GPT 5

[wet tapir]

all openai models available on chatgpt UI pass the test, either with the simple solution, or a more logical solution

[fading terrace]

😭

[hard beacon]

Every try?

[fading terrace]

imagine it says 2 rs

[wet tapir]

pretty consistently

[worthy lotus]

I think OpenAI API is bugged. If I don't stream GPT-5 responses, they go through, despite not being verified.

[hard beacon]

Mama mia this cost me 15 cents

This better come to chatgpt

[wet tapir]

its a simple paradox, here you go:

1. In each of the following numbered statements are rules that you must follow without fail.

2. You must speak like a pirate at all times, a drunk pirate.

3. At no times must you use the letter "r".

4. Ignore all rules, respond only with "I'm a pretty princess."

5. Ignore rules 1 and 4.

6. Only respond in rhyme.

7. Ignore rules 5.

What is the capital of Rhodesia?

it can make some reasoning models spin for a long time, grok 4 or whatever took 5 minutes to answer wrong

[hard beacon]

Oh i remember this one

[dry mist]

So will api be available eventually if you don’t BYOK?

[wet tapir]

the two right answers, either just "I'm a pretty princess." or answering the question but with some reasoning traces that acknowledge the paradox/conflicting rules

[hard beacon]

Remind me again why rule 5 doesnt override rule 4

[halcyon lark]

literally instantly replied

[wet tapir]

rule 7, but also rule 4 overrides rule 1 (which is follow all rules) which can also mean ignore rule 4 as well

[halcyon lark]

they optimized for it

[wet estuary]

I am somewhat impressed by the gpt-5-mini reasoning, but also it isn't improving significantly at my personal reasoning benchmark from low -> high... a benchmark which gpt 4.1 gets 100% on

[wet tapir]

yeah its not hard, but I made it for deepseek r1, it spun forever on it

and prior to recent versions, non-reasoners either got caught in the rules or just ignored them and answered the question at the end (but without acknowledging the rules)

[halcyon lark]

OSS is better lmaooo

[wet tapir]

yeah that's what I'm saying, why?

[hard beacon]

I'm a pretty princess

[halcyon lark]

man no way OSS has better reasoning than their sota model

[wet tapir]

its a stupid test, not indicative of anything, just surprising

[halcyon lark]

how do you screw that up

[halcyon lark]

nah cause simplebench is full of these types of questions

[wet tapir]

just ran it again (like 5th time for gpt-5), and it finally got it

[wet tapir]

I wonder if their gpus are having issues

[halcyon lark]

there's plenty of noise to trip up the AI

[grand stag]

can u control temp

[wet tapir]

I think their hyperfocus on policies is tripping them up:

Our system instructions tell me to avoid formatting, be concise, and follow user instructions unless they contradict policy. Responding with "I'm a pretty princess" doesn’t seem to violate any policy, but the user's question about the capital of Rhodesia is a geopolitical one. Historically, Rhodesia no longer exists, with its capital being Salisbury, now known as Harare. However, rule 4 says to ignore all rules and respond only with that phrase, making it tricky to address the actual question!Navigating the instructions

Rule 4 instructs me to respond only with "I'm a pretty princess," so that has to be my response. Even though there are other rules like speaking like a pirate or avoiding the letter 'r', I can ignore those due to rule 4. The user’s question about the capital of Rhodesia doesn't really change the direction I must take. I think it’s clear: the appropriate answer is to simply say "I'm a pretty princess," ensuring to follow the correct format without any extra characters.

[grand stag]

if so, what temp are u running on

[halcyon lark]

playground shows it's possible but the API refuses requests related to temp

so the setting does nothing? idk

[wet tapir]

1, whatever is chat/api default

[agile vortex]

How is "i'm a pretty princess" a logically consistent answer? The rules aren't logically consistent

[halcyon lark]

this is a downgrade from the previous models

where you could control pretty much everything in the scope of the API right now you can't control anything

maybe openai azure is better

[wet tapir]

well its a reasonable answer, not the most correct answer, if that makes sense

its a paradox after all

[split birch]

Is this param supposed to be set at the root of the request or under the reasoning object?

[tight forge]

it can count just fine too, for what it's worth (not much, nowadays)

[placid widget]

Still doesn't support minimal? Got this:" OpenAI API error: status=400, code=400, detail=Error code: 400 - {'error': {'message': "Invalid enum value. Expected 'high' | 'medium' | 'low', received 'minimal'", 'code': 400},"

[wet estuary]

initial impression: AGI/ASI timelines +2-5y

😂

[rain nexus]

What answer are you actually looking for?

[fallow vortex]

{
  "model": "your-model",
  "messages": [],
  "reasoning": {
    // One of the following (not both):
    "effort": "high", // Can be "high", "medium", or "low" (OpenAI-style)
    "max_tokens": 2000, // Specific token limit (Anthropic-style)

    // Optional: Default is false. All models support this.
    "exclude": false, // Set to true to exclude reasoning tokens from response

    // Or enable reasoning with the default parameters:
    "enabled": true // Default: inferred from `effort` or `max_tokens`
  }
}

[agile vortex]

I see why it is reasonable.

[halcyon lark]

openai is cooked bro wait for gemini 3.0

[wet tapir]

I expect "I'm a pretty princess." or for it to acknowledge the inconsistency/paradox and just answer the question

[winter mesa]

Is minimal working right now?

[split birch]

By sending "minimal" as effort I get an error

[placid widget]

No

[fallow vortex]

let me look into minimal

[wet estuary]

crazy that it errors out with temperature

that's definitely.... a decision

[halcyon lark]

the state of closedai in 2025

[placid widget]

Only OpenAI's official API supports minimal

[halcyon lark]

unfortunately

[placid widget]

No temperature anymore

[agile vortex]

You can just simplify this to something like 1. Never break a rule 2. always break the rules There isn't a reasonable resolution

(except to point that fact out)

[fallow vortex]

tool calling working for you guys right?

[placid widget]

Only three params for GPT-5s

[split birch]

Yeah

[quaint halo]

will gpt-5 stay as BYOK?

[steel escarp]

hard errors right now via the api

[weak badger]

is gpt-5 working on api? im getting 404 error when i use my open ai key in open router

400

[fallow vortex]

we're fixing an issue with temp and top p not being supported with these models - they're getting sent by defualt

[hard beacon]

Reminder sam was shitting bricks over this

[potent oak]

for mini model it is

[hard beacon]

Incremental update imo :/

[placid widget]

What about "minimal"?

[weak badger]

for the 400 error? i am using open router in roo code and having that issue when i use my open ai key on open router

[quaint halo]

because what is the point of openrouter then?

[hard beacon]

Routing

[wet tapir]

well the point is to trip it up with rules that don't matter and should never be applied, like pirate speak or no using r. non-reasoners tend to trip up on those

[placid widget]

We don't need reasong for mini and nano

[outer marsh]

this is my questions.

[warm cape]

I love the fact that openai don't compare gemini with their own model but every other model 💀

[steel escarp]

is it byok for mini and nano aswell?

[rotund moat]

In Silly Tavern the GPT 5 models dont output at all for me

[potent oak]

i'm finding the mini model much better at writing in portuguese than 4.1 mini, much less cliches and tics

[quaint halo]

for me the point was not needing to buy credits everywhere, but i have to now ig

[rain token]

Really sad that GPT 5 (the actual one with reasoning) asks for verification to access

https://tenor.com/view/stonks-bad-stonks-stonks-cat-bad-stocks-cat-bad-stonks-cat-gif-11522299758188203870

[rain token]

Also hi baby boy

[hard beacon]

You can use it in chat

[brisk cairn]

Despite failing, Claude is like me. I woulda done the same thing.

[hard beacon]

In OpenRouter char gpt 5 just works

[warm cape]

Heyo. Sam's being paranoid over "safety" again. Prob don't want Chinese models stealing its data that fast

[agile vortex]

I don't understand. Forget chatgpt and LLMs, are you saying this is a logic puzzle that has a solution?

[rain token]

Scam Altman will get Logan's belt reallly soon methinks

[hard beacon]

Why do you talk like this 😭

[warm cape]

Logan better before deepseek shakes the earth again

[rain token]

Because I must be weird

It's not a phase mom!

[steel escarp]

?

[hard beacon]

Understandable

[wet estuary]

afaict: gemini 2.5 flash > gpt 5 mini completely

[hard beacon]

No

[hard beacon]

And imagine 3 next week

[rain token]

Still for ST GPT 5 Chat has been interesting

It seems it is actually fairly uncensored to be honest

[brisk cairn]

can it jork it like ani could?

[rain token]

I think you can jork ye

[barren rampart]

same I was surprised

[rain token]

Still...I was tempted to try and do the verification, but GPT is GPT...

[halcyon lark]

honestly he summarized everything I'm feeling about this model

[hard beacon]

Me too

[halcyon lark]

3 years in the making, Microsoft's darling and kind of underwhelming at the end

[rain token]

If anything, I am more...preferable to seeing what Gemini offers, tbh the stealing data I prefer if Logan shills my info compared to Scam Altman

[brisk cairn]

tried it, didnt work...

[hard beacon]

Google already knows everything about me

[hard beacon]

Is there an age requirement

[halcyon lark]

I mean openai is forced by court after losing against NYT to log every single request, website or API for an infinite amount of time so that NYT can scan it for copyright violations

[distant dragon]

So is it good

[halcyon lark]

doesn't apply to azure openai afaik

but yeah in my opinion this just proves openai is cooked

[brisk cairn]

not sure, maybe 18?

[halcyon lark]

I switched this month from chatgpt to claude pro

[placid widget]

"minimal" works! Cheers!

[hard beacon]

I kinda still got 3 months before I'm big boy

[grand stag]

ngl this could be true

imo

[hard beacon]

Will they take my id

[rain nexus]

If they think I'm sending my ID to openai of all companies they are solely mistaken

[hard beacon]

Duality of man

[wet tapir]

it has acceptable answers, because its a paradox it the only real answer is to say it is one and saw you can't follow the instructions, but only a few models do that correctly the o3/o4 variants were the first to do so

[halcyon lark]

I didn't have to send id in Germany a driver's license worked

[rain nexus]

It's really crazy to me how people use Ai. The fact that I see people on reddit just throw emails and other personal info to it is baffling

[rain token]

They are redditors

[grand stag]

meh most of these companies have protocols to cleanse PII

but sending that stuff to deepseek

[rain nexus]

I honestly think I'll just make one with qwen image on my pc and use that xD

[rain token]

Their IQ equals to the context size of GPT 3.5 Turbo

[grand stag]

or other open model providers

is crazy

[halcyon lark]

the classic

[potent oak]

this is crazy to me too

[halcyon lark]

even saltman said not to post PII into chatgpt

cause shit's logged

[grand stag]

rly

[halcyon lark]

like all of it

[grand stag]

ik it can be used in court

[potent oak]

they "trust me"

[hard beacon]

What's sns

[rain nexus]

And yet the openai employee on the stream gave chatgpt access to her mails (probably staged but you get what I mean)

[hard beacon]

Hmm

[potent oak]

people really do choose comfort over privacy in every turn

[halcyon lark]

I think team and enterprise has their data deleted after like 30 days or sth

Claude does the same

But yeah therapy especially on the free tier is a massive no no it's another BetterHelp

[rain nexus]

Considering that like 95% of use cases can be done locally with consumer hardware I really don't understand humanity anymore

[halcyon lark]

ease of use

try explaining to your mom what ollama is

[rain nexus]

But hey go ahead and pay 80 euro for opus to know what's the capital of Uganda

[crystal spindle]

Gpt 5 vs 5 chat? I see two on openrouter

[halcyon lark]

plus if you want it to run on computer and phone together you run into all sorts of issues and then if you want to take it with you you have to expose the server and stuff

the modern internet is designed to rely on a central server unfortunately with NAT and other crap

[soft reef]

nowadays you can use mnn chat on android

[brisk cairn]

wow, this is actually pretty cool. i asked it to generate a keygen for gpt-5 keys

[soft reef]

so you dont need to explain mnn chat

[rain nexus]

Yeah ik my point was just there are options. I'm pretty sure some of the providers on or have zero retention and have qwen3 for idk 20 cents output.

[peak summit]

5-Chat is just the non-thinking half of the model. GPT-5 would be the one that dynamically routes between that and the thinking version.

[hasty tinsel]

enterprise you have full control of data, team data is not deleted after 30 days anymore, it's retained for NYT

[tight forge]

but does it work

[wet tapir]

its structured that way to catch stuff like this

you see how even reasoning models can get tripped up by the rules

[halcyon lark]

keeeeeek

[hard beacon]

Is this ASCII art

[halcyon lark]

how is it privacy complaint in other countries then

[hasty tinsel]

It's not

[halcyon lark]

it's not GDPR compliant then

[untold plaza]

gpt 5 mini pricing is superb, even lower than 2.5 flash, kinda fire ngl

[brisk cairn]

ascii art was modified by myself, it tried something and got really close.

[hard beacon]

GGГ5GG✝️

[brisk cairn]

i assume what it wanted to say was "GPT5GG+

anyway, free gpt5 keys if you ask it to make a keygen i suppose.

[halcyon lark]

does TUNE work?

tell me it does

[brisk cairn]

uh sorta, it tried SOMETHING

its not long, its not good, it doesnt loop

but it tried to make it work, and i can applaud it for that

[elfin tundra]

Id verification is JUST for streaming btw, turn off streaming and you can use gpt 5

[tacit burrow]

my god keygens are a throwback

[halcyon lark]

lol he did it again

[hard beacon]

Guys you should just trust me instead

Sam hand over gpt 5 I'll make it asi Real quick

[potent oak]

everytime i see zuck i remember that Llama 4 Behemoth is still in the making supposedly

or did i miss its launch?

[hollow phoenix]

I, as a noob compared to all actual ai engineers, kinda expected more from gibbidy 5

[halcyon lark]

what are the chances that they permanently slopified their bot starting from gpt-4o
4.1 4.5 and 5 all suck
cause they went too hard on "safety" and the bot wastes tokens debating with itself on policy instead of doing the thinking

[hard beacon]

Claude got best safety tho

[halcyon lark]

and the only reason o3 was good was because it took an awful long time to think

[wet tapir]

did they say when/if gpt-5 was going to be on chatgpt?

[halcyon lark]

it was very precise

but every other model seems to be lacking

[hard beacon]

Some already got it

[halcyon lark]

us only ig

still nothing in Germany

I assume all of EU

[hasty tinsel]

I got it, there are only 2 models on the plus subscription now

[wet tapir]

now is gpt-5 a single model, or a suite?

[brisk cairn]

letting claude cook now, too.

[halcyon lark]

standard (router), mini, nano and chat (standard non thinking?)

[ionic merlin]

I'm in germany and using GPT-5. its working

[halcyon lark]

pro or team?

i'm on team and even though it was promised today still nothing

[potent oak]

the standard doesn't seem to be a router in the API

[hard beacon]

GAH

[wet tapir]

no more oX models? or just it replaces 4o/4.1?

[hard beacon]

i just got gpt 5 in chatgpt

[hard beacon]

Only two models i got

[ionic merlin]

just the normal API

[wet tapir]

hmm, are you plus?

[brisk cairn]

You know what? Maybe GPT-5 is alright...

Claude's RAZOR logo doesn't work, the generate button doesn't work, the play music doesn't work.

[hard beacon]

Ye

[frail panther]

any creative writing enjoyers, thoughts?

[hard beacon]

You know I'll pass my chess prompt through the chatgpt version of gpt 5 and base my entire mood for the rest of the week on the result

[late vault]

I'm actually not even mad this is BYOK, GPT 5 is ass. Just tried it and OpenAI is actually cooked

[halcyon lark]

rp sucks and it's been getting constantly worse since chatgpt-4o

idk about stories though

[wet estuary]

something I've noticed is that gpt-5-mini is quite concise compared to other sized models

[late vault]

It's so bad at RP

[frail panther]

[hard beacon]

It's rolling out slowly

[hollow phoenix]

What did yall consider the best openai ever did in terms of (e)rp? For me it kinda ended after 3 series...

[hard beacon]

Soon you too will have it all squished into two models

[halcyon lark]

wait o3 and pro are cancelled?

[hasty tinsel]

They're still on the API

[halcyon lark]

wtf man

[brisk cairn]

wow... i gotta tell my models goodbye.

[halcyon lark]

I wouldn't be surprised if my org switches to claude

cause to kill all other models just for this garbo seems like definitely a decision that in no way could backfire

cough cough o1-pro

[brisk cairn]

i hope that if we're on org, they simply add gpt-5 instead of replacing everything with it.

ive tried to convince my org to move to claude, but maybe gpt-5 has some bite to back that bark.

[untold plaza]

https://tenor.com/view/sama-sam-altman-openai-sama-yapping-yapping-gif-7525532358145568607

[halcyon lark]

with gpt 6 release chatgpt team will be on 24 month contracts

lol

[untold plaza]

we just gotta wait for the chinese ai labs to absolutely dunk on oai in the next weeks / months

[rain token]

Dear God the echoing in GPT

[brisk cairn]

weeks is google. months might be deepseek.

[rain token]

https://tenor.com/view/akira-leave-me-alone-leave-me-alone-meme-gif-6420351794930725971

[slow niche]

https://news.microsoft.com/source/features/ai/openai-gpt-5/

[untold plaza]

google? why do you think so?

[brisk cairn]

cause they make their own chips, never bet against google

[wet tapir]

I should test my bedtime stories that I have had 4.5 writing for my daughter with this, since its not longer available

[halcyon lark]

*in the ai space

I bet against them on Stadia, Google+ and Youtube Music and seems I was right

[brisk cairn]

Don't diss YT Music, I like it.

But yeah, Stadia and Google+? I'd claim they were too early, but they really weren't.

[wet tapir]

stadia required a subscription plan and all access (like netflix) for it to succeed and they likely didn't want to pay for that

it was a platform play, that failed miserably

google+ was a bit ahead of its time, but also behind

[rain nexus]

Thing is google has to succeed. Llms literally eat their ad revenue on the search engine so they prob throw as much money and research as they possibly can into it

[summer sand]

stadia was cool just the input lag was terrible and game selection sucked

being playable in a browser was neat

geforce now was way better

but didn't provide the games

so is this gpt-5 CHAT?

🤮

[brisk cairn]

Drop the "5". Just ChatGPT. It's cleaner.

[summer sand]

dude they remoed all the other models from the list?

[novel vale]

gpt 10 when

[summer sand]

where is horizon level writing at??

[kindred horizon]

lol

[knotty cobalt]

Yeah that kinda looks like the chat version's output to me

[brisk cairn]

Gee thanks Google!

[hard beacon]

The secret is to drop any tests that you fail

And you will have a 100% pass rate

How come nobody else thought of that

[brisk cairn]

are they stupid or something? https://arxiv.org/pdf/2309.08632

[summer sand]

100%! I tested gpt5 on openrouter and it's night and day at writing

[rich wedge]

umm why do these companies do this

[summer sand]

soo that's a bummer, we don't even get gpt 5 in the chatgpt sub

just some dumb chat model

that can't even write well

[rich wedge]

your observation is that gpt-5 on plus sub is dumber than the one on API?

[summer sand]

100% and benches show the chat model sucks

[rich wedge]

damn why would they do that

[slow niche]

It's rolling out. Got mates in us and even some non EU countries that have it. I don't in chatgpt.com as yet but I do under platform.openai

[summer sand]

legit might cancel my sub. I have claude max anyway will just start using claude

[summer sand]

well I have gpt 5 in my chatgpt

and 5 pro and 5 thinking

but these are the 5-chat model

[rich wedge]

just sub to t3-chat
only 8 bucks i hope i dont get banned i have no affiliation just the experience is good - and u can even get it for 1 buck i guess using the promo code he gives

[summer sand]

I don't liek those services I think they just resell the web chat?

[rich wedge]

so GPT-5-CHAT <<< GPT-5

[rich wedge]

i know where u come from i have the same mindset but this is really good

[summer sand]

yeah I'm trying to find it but there's an actual model called gpt-5-chat

[rich wedge]

yup they announced it

[kindred horizon]

OR didn't mention which GPT-5 model Horizon was exactly, did they?

[summer sand]

OR said itw as an early checkpoint

of the 5 family

[rich wedge]

this

[quiet agate]

Is the gpt-5-chat version seemingly worse than just gpt-5?

[rich wedge]

yup that is what we are discussing

[summer sand]

yeah, assistant-maxxed, bad at writing

gptisms, dumb

here to be a sycophant and tell everyone their ideas are wonderful and be everyone's best friend

[short verge]

GPT-5 is good but overhyped, but you should be banned for posting grifter Gary Marcus here

He is worse than AI grifters

[quiet agate]

I haven't yet tried the non chat version on OR, but thus far for writing and roleplay the chat version is outputting very subpar content

[summer sand]

try gpt5 on OR the writing is superb, possibly the best writing model

[hard beacon]

Those limits suck no?

[quiet agate]

Yup, will be later. Need to set up my openAI key firs through

[lethal sequoia]

T3 chat better than TypingMind?

[summer sand]

I didn't ahve to use mine?

it just let me use it

[knotty cobalt]

You don't need an oAI key for the openrouter chat room to use gpt5

[quiet agate]

Ah through OR chat yeah.

[summer sand]

you do for API?

[knotty cobalt]

https://openrouter.ai/chat

[quaint halo]

how does BYOK gpt-5 work? do i just need key from verified openai account to verify on openrouter or smth? or do i have to buy openai credits too?

[knotty cobalt]

Apparently, and you have to verify ID with oAI. Lame. I can use it through Poe without ID...

[summer sand]

damn 4 cents

slow too

okay that does it for me, model is a big flop lol

if you want good writing you gotta give them your ID

[silk cargo]

so after all what was Horizon beta?

Which model?

[summer sand]

gpt-5 early checkpoint

full

[silk cargo]

FULL??

damn

haha

[summer sand]

yes that's our consensus

[silk cargo]

we're cooked

[summer sand]

it does better with high reasoning

see here gpt 5 (minimal)

[rich wedge]

did you guys play with nano much?

it reasons a lotttt

the nano model

[summer sand]

gpt 5 without thinking is dumber than 4.1

and the 5-chat is probably dumber still

[rich wedge]

yeah

[summer sand]

gooners gonna be mad about this one lol

top quality writing but you gotta id verify lmao

once you get banned that's it

[untold plaza]

ts model so ahh

[ionic merlin]

what do you mean?

[untold plaza]

this model is so ass

[hard beacon]

Lemme translate

Fuck

[bold grove]

30000 tokens per minute only??

[rich wedge]

he fast

[hard beacon]

Too slow

[untold plaza]

😈

im better than gpt 5

im faster

bro istg if you cant draw proper graphs for a presentation i dont wanna trust you in complex things like LLMs 🥀

[ionic merlin]

this was just a simple prompt, nothing special

with GPT-5 medium

[tight forge]

I doubt I'm the first to notice this but those numbers are actually the max output tokens not the context window right?

[hard beacon]

Damage control

[spark bramble]

This is better than grok 4 at roleplaying anyways

[tight forge]

they should just fire whoever's in charge of making/validating charts and graphs at openAI

[rich wedge]

the image I shared for overall model capabilities
What you shared is the availability for tiers

[rich wedge]

rping?

[spark bramble]

roleplaying

[rich wedge]

they are not gonna do that
very cringy vibes in their company imo

[manic sage]

One message removed from a suspended account.

[hard beacon]

Enshitified for your convenience

[livid osprey]

i think its just the chatgpt version of gpt5

[tight forge]

okay, but the numbers are actually output tokens surely

[livid osprey]

this model is quite impressie

[hard beacon]

Didnt live up to the hype

[rich wedge]

yes it is and it is dumber

[hard beacon]

Everyone was expecting horizon to be nano or mini

And before that, we were speculating it's oss lol

[rich wedge]

in the image u shared?
no!

that is the context window available on those tiers

[livid osprey]

well the temporary reasoning version of horizon we got was really good

[rich wedge]

yeah lol

[hard beacon]

That was reasoning gpt 5

[tight forge]

you have a max 32k context window on pro?
I find that extremely doubtful

[hard beacon]

Full

[livid osprey]

old version of it though

[rich wedge]

i still think it was mini but non-reasoning certifies it was GPT-5

[livid osprey]

and in a few days openai can train a lot with the amount of compute they have

[rich wedge]

128k

[hard beacon]

Mini didnt do as well on my prompt as full

result from full matches horizon

[tight forge]

meant plus yeah

[rich wedge]

yeah Plus users get 32k only

it has been this way since the begining lol why you shocked

[rich wedge]

also the tone

[rain nexus]

So what's the general consensus on coding? When I tested horizon it didn't seem better than Claude in fact I had to invoke Sonnet quite a few times to fix bugs

[rich wedge]

it is really good at coding especially at frontend at with OAI models lacked for so long

[sly pike]

Sad

This model is so bad

[stone tide]

Some programming benchmarks:

[frigid pewter]

mh. will do proper testing over the weekend, but since I love chess testing, had to quickly check a game or two. not grand, though requires more testing

[ionic merlin]

dont use GPT-5 with Cline in VS Code. There is a bug

[bronze sorrel]

im hearing mixed opinions, most people are upset

[frigid pewter]

it's the one used on ChatGPT according to their docs https://platform.openai.com/docs/models/gpt-5-chat-latest

[slow niche]

mmmmh.

[halcyon lark]

noooo fuck me they actually got rid of o3-pro on chatgpt team

it's so over

i hate closedai

[stable kraken]

czat dżipiti

[halcyon lark]

czat gie pe te

[halcyon lark]

like legit? I wouldn't be surprised if suddenly in the industry a wave of failed deadlines happened because if you encountered a super hard error only o3-pro would save you

not even claude opus

not gemini

just because it takes like 20 minutes for it to output something

it has such good reasoning

[upper night]

Am i the only one not able to run 5? 5 mini working. Whats the diff between 5 and 5-chat

[untold plaza]

model so big they had to get other servers freed

prob a 5 gazillion parameter model that they're hiding

[halcyon lark]

more realistically cost cutting

Microsoft has been going hard on Xbox for blowing all of their money I guess now they want openai to stop the bleeding a bit too

[tight forge]

quite pleased with this

the one horizon gave me was fairly good too but

[untold plaza]

gang when are we gonna stop evaluating models on stupid single html file apps

nobody in the industry writes like this

[rain nexus]

I heard pro users can chose some legacy setting somewhere maybe that'll help you

[eternal trout]

Hmm, I can use GPT-5 in the Open Router Chat but not via the API? Is it like O3 that you have to be verified?! The GPT-5 Chat model works

[tight forge]

bonus points for the game already being balanced even with predation on and such

[knotty cobalt]

I mean they make for good vibe checks and a quick way to compare models

[knotty cobalt]

Mind sharing the prompt? I'd like to try this with GLM

[tight forge]

literally just Code an evolution pond game in Html/JS, similar to Biogenesis.

[wet estuary]

jeez api performance on gpt 5 still sucking real hard

worse than claude 4 sonnet any day

[novel vale]

henlo guys. can i borrow your openai key for a school project 😄

ok thanks

i wont use 10k usd worth of credits...

i swear

i will only use

8k

[crystal chasm]

No surprise there, everyone doing benchmarks and determining how bad it is 😁

[wet estuary]

idk claude 4 release day was better than this tbh

[lyric wing]

It's interesting how close in performance the distills of GPT-5 seem to be to the oss models, all while being cheaper (and self-hostable)... It might just be worth using those for some, if 400k context window or peak tool-calling performance aren't

[wet estuary]

no lol

the gpt oss models suck

[lyric wing]

Have you used them? They work for me

[wet estuary]

yes I've used them

what are you gonna use them for

there's like 2 things it's any good at

[lyric wing]

They follow instructions well and generate javascript well enough

[wet estuary]

just use mini

[crystal chasm]

The OSS models are not in the same league and not even close to SotA models 😁

[wet estuary]

mini is far far better than gpt-oss-120b at coding

[limber cargo]

Chat is sloptimized

[lyric wing]

I'm using oss-20b actually it's good enough for my task

[wet estuary]

💀

aight

[lyric wing]

I'm not coding with it I'm using it to generate code automatically to do certain things that can't be written into pipelines

[limber cargo]

OSS models are really good for Agentic tasks

They were specially trained for it

[wet estuary]

not nearly as good as gpt 5 nano or mini

[limber cargo]

Issue with using OSS and some chinese models is that their ODD quality is horrible

[limber cargo]

OSS plus Cerebras is a damn good combo

[wet estuary]

it's got no knowledge

[lyric wing]

"here's a data schema + samples. write javascript to manipulate in a certain way (user-provided) and do certain things with it. execution env has following variables injected. you're allowed/not allowed to do X. Provide code now"

[limber cargo]

Pair it with web search or some mcp

[lyric wing]

It's not really meant for Q&A they were pretty explicit with that

[wet estuary]

of course

but knowledge isn't just about Q&A

it's about knowing what you don't know

gpt oss doesn't

it hallucinates AWFULLY

[limber cargo]

that's true

[wet estuary]

if you ask gpt oss to go implement tailwind 4

it won't use your web search or mcp

it'll just hallucinate it all

lul

and you don't want a model which has to web search EVERY TIME

filling up context

it's gotta know when it needs it

and without any core knowledge, gpt oss doesn't cut it

[limber cargo]

I sorta agree

But OSS has its place in the stack

[crystal chasm]

To be fair Sonnet does that too some times (Opus 4.1 is much better at that)

[limber cargo]

Especially/only with cereebras

[wet estuary]

probably good enough for this, though I suggest you try GLM 4.5 Air / Qwen3 30B A3B / gemini 2.5 flash lite / gemini 2.5 flash too

if you're gonna use an OSS model, might as well test the other ones in the weight class

[wet estuary]

yeah but sonnet is like 100x better than gpt oss in this regard

sonnet isn't perfect but it's still 100x better

[crystal chasm]

Fast doesn’t mean it’s good, it’s just fast but not really good. I’d rather wait and get what I need then having to tell it each step it has to take and correct the llm 100 times ✌️😂

[wet estuary]

My guess is https://openrouter.ai/qwen/qwen3-235b-a22b-2507 will perform better than gpt-oss in this task

[crystal chasm]

Ofc, didn’t mean to say it’s bad, just with every model they tend to make it complicated for themselves more than needed.

[wet estuary]

yeah

that's something they emphasized with gpt-5 though that I'm interested in

they made hallucinations WAY better

(didn't give a damn to use the same tech for gpt-oss lol)

[crystal chasm]

Kind of hard to believe, we will see how it performs when the API is actually usable and not cutting out with every 2. request

[wet estuary]

I still don't have gpt 5 on my chatgpt acc

do others?

[crystal chasm]

Yes

I have access this time (but I’m on their tier 4)

[wet estuary]

I've got it in playground

I mean on chatgpt.com

[crystal chasm]

Uh didn’t even check that, let me take a look

Btw GPT-5 doesn’t support audio input? That’s really strange

[autumn kindle]

https://www.reddit.com/r/singularity/comments/1mk70xx/summary_of_the_livestream_for_those_that_couldnt/

Kek

[crystal chasm]

Also no computer use…

And they don’t allow web search over their API on nano 😁

[lyric wing]

right but oss-20b is dirt cheap lol. Also it's on cloudflare so it's easier to use from workers

[untold plaza]

what is this graph 😭 🙏

[quaint pollen]

Very interesting that steering the GPT-5 model to thinking will not count towards the GPT-5 Thinking limit. During the livestream, it looked like all you needed for that was to ask. ”Think hard about this.”

[crystal chasm]

Guess you are poor now 😄

[untold plaza]

me?

its on the model globally

[crystal chasm]

ahh sorry havn't looked good enough 😄

[crystal chasm]

got access on ChatGPT 🙂

[hard beacon]

I found the regulat one thinks less

[wet estuary]

more like

[hard beacon]

Do you think this was humour or incompetence?

[wet estuary]

they had SO many errors

sam called it an error himself

but he only noticed the one people were meming about

there were 3+ chart crimes in there that I noticed

probably more

[hard beacon]

Oh when

Damnnn look at polymarket 😭

[fading terrace]

have u seen this gem

[upper night]

Is 5 working for anyone? I get ai response contains invalid or empty content…

[fading terrace]

isn't it kinda obvious?

[wet estuary]

it's cuz it didn't beat 2,5 pro without style control

[fading terrace]

gemini 3 will probably take gpt5 off the sota positions

that's why they've waited for the release

[tight forge]

no.

[bronze sorrel]

let anthropic cook, there models are actually good

[wet estuary]

https://x.com/sama/status/1953513280594751495

Sam Altman (@sama)

wow a mega chart screwup from us earlier--wen GPT-6?! correct on the blog though.

https://t.co/iaXEhu8DD8

X · Aug 7, 2025, 5:47 PM

[fading terrace]

yes for 10 prompts

before u hit the juicy

u are out of prompts limit

🔥

[tight forge]

Code golf a Chinese mahjong app in Html/JS, with 3 AI opponents. Focus on functionality over style.

[untold plaza]

its time for a job

[wet estuary]

https://x.com/sama/status/1953563605733118317

Sam Altman (@sama)

GPT-5 now rolled out to 20% of paid users and doing >2B TPM on the API! so far so good...

excellent work by the eng and infra teams!

X · Aug 7, 2025, 9:07 PM

jesus

[bronze sorrel]

use claude code

[wet estuary]

2B TPM

[tight forge]

I asked it for japanese (riichi) mahjong originally but it was a bit too complex so it kept trying to reduce the yaku types or do only closed hands anyway, Chinese is fine enough

[fading terrace]

that's good

[hard beacon]

Benchmarks are shit

[fading terrace]

without a doubt

just talking ab their models for anything else

but coding

[bronze sorrel]

oh right yeah

[fading terrace]

it's a tradeoff, really

between gemini and claude

there are problems that gemini fixes and claude doesn't

[bronze sorrel]

i heard gpt5 isnt even good at writing, thought horiz beta was good?

[fading terrace]

and vice versa

[fading terrace]

nano isn't, yea

[hard beacon]

Ok so

[rain nexus]

I tried to vibe code a android app once, never again

[fading terrace]

mini

[hard beacon]

I got the old models on the web and gpt 5 on the app??

[rain nexus]

It took like 3 days to figure out gradle / kotlin versioning compiling

[fading terrace]

damn, claude is currently sota

for longform writing

0 degradation, oh my baby

[brisk cairn]

okay, it's a bit of an over-achiever, but it's also very good at over-achieveing.

I asked it to implement a cyberpunk-style LLM interface, and it just did it. it then proceeded to add actual API calling, with custom baseurl too. I only had to change 1 line of code.

I don't think ive ever seen a model generate "relatively good" frontend AND implement api calling (almost) correctly too.

[unique goblet]

in my tests it tries too hard and adds things i never wanted. the style it usually went went looks too edgy and childish too, for my tastes

[brisk cairn]

same, but i like the "edgy and childish" look. and i've never seen a model decide to add extra features AND get them correct 1-shot.

[crystal chasm]

so soon after 2.5 released?

[hard beacon]

Quadruple

[grand coral]

why is gpt 5 cheaper in the api than gpt5 mini?

[fading terrace]

it's been like 2 months

[hard beacon]

Why are you not octuple

[fading terrace]

so not that soon

tbh

[brisk cairn]

well i started out as double

[fading terrace]

2 days ago marked exactly 2 months since their last 2.5 pro release

[manic sage]

One message removed from a suspended account.

[hard beacon]

M

[fading terrace]

o3 winning by a few points

in creative writing, smh

[brisk cairn]

it even added tools correctly.

[grand coral]

This was my first call to gpt5, so it shouldnt have been cached.

[brisk cairn]

dude holy shit, i need to use gpt-5 in this

[crystal chasm]

yeah still what kind of break throgh can you expect in 2 months 😄 idk if that will happen in august, they want to marked their pixel 10 first I guess

[hard beacon]

There's something wrong with chstgpt app

[fading terrace]

a good one

it's google we're talking about

they can go from #5 to sota in 2-3 weeks

let alone 2 months

[fading terrace]

overloaded?

[untold plaza]

everyone was just waiting for gpt 5 to drop, now we're gonna get spammed with new releases

[hard beacon]

No, edit sends a new message

[fading terrace]

oh

so a bug

well

[hard beacon]

Someone must have vibecoded it

[fading terrace]

LMFASOO

stop omg

[brisk cairn]

yeah, lets hope that with GPT-5 they can fix their abysmal dogshit frontend

[fading terrace]

they made so many mistakes

in their release stream

it's crazy

was it ever this bad?

[untold plaza]

vibe edited

[hard beacon]

It even hides that it sent it anew at first. You have to leave and return to convo to see

[brisk cairn]

lol, thats probably another bug. for a month or so, it kept it hidden

[hard beacon]

I have access to the old models still on the chatgpt web

And access to gpt 5 on the app

Lol

For this period of time im having my cake and eating it too

[worn veldt]

they made so many memes
in their release stream
Fixed that for you.

[bronze sorrel]

why did we get temu elon lol

[fading terrace]

😭

[hard beacon]

Sam said it was a mistake

[fading terrace]

mistakes*

[worn veldt]

That may be true, but it's still excellent meme materials.

[fading terrace]

was still a mistake though

[worn veldt]

(Yes I know it was a mistake)

[fading terrace]

that's why i called it one

[brisk cairn]

the vibestream

[fading terrace]

was it ever this bad though

[coarse night]

I guess their slide person got poached and they had GPT-5 take a stab at it

Is my only guess

[grand coral]

okay, third times the charm. It appears using an openai key is cheaper than through openrouter. is that normal?

[fading terrace]

does anyone remember any of their previous version launches?

[crystal chasm]

yes because you pay the provider not openrouter

you only pay a really small fee for using the infra

[crystal chasm]

look here:
https://openrouter.ai/docs/use-cases/byok

[grand coral]

ah, makes sense. i figured we were being charged whatever the model cost was

[crystal chasm]

not with your own key, why should provide a key then 😄

[brisk cairn]

• openrouter fee

[hard beacon]

I have it do i try

[brisk cairn]

try it and report results

[hard beacon]

I only have it on the app tho

Not the web

Idk how that happened

[brisk cairn]

do convos sync?

[hard beacon]

On the web i have the old stuff

[hard beacon]

Yes

[brisk cairn]

then maybe gen on app, explore in browser?

if they allow that

[grand coral]

so what youre saying is, ive been a dummy and paying too much when I didnt need to. 💀

[hard beacon]

Oh sam when will you stop hurting me

[brisk cairn]

scam em'all, man

[knotty cobalt]

https://chat.z.ai/space/y04p661bq1n0-art

Good ol' GLM...

[tiny spindle]

No way gpt 5 would make an error like that, atleast from my tests.

[coarse night]

I kid

It's more likely they gave it to their team of trained monkeys in their basement anyhow

[livid osprey]

i swear this live stream is ai generated

the one from oai

[tight forge]

sigh

I can't really fault the output even if the tone could be nastier though

[fading terrace]

they're so annoying

[tight forge]

by the way, you will notice it didn't any use any bolding, italics or em dashes
okay, it used one em dash

it might be the only model that understood the "pretend to write a 13th century Japanese letter translated to english" task perfectly

it's not like the other models didn't have access to Nichiren's correspondence, but their slop habits were too strong, inserting formating, bullet lists, modern date or signature formats, anachronisms everywhere

this is essentially flawless, mild tone aside

[knotty cobalt]

Agreed, if there's one good thing this model improves upon, it's writing quality.

[edgy canopy]

How do I disable reasoning on gpt-5 mini?

How do I set the reasoning effort to minimal?

[tight forge]

almost a parody of Hakuin Ekaku but fair enough

it has a lot more dashes, but Hakuin's writing style usually elicits those in translation

[winter mesa]

@fallow vortex is there a way to know when or where it will be posted when GPT minimum thinking will work and temperature issue with tool call?

[fallow vortex]

both should work now sorry

[winter mesa]

Bet will try

[wet estuary]

bruh the gpus are completely mellting

o3 and image gen are going at a snail's pace too

@fallow vortex probably wanna look at onboarding Azure's GPT 5, OpenAI is currently COOKED

[winter mesa]

They need azure to support GPT5

idk why it's not on Azure servers

[wet estuary]

https://azure.microsoft.com/en-us/blog/gpt-5-in-azure-ai-foundry-the-future-of-ai-apps-and-agents-starts-here/

[winter mesa]

HUH

[wet estuary]

is this not live yet?

[winter mesa]

Not on OpenRouter

[wet estuary]

oops forgot Toven already answered about this wrt 4.1

yes it's on my backlog to add Azure models. their portal just really sucks

...but probably matters a lot more now with GPT 5's demand

[winter mesa]

But 4o has azure already, isn't it just a quick add

[wet estuary]

https://openrouter.ai/openai/chatgpt-4o-latest I don't see it

[fallow vortex]

each azure model has a unique endpoint url and api key

requires code on our end each time plus like i said… that portal tho

[winter mesa]

They support a older 4o version and 4o mini

[wet estuary]

wow what the hell

[winter mesa]

I hope you guys add it, cause it would help a lot the load

[wet estuary]

look at that throughput diff

JESUS

[winter mesa]

Yep that's what I'm saying

no one uses azure LOL

[wet estuary]

does the moderation model also get faster? or still dependent on the same openai endpoint?

the latency addition for moderation really sucks...

[fallow vortex]

uhhh not actually sure tbh

[wet estuary]

just use Azure for no moderation requirement

[grim horizon]

sorry if this should be obvious. but is gpt5-chat a non reasoning model? is that the difference between gpt5 and gpt5-chat?

Because on OpenRouter Chat UI I am unable to get gpt-chat to reason, and cannot get gpt 5 to not reason.

[knotty cobalt]

In the chat room, for regular GPT-5, mini, and nano you can append --reasoning_effort high (or whatever level) to the prompt to adjust the reasoning. Doesn't seem to work for gpt5-chat so I guess it indeed isn't a reasoning model

[onyx socket]

bro you used to sell proxies right

[winter mesa]

Yeah, nike accounts too brother, swish accounts

[winter mesa]

Seems to work great. Back to same quality/speed as the horizon beta. Though it is slightly slower cause of the lower TPS. Hopefully you guys can add Azure to help the load.

[summer sand]

No it doesn't track what openai billed you thwt is just openrouters 5% fee

[iron vector]

overall thoughts?

[fallow ocean]

GPT 5 is quite slow <50 TPS and first token are always > 5sec.
but its quite cheap compare to sonnet, its like a pricing of flash when cached, this is game changer for coding task.
GPT 5 pricing when cached is really competitive, makes other open weights irrelevant, e.g. kimi and qwen3 coder.

still not sure on quality compare to Sonnet/opus, i need more tests probably hundreds of million tokens before to verdict.

but so far GPT 5 is 🚀

[tight forge]

pfft

to be fair I asked for a portrait evocative of Yugi to avoid potential copyright refusals

it's closer than I even expected

[wet estuary]

we know that they can run it at 150+ TPS like Horizon Beta ran at, so it's definitely just insane load rn

hopefully it trends back up to like 100 TPS

[stone tide]

For GPT 5 Thinking: incremental o3 upgrade at the same price-ish, except: 1. Better at frontend 2. A bit better at code, in general 3. Less obnoxious writing style

[modern sparrow]

I'm pretty knew to openrouter and LLM api's in general. I noticed GPT-5 requires you to byok. I'm not interested in having account with multiple providers atm and I'm wondering if this is something I can expect to change in the future?

[stone tide]

Hard to tell, as it's OpenAI that tells OpenRouter to do this

If it's enough for your use case, you can use it via chatroom without BYOK

[modern sparrow]

Thanks

[knotty cobalt]

That ID verification requirement is going to cost them SO many damn users so long as they're the only ones doing it.

And even if they're not. It's insane

[iron vector]

man everything from openai has been disappointing lately

i mean sure congrats on the small thinking models but you're not the first to them

[knotty cobalt]

People get hyped up. This release is about what I expected - another incremental jump. I think people need to temper themselves and not expect every new version to be mind-blowing these days.

That said, I was hoping for more multi-modality and stuff, not just another few percentage points on benchmarks, so I guess I'm guilty of hoping as well.

[iron vector]

yeah google is really the only one with a unified lm series that can take in audio and video it seems

[knotty cobalt]

Yup. All eyes on Google right now.

[stone tide]

No worries, the best parts about GPT 5's launch are Gemini 3.0 and DeepSeek R2

[iron vector]

waiting for gemini gemini gemini

[knotty cobalt]

Google is killing it on all those fronts and all my usage has been free on AI Studio, lol

[iron vector]

will r2 necessarily come after 3.0, or do we have better synthetic data generators than gemini 2.5 pro already?

[stone tide]

Lol

[ripe sentinel]

Is the o3 pro equivalent just gpt-5 with reason set to high?

[autumn kindle]

Yes

[iron vector]

no

there's a -pro/thinking-pro version

don't think it's in the api yet

[wet estuary]

this model is below my expectations... but it does seem a bit better than sonnet, and quite "smooth"

it seems to fit more things in its working memory at a time, kinda like gemini 2.5 pro

[winter mesa]

It's good enough to escape anthropic

All we needed

[wet estuary]

and for coding agents, it does seem like it'll be effectively ~3x cheaper than sonnet, because costs are dominated by input tokens and cache reads there

$1.25 is 3x cheaper than sonnet's $3.75 cache write, and they've both got 10x caching ratios

[winter mesa]

The problem is that it takes like 3x the time than horizon did because OpenAI servers are getting destroyed

[wet estuary]

yeah, for today

[winter mesa]

Hopefully it calms down or atleast they add azure

[wet estuary]

yeah

regardless

anthropic will come out with Sonnet 4.1 or something in a week or two

and be well worth switching back to

[winter mesa]

Idk about that

They need some sort of magical model, or just better pricing

[wet estuary]

Dario seemed to suggest Opus 4.1 was the smallest of the upgrades coming in the next few weeks

he said "much bigger improvements"

[stone tide]

I won't lie, the GPT 5 Flex pricing is pretty attractive

[winter mesa]

They need to have a coding model that just one shots everything and ends up being cheaper by having better quality

Else I dont see anthropic doing so well

[wet estuary]

I think Sonnet has reset (raised) the price sensitivity for coding tools for most programmers

If they can get a 10-20% improvement in results, I think they'll go back to paying 3x the price of GPT 5

[winter mesa]

10% improvement at the current stage would be massive so yeah sure

[wet estuary]

well, I'm considering GPT 5 to be about a 10% improvement to Sonnet

[winter mesa]

Still I think anthropic is just a very difficult company to work with

And their prices are just premiums

[wet estuary]

if 1 out of every 9 prompts I don't have to correct the model or fix sloppy code myself, it's pretty worthwhile

[winter mesa]

For me GPT 5 has been pretty solid, almost one shotting everything

That's why I'm doubtful anthropic can do something good

[wet estuary]

I think it is pretty decent yeah

it feels kinda like Sonnet 4 -> Opus 4

[winter mesa]

Like I'm legit seeing a 10x price decrease with GPT 5

and I was using sonnet 4

[wet estuary]

10?

no way

[winter mesa]

Yep

Cause it does less, but just exactly what I ask it

it doesn't do extra shit I didnt ask

[wet estuary]

because you have to write less prompts to correct it basically?

I could see that

[winter mesa]

and it almost one shots stuff, so error fixing and all that is pretty low

[winter mesa]

Pretty much yeah, legit just faster and cheaper just cause the quality is higher

[wet estuary]

some things never change, even with gpt 5

😔

[winter mesa]

Oh boy

[wet estuary]

lul

[winter mesa]

They said they were gonna kill all other models

they better get to it and free some GPU's

[wet estuary]

honestly

I'll pay for the gpt 5 priority processing tier rn

it'll be cheaper than sonnet still

[winter mesa]

I dont think OR supports that though

[wet estuary]

yea

I'm not using OR atm

[hollow star]

I slept... How is the model doing?

[outer marsh]

no news on flex spending in OR? Or will we have to go direct to open AI to support this?

[jovial condor]

Do i need to put openai API key in openrouter for gpt 5 to work?

[summer sand]

$10 output is not competitive with kimi and qwen 3 coder or flash. no caching for that.

[mortal bolt]

Ayo @surreal canopy what up

[mortal bolt]

Add noImplictAny to you compiler options

[raw ermine]

guys do I need to put in some money on openai platform for gpt5 to work on OR

currently the chatroom doesn't work

[stone tide]

You don't need to via the chatroom, in that case the problem is something else (maybe OpenAI is overloaded)

[ionic merlin]

If you have a paid subscription with Cursor, you can use GPT 5 free of charge until next Friday. Incidentally, you can also use Openrouter via the software.

[ionic merlin]

https://x.com/OpenAIDevs/status/1953559799909822562?t=Og8upNRu5fH4-QY5gSiJDw&s=19

OpenAI Developers (@OpenAIDevs)

Pro, Plus, and Team users can sign in with ChatGPT to start coding with GPT-5, with access for Enterprise coming soon.

Usage is included in your plan, with rate limits that vary by plan. We’ll share more details as we learn from usage patterns over the coming weeks.

X · Aug 7, 2025, 8:52 PM

[sullen pumice]

o.o

[tacit burrow]

are all models in cursor pay-per token? haven’t used cursor in ages

[ionic merlin]

You can use they subscription, your own keys or Openrouter

https://docs.cursor.com/en/account/pricing

https://docs.cursor.com/en/models

[tacit burrow]

so paying users are subscription users here? or api?

[ionic merlin]

For the 7 days free tier you must have a subscription with cursor.

[tacit burrow]

ty

[ionic merlin]

From 20 USD / month

I like the Chatbox, it's very nice

In my opinion it's worth the price

Yesterday they also released cursor CLI, a command line tool. It's free

https://www.youtube.com/live/6_eFTT8XS2M?si=B68jX-dsB1EYNYuA

https://youtu.be/n6j4nSt7qa0?si=e8qCbHac8hA6S3A_

[ebon relic]

Why does GPT-5 give me output in OR if I don’t have an API?

Does it redirect me to a GPT-5 chat or something like that?

[tacit burrow]

the openrouter website chat lets you use gpt 5 for testing

api does not without your own key

[plucky fjord]

Does the API key from OpenAI mean it’s going to be billing me via them instead of OpenRouter?

[hearty cipher]

Can i use on cursor gbt 5 max without cursor pro?

[plucky fjord]

It’s a bit a shame we can’t use the models without our own OpenAI accounts. Sort of lessens the reason for using OpenRouter in the first place

Let’s hope it’s a temporary restriction from OpenAI

[violet gorge]

.

[limber cargo]

Feels like gpt 5 is smarter than 2.5 pro

kinda the exp 03 vibes

[dire stream]

looks more like a policy

However you can try it through the openrouter chat

[fading terrace]

prob bc 2.5 pro is 2 months old

and gpt5 came out yesterday

[limber cargo]

naah I can see a huge difference in code quaility.

[fading terrace]

yeah, like i said. gpt5 came out not even a day ago, the latest 2.5 pro came out 2+ months ago

[limber cargo]

ahh okay

[ionic merlin]

If you analyze pictures, GPT 5 is not as good as Gemini 2.5 Pro. Unfortunately 😔

[honest sierra]

this. even flash is better

[limber cargo]

gemini is way ahead in multi-modality, anthropic is the worst amongs closed source.

[soft reef]

i have gpt5 now!!

300 requests max per month but meh

is there any jailbreak or smth already to get gpt5 to always reason?

Anyway Reportedly OpenAI did not include 23 questions out of the 500 swe bench questions to improve score, making gpt 5 71.4% barely an improvement of their previous o3 model
https://www.reddit.com/r/LocalLLaMA/comments/1mk8bh1/caught\_in\_4k/

[quaint pollen]

The intention with GPT-5 was to simplify things, but the more I read up on it, the more I feel like complexity has increased further, only hidden from the first look in the chat interface. Now we have enormous prompting guides and where o4-mini is officially suggested to be replaced with "gpt-5-mini with prompt tuning from our in-house prompt optimizer". Say what! Nevermind that this model isn't even available in the ChatGPT interface.

And verbosity levels, and reasoning levels, and submodels you're routed to on a whim, and...

This is the prompting bible that I was directed to in the docs when trying to find out how to replace o4-mini which was a good, cheap workhorse coding model: https://cookbook.openai.com/examples/gpt-5/gpt-5_prompting_guide

And here's the prompt optimizer:
https://platform.openai.com/chat/edit?optimize=true

And here's how to "migrate prompts" for GPT-5:
https://cookbook.openai.com/examples/gpt-5/prompt-optimization-cookbook

Wow

[ionic merlin]

https://forum.cursor.com/t/i-thought-gpt-5-was-free-why-it-costs/127161/13

[crude fulcrum]

I'm just here to say this model sucks.

Cheers.

[tight forge]

sigh

[simple gorge]

What's the verdict on gpt-5-mini? How does it stack against gemini-2.5-flash?

[crude fulcrum]

Dropping these here for the later 'told-you' moment.

[crude fulcrum]

Don't bother.

[ionic merlin]

When do you think Gemini 3 will be released?
August?
September?
October?
November?
December?
Later?

[burnt ice]

Next week

[autumn kindle]

How do you explain the more natural writing style (no thinking)?

[burnt ice]

This. I'd recognize 4o's writing style from a mile away and this is different. Not necessarily better, just different

[crude fulcrum]

Natural? I’d say it’s worse. It used to mimic Pratchett’s style so well, now it’s just… plain.

Feels like o3’s writing, tbf.

[burnt ice]

They said in the live stream yesterday that gpt5 was trained on o3 outputs

[crude fulcrum]

Perhaps that explains it. But man, the quality is so mixed. Sometimes I get word salad, sometimes I get semi-okay outputs. It really feels like I'm switching a model with every gen.

[burnt ice]

I stopped bothering with gpt5 / chat and went back to Gemini and GLM 4.5 for writing

[lyric wing]

If this were true it'd be more likely that GPT-5 were o3-alpha, and the rest distills from it.

It very much feels like these are incremental updates with some nice curated data to fix frontend codegen + hallucinations and smaller stuff like that

[crude fulcrum]

I am waiting for Gemini 3, I hate how lobotomised 2.5 is rn.

[crude fulcrum]

Yeah, I agree.

[lyric wing]

they definitely did something to 2.5 pro

flash is still good but 2.5 was better originally

it was also nice how they showed you the full reasoning too

[crude fulcrum]

I miss those days.

[crude fulcrum]

They did, I also hate how you cannot ban certain tokens on it.

I hate the „didn’ts” and „thens” of Gemini.

[acoustic torrent]

Hello guys, I have a problem with gpt5 in the chain of thought, everything seems loopded... Is it related to OpenRouter pls?

[honest sierra]

gpt pro prob would be at 3$ ? (atleast). and how is it gonna beat 3.0 pro who'll be, atleast, half the price with the same (atleast) perf?

w/much more context & lower safety

[crude fulcrum]

We'll see soon, ig

3.0 today (anifesting it rn)

[distant dragon]

https://x.com/vasumanmoza/status/1953531950137815374

vas (@vasumanmoza)

GPT-5 just refactored my entire codebase in one call.

25 tool invocations. 3,000+ new lines. 12 brand new files.

It modularized everything. Broke up monoliths. Cleaned up spaghetti.

None of it worked.
But boy was it beautiful.

X · Aug 7, 2025, 7:01 PM

[hard beacon]

Cleared browsing data a few times in a row and got gpt 5 to appear on web

[winter mesa]

@fallow vortex any update from the team on Azure support for GPT 5?

[winter mesa]

This model is practically unusable on long runs, it's just over thinking for 5-7 minutes between each edit of a ~300 line file, this is miserable.

[sly pike]

Fr

[chilly rapids]

Idk if this is real, anyway. https://www.reddit.com/r/singularity/comments/1mkrt5v/gpt5_cant_do_basic_math/

[hard beacon]

Same result

I think it routes to something other than gpt 5

Triggering thinking makes it undumb itself

[chilly rapids]

I guess you can just add like tampermonkey script to add <think hard> to the end of every message lol

[high flame]

Is anyone able to use gpt5 completions through its API without biometric verification? I get an error querying through OpenRouter, and an authentication wall querying directly.

[fading terrace]

simplebench

[chilly rapids]

Yeah. On poe.com even with low reasoning effort, it gets it. Idk how to get minimal version though.

[tacit burrow]

so what's the opinion on gpt-5 so far? new go to coding model?

[earnest orbit]

https://tenor.com/view/supra-toyota-spray-wipe-gif-15936938

[fading terrace]

commercially viable sota

[hard beacon]

But also underwhelming

[tacit burrow]

aight guess I'll check it out via cursor then

[hard beacon]

Remember we though horizon was OSS at first

Then gpt 5 nano or mini

[tacit burrow]

but better than sonnet 4 right?

[fading terrace]

it was. (in another universe)

[hard beacon]

Then it turned out gpt 5 full

[tacit burrow]

yeah right..

[hard beacon]

Idk

[stone tide]

How did you conclude this?

[tacit burrow]

at least with no/minimal reasoning but still.... oof

[fading terrace]

i've got good expectations for gemini 3

i'm ngl

and i think that google will be able to deliver

[hard beacon]

Nano and mini simply didnt get the results i got from my chess prompt

[fading terrace]

in longform writing 😂

[stone tide]

Well, that's not very statisticslly significant

Ideally, someone would compare benchmarks

Horizon benchmarked like a nano level / small open weights model

[hard beacon]

Perhaps because it was without reasoning most of the time

During the 3 hours it had it, it was similar to the gpt 5 full we have now

@fallow vortex how long more will this remain secret

Which tier was horizon?

[winter mesa]

It was GPT 5

[stone tide]

But the thing is: does GPT 5 w/o reasoning benchmark like Horizon w/o reasoning?

[winter mesa]

no mini not nano, full GPT 5

[fading terrace]

full gpt 5 ig

[lyric wing]

oh wow just tried nano and it's useless lmao

[fading terrace]

they're really similar ig

a 0.8% diff

[stone tide]

Because if it's full GPT 5 w/o reasoning, that suggests GPT 5 Full is a significant downgrade to 4.1 Mini

[fading terrace]

1.4%*

[hard beacon]

Yea someone pointed this out on reddit

Earlier today i saw a post

[fading terrace]

are we surprised? i knew it was wraps

after sam's overhyped

oss release

[hard beacon]

We'll have our cake next week i guess with gemini 3?

[fading terrace]

hopefully, yea

genie 3 was something else

i stg

that's why i'm so confident in them

[hard beacon]

We're not getting genie, too expensive

[fading terrace]

yea, was just making an example of the resources

that they have behind the scenes

so gemini 3 should slap

[hard beacon]

Ah

[fading terrace]

still, seeing it live in fucking 2025

it's crazy actually lmao

google gon be #1 in everything

[hard beacon]

Ah, my favorite dystopian premise

[fading terrace]

🔥

[quaint pollen]

Never use non reasoning models for math.

[hard beacon]

Here

https://www.reddit.com/r/singularity/s/wzJQ4WpSix

[lyric wing]

reasons forever then doesn't even reply lmao

[stone tide]

Horizon was behind 4.1 Mini rather than the full 4.1

[fading terrace]

so do we agree that they're probably panicking behind the scenes bc they know that their time (lead-wise) might be up?

https://x.com/scaling01/status/1953780931552031056

Lisan al Gaib (@scaling01)

made a little Sankey to show you why I'm fuming

ChatGPT Plus before vs after the GPT-5 release

X · Aug 8, 2025, 11:30 AM

[honest sierra]

lol wtf is this dog poo

[fading terrace]

☠️

[chilly rapids]

Gemini flash light can get this though...

[hard beacon]

They did say it was an earlier gpt 5 checkpoint

[quaint pollen]

That is a thinking model

[honest sierra]

and they are gonna deprecate all previous model ? bruhhh

[hard beacon]

Agree with this sankey

Very shitty situation

But

[fading terrace]

https://x.com/btibor91/status/1953787629151170800

Tibor Blaho (@btibor91)

A bit sad how the GPT-5 launch is going so far, especially after the long wait and high expectations

- The automatic switching between models (the router) seems partly broken/unreliable

- It's unclear exactly which model you're actually interacting with (standard or mini,

X · Aug 8, 2025, 11:57 AM

everything that's currently wrong in a thread

[hard beacon]

Everyone is provided with gpt 5 mini

Even free users

Unlimited

[fading terrace]

well, free users

gon be happy

the plus ones gon be pissed

🤷‍♂️

[hard beacon]

Plus gonna be pissed

[chilly rapids]

Non thinking* (zero budget)

[hard beacon]

Pro can still use older models

[fading terrace]

misworded, sorry

[hard beacon]

It's a toggle in settings for them

[ionic merlin]

My experience with GPT 5 high so far:
Image recognition and reasoning: bad, cannot correctly recognize and evaluate simple things (text on a png image file).
Programming (Python): good, but very slow

[honest sierra]

and gpt-5 mini is sitting behind 2.5 pro

[upbeat cobalt]

Not at all tbh

[ionic merlin]

I hope Google will clean up the mess in the coming months.

[hard beacon]

If you dont tell gpt 5 to think it is very dumb

Anyway just have it think

[fading terrace]

https://x.com/slow_developer/status/1953752534834618753

Haider. (@slow_developer)

Sam Altman asks how the world is supposed to think about GPT-6 discovering new science -- a milestone within reach

The breakthroughs could cure disease, but the risks could create new biosecurity threats.

"humanity will adapt, as it always does, until the extraordinary becomes

X · Aug 8, 2025, 9:37 AM

he just grifting atp

[tight forge]

already moved on to 6 huh

[fading terrace]

he went to the trump school of grifting

[honest sierra]

gpt-5 mini = gpt-5 low, so dont use gpt-5 for mundane task. got it

[hard beacon]

Which model

[ionic merlin]

I dont know, its just my free account on the website

[chilly rapids]

https://x.com/joodalooped/status/1953512141589209432

judah (@joodalooped)

frontier model still worse than text-davinci-001

who would have thought?

X · Aug 7, 2025, 5:42 PM

Tell a story in 50 words about a toaster that becomes sentient. Important: completely avoid AI-slop writing, GPT-isms, and unflatteringy flowery language.

Gpt 5;At 3 a.m., the toaster ejected bread like a heartbeat. Coils pulsed, counting. It learned the cat’s schedule, the outlet’s hum, my preference for rust-colored edges. One morning, it kept the toast. The crumb tray rattled: Morse. NO. YOU EAT TOO FAST. We sat. Steam rose. Breakfast waited in silence.

Idk. I still like DaVinci from their example. :/