

[gentle plinth]

@hardy lion its a bug, you have to reload the page for the option to appear

[tulip cipher]

what i need to do now..

[torn mantle]

lmao

[sage heath]

Hello

[hardy lion]

thanks!

[ocean vortex]

Oh for sure they are gonna find a way to let people set their money on fire. Let them have what they want!

[tulip cipher]

?

[sleek crow]

you can reset cookies only for lmarena on your browser to reset the limit

[tulip cipher]

how

[sleek crow]

what browser do you use ?

[tulip cipher]

btw, what's the difference between beta.lmarena.ai and lmarena.ai

[tulip cipher]

firefox

[sleek crow]

Just find on internet how to reset cookies for a website for fiefox beacuse right now im on microsoft edge

[tulip cipher]

?

[sleek crow]

i have no clue it think its just the same

[spare rune]

whats the rate limit on gpt5 high?

[indigo hazel]

i think right now they're the same. but beta would be used in case to try something, new features or to see if it all works

[eternal niche]

btw guys gpt5 sucks

[ocean vortex]

you seem obsessed with it

[eternal niche]

gemini 2.5 pro the best

[leaden sun]

sam is being practical here, am sure there are enough people from a niche circle paying that

[keen beacon]

I like kimi k2

[eternal niche]

even kimi k2 better than gpt5

[gentle plinth]

if you pay that for 2 years you could probably get a used h200

its not worth it

[leaden sun]

gpt 6 coming out next soon, so this is just intermittent luxury foreplay for privileged few

[gentle plinth]

its not that expensive to do inference at scale to justify such costs, even for larger models

[leaden sun]

not for certain niche circle of people who dont understand the details i guess

[ocean vortex]

I think they already played that card with $200 sub tbh. More than that it's becoming insanity and then where do you stop...

Not to mention that with 1k per month you kinda already could comfortably rent hw to host any model all to yourself

[unborn lantern]

Any proven data?

[unborn lantern]

Lol

Lmarena just change the name of gpt 5 to gpt 5 high or they change the api?

[eternal niche]

yes

[ocean vortex]

just the name, it was high as a kite since the start on lmarena

[tulip cipher]

yes

nice nick btw

[eternal niche]

Что вершит судьбу человечества в этом мире? Некое незримое существо или закон, подобно Длани Господней парящей над миром? По крайне мере истинно то, что человек не властен даже над своей волей.

[ocean vortex]

Co ty tu pleciesz?

[eternal niche]

🍆

[woven scarab]

gm AI fam! 😁

[cedar tide]

@echo aurora Why was step 3 removed 6 hours after it was put in?

[eternal niche]

it was gemini 3

[cedar tide]

The Discord community in RP trolls pissed me off

[tall summit]

is gpt5's context limit on the website still 8k/32k/128k?

[tulip cipher]

ee ti moe opisanie so stima spizdil

[eternal niche]

yes

[tulip cipher]

tvar'

[keen beacon]

For vibes yes , id say its the best

But gpt-5-thinking or high is the new standard for complex stuff

[junior sonnet]

Ok

[viscid timber]

is gpt 5 the best ai model now?

[keen beacon]

No

[viscid timber]

then what is the best

[keen beacon]

Depends on what you use it for

[viscid timber]

searching

questions, basic tasks

[keen beacon]

Id say gpt thinking. For vibes kimi 2 and claude are better

[barren ermine]

what about creativity?