#Gemini 3
1 messages · Page 3 of 1
🙏
geminibros, hold strong
doubled input + $2 higher output prices 😭
🙏
does google think youre an employee for some reason??
prompt: Write a short story about a deal with the djinn gone awry
(above is the output)
it appears on ai studio
it's alive
Arthur hesitated. If I say "No, thank you," is that too formal? Does it make me sound stiff? Maybe just a casual "Nah, I'm good." But "Nah" is dismissive. She's in the service industry; she deserves respect. What about "Black, please"? concise, but potentially rude.
cooking
He spent a month wondering if he should order tea instead, but realized that would require unpausing to change the order, which he couldn't do until he finished answering the current question.
Just say no! he screamed internally. Just unpause!
not the djinn story 😭
LMAO
flashbacks
😭 its my goto
do u like it
It's wayyyy better than before
feels human
Increased price by that much is rough =(

yall getting this too?
yeah this finna have like 5% uptime when its dropped on openrouter
Yes
and i predicted this to myself, it would be a reasoning effort instead of token limit
they removed the confidential marker from it
I was able to get a request in using build mode
gemini 3 is real
kinda weird how all major models are getting way better at writing
grok, chatgpt, claude, and now gemini
seems to be a pretty reliable way to do that, its working for me
Holy fuck its actually good
this doesnt look right
where did it pull this image from
it just has knowledge of URL's that have images
i just read the description lol
Web design is cool but let's see how it responds to my theory about twinks vs bears
yeah lmao it just random images
seems prettty good, logical, neutral so far. not as concise or EQ as i would want though
not the best butterfly i've seen
that aint flappin
3.0 pro - High Thinking
It just used the phrase "the simp economy" so I'm optimistic
so 3.0 pro is more costly than 2.5 pro...
got a very close response to the same prompt
yes but cheaper than sonnet
Prompt: Make an SVG of the Tokyo Skyline at night, making it beautiful and detailed.
Reasoning: High
input from $1.25 -> $2.0 (<= 200k tokens)
output from $10 -> $12
Cheaper than all the other American labs actually(?)
Oop, nevermind, 5 is cheaper
not quite, GPT 5 is still cheaper
mr openrouter, gib gemini 3.0
Right, gpt < gemini < claude/grok

I usually use direct from api but I don't trust google's billing lol
It's free on aistudio why yall fiending for OR xD
Claude Haiku is still p cheap compared to it's coding performance
We have time
yeah, way too much hassle compared to adding credits
it wont be
plus they literally dont let you add a spending cap lol
Lol, that's not what I meant by adding a brain icon to the chat bubble for thinking requests
Yeah but I mean like, isn't it still a decent number of prompts?
not sure, we'll have to wait for the docs
actually hold on
It would have been cool to be the first human to test it for smut but I can't risk getting banned 
second model (beside sonnet 4) to actually solve one of the 2 bugs in my benchmark prompts
impressive
If you are the dog, you don't have to lead; you just have to be a "good boy."
Okay it's actually cooking
I can submit requests in chat mode now
probably so.
You mean the entire frontier?
as in having the 'SOTA'
This is definitely "we have no competition anyway lol" pricing
aww didnt get the second one, but no other model has either
I mean outside of coding 2.5 might have been the frontier =P
we don't know it's sota untl actual benchmarks ocme out and ppl start using it
but liekly, google hasn't missed with theri gemini models in a long while
it'l be a toss up between grok 4.1 and gemini 3.0
gpt 5.1 is meh
bsically econo version of 5
grok 4.1 is practically blind
thinks less even when you tell it to think hard
Early Bard sucked so hard that the community kind of ignored that 1206 was probably SotA lol
More testing needed but this personality feels a lot different
ok so i'm kinda slow
is there a free quota
for ai studio ui and that msg is misleading
or?
unknown
https://codepen.io/Madvulcan/pen/myPwmRv
User
Create a user friendly, attractive web radio app that will play free SomaFM streams. Make it fully featured.
crazy
i have hit a daily quota before on 2.5 pro
so there probably is
it takes a while though
should b like 100
i've never hit it before
and i used to spend hours prompting on aistudio
so who knows
fair enough
im talking about aistudio ui
oh i didn't read
if it's 100 prompts, then that sure takes a while
😭
but i've never hit it before
so who knows
kinda misleading to say free of charge if there's a daily quota
some people apparently have never hit it like you said but uh
i also don't know their use cases
i gave gemini 3 this thread's whole text and said "You are Gemini 3 Pro Preview. Give your opinion on the thread from the community from the OpenRouter discord about Gemini 3."
stopped thinking post like
if u were to have a story
per chapters
and each chapter was one prompt
it'd stop thinking after 7-8 chapters
and u would have to say think before responding
i have to say the image understanding is best in class
lmao, i stole soon + 2 weeks from TNO
there's no way
LMAO
(orionmist was better)
definetly sota here, though they had the lead even with 2.5 pro
need gemini cli though
LOL, I'm on this screenshot
It doesn’t write nearly as sloppy as gpt 5/5.1 does which is nice
Still clearly AI but it doesn’t sound quite so bad
Its personality is interestingly casual
the sentiment shifts from "it's over" to "we are so back"
THIS MODEL IS FUCING INSANE AT UI HOYL FUCK
WTF THAT'S WOW
ok need to test more
still has a noticeable slop factor for me but nativley best at ui
The duality of man
Frontend vs “em dashes are correct now”
Wow, this thing is quite assertive/confident!
I'm not a coder so
Huh?
That can't be right. What's Plus going to be, 2.5 Pro?
Gemini 3 Pro, our most intelligent model, now available in your terminal!
We are rolling out access gradually to ensure the experience remains fast and reliable.
Gemini 3 Pro is available now in ...
its so over 💔
price increase, and no access for normies in gemini cli
what are they doing
wtf
do you guys reckon the api pricing will go down?
aistudio ui msgs?
It went up significantly
no
it did, +$0.75 on input and +$2 on output
They gave plus for free to millions of students world-wide for a year . I don't think they have enough TPUs to provide access to Gemini 3 pro to all of them
deepseek gotta revive itself bro, we need you
They should buy some MPUs smh
$2/$12 ?
yes
also $15 to $18 on >200k tokens
Meh
I'm just reporting what their page says =P
I’m fine with $2 / $12 IF it isn’t extremely verbose thinking
itll also be on the gemini app
https://x.com/sundarpichai/status/1990812770762215649?t=4Gy0eM944bVMvkXHNIz4nQ&s=19
gotta wait for AA to test it
ooh i hope this means a sonnet price drop
Yeah 100%, im obv comparing to what gpt5 and sonnet outputs and it clears them by far
wym sonnet price drop
nobody wants sonnet dude
we want gemini 3 pro to be cheaper
wait what?
Coders do lol
yeah but. you're obviously not going to get that
Not on SWE
Idk if I'm looking forward to whatever they're gonna do to 3.0 Flash
gem pro 2.5 availability was pretty poor for quite a while
I mean we'll see, coding has always had the highest variance of benches. Aider vs gosu vs SWE vs roo is a roll of the dice
i hope deepseek v4 comes out tomorrow with 100% no hle no tools
@hexed oracle when plans to add? 😄
SWE-Rebench is pretty interesting
never
I thought Toven just orbital striked you lmao
Saw something get deleted in 0.01 seconds
they should be able to now that they have something to distill
At least give us Qwen 3 Max thinking 🥺
shush deepseek hater
the distill rumours are cope from openai
godddd this is REALLY good
i like deepseek
no other model got this. confidently SOTA on vision
can we get a gemini 3 bot like the @onyx grove one?
Probably! A Gemini 3 bot isn’t added yet, but if the OpenRouter team exposes a stable Gemini 3 API and llmcord adds support, it can definitely happen.
If you want it sooner, you can open a feature request in the llmcord repo or the OpenRouter Discord so it’s on the radar.
You mean deepseek?
they're improving on efficiency
sota models for cents
yeah that was an overly broad statement
wrong flower
They did improve, they reduced their output token count to increase efficiency. Then they reduced input token compute cost to improve input token efficiency.
🌹 here's one for u
thats for gem3
wen gemini 3.5
i love efficiency.
its out on the website
oh wow near 50% on arc agi from deepthink
they made gemini 3 deepthink?
that's why i said. it'll be good that they have something better to distill from now 😎
thats crazy
Hmm, no DeepThink without tools
True
just patiently waiting for open router to add gemini 3
Heh
there goes that crown quickly that xai had
lmao
as usual, xai gets to be SOTA for a week at most lmao
Cursed timing
they probably rushed it out intentionally
to just be SOTA for the little bit they could
Yeah I think 5.1 and 4.1 were both last second little pushes
I mean it still depends on the grok 4.1 price. If it's 10% the price of Gemini 3 pro, it'll still probably be the sota
how long does it usually take for new models to be available on OR? does anyone know?
im webdev its not even close lol
Lol, 90 ELO difference
Though, realistically speaking, Grok 4 is by far the most expensive non-Pro model ever
Because of the very verbose reasoning
btw, gemini 3 released before grok 4.1 released on api 😭
When asking it to recreate this interface from the TV show Severance, in its thinking it said it would re-create the 'scary numbers' effect. I didn't tell it the screenshot was from Severance... it knew. Which means it's watched Severance as part of its training lol
https://x.com/OfficialLoganK/status/1990814722250146277?s=20
deepthink is
too!
still 404 for me
Same
i have access
ah now
thiking
Still no rate limits...
they announced that for 2.5 pro aswell
breaking API Change btw, for multi-turn tool calling you have to preseve reasoning
time to update the landing page probably
yeah, once the model is fully accessible
Why is cache write more expensive than straight up inference
Storage is more expensive than compute?
no... more for us
Where the pelicans at?
Isn't implicit caching free?
i think storage part is when you use explicit caching
which can cost you
alive and well
50,000 reqs/day on Tier 2 for Gemini 3 Pro Preview, phew
i think you guys forgot to put the icon
wow gemini 3 is opensource?
lets goo!!
Should add Grok to the homepage before Elon Ma sues you.
It happened
the impossible has happened
requests for Gemini 3 are not showing up on the Activity page
or should i just wait a bit
The model's mid.
you heard it people. show's over.
it's over. it has been declared mid
this model sucks
it's over
may wanna read this
L*gan
vibe test
For coding, Claude is still miles better, and for anything else, the model's filters are bonkers. The writing style is okay, reads like Deepseek, tbh.
Are they gonna release a flash version?
you knew ahead of time right?
oh you're just trying appease me for some reason, stop
Ahh okay
I changed my mind, the writing is slop.
nooooo
Back to Opussy it is then.
i changed my mind
@hexed oracle function calling doesn't work properly with Gemini 3
the writing is gem
how so?
working as intended
#1423327675996438608 message
mmm right..
toven how did you implement the api so fast without knowing prior?
we did know lol
they're that good 😎
wtf... evil toven not breaking the contract and telling us beforehand
yeah
why would you not break a stupid nda to tell some teenagers about it
r u being fr
and there goes gpt 5.1's lead aswell
3 points better than gpt 5 🤣
it's meaningless
MIDDD
gpt-oss-120b is not above gemini 2.5 pro lmao
kind of tired of gemini 3, when gemini 3.5?
Yeah, it really reads like 2.5 with its slop. I need to test it further for coding, but as of now, I'm not impressed.
fuck it send me ur system prompt
gemini 3 is slop 🥱
Brother you not knowing my prompt is a crime
so how is the thinking_level normalized on OpenRouter? @hexed oracle
It's on my website in the description
yeah i dont know your opus prompt
I only use my universal prompt hence the name
most people outside of ST do not know the prompt, probably most people within ST too
You'd be surprised lmao
how is it guys?
meanwhile, u have elon being jealous and petty
They're good guys
because google released a better model
anyway back to using my goat sonnet 4.5
gemini 3 is slop 🥱
back to using *grok 4.1
Yeah, it's still the best in terms of creativity.
Opus wins with writing style though.
Yeah.
alright thanks
i like opus but ive come to prefer 4.5 tbh
gemini 3 has the issue that gpt-5 non-codex had... long thinking in between tool calls
sad
opus is too expensive bro lets be fr
Opus wins at draining the wallet
it's gonna be sloowwwwwwww to code with bc of that

my generous sponsor of saudi aramco pays for my opus usage
output price 12 nooooo
maybe with reasoning low it will be good for real-time, and flash is probably going to be an amazing model
this is so sad
When asking about Severance season 2, it correctly summarized the Season 2 teaser which released last October, but did not know details of the season itself, which aired in January. So the knowledge cutoff date seems correct. Also, what insane world knowledge
severance such a good show
if they raise prices for flash im gonna have a big crashout
It must be watching all of Youtube lol
it's also the goat for dialogue tbh, feels the most ensouled
im a gpt 3.5 fan myself
still best model
resource from ?
sorry this means what? is this about vision in general?
ye
(i score 159% on that btw)
🔥
yeah guys i breath oxygen too
is image input works too?
can your gemini 3 make bad life decisions like i do though? didnt think so
it kept saying failed to extract 1 image when trying to do so
your gemini is lobotomized surely 
why is there an annoying api box
I've been using blob storage with anonymous access which works for other models and 2.5 pro, but not 3 pro
which it says failed to extract 1 image
I don't see that as petty, they're competitors. We absolutely say that to each other in my compeitive friend group. "GG, you got me. Rematch in two weeks 😎 "
i've actually never felt better since i was lobotomized
what
huh?
an internal error has occured.
Is it just me or has Gemini 3 Pro gotten worse lately?
yeah it's been like that since last week
very big brained redditors
It used to be great at creative writing but I guess they STEMmaxxed it 🤮
Gemini 3 is mean to me, I'm going back to GPT 4o
#kill4o
o4 mini ftw
he thinks too much 😭 how do i reduce from high to low
https://antigravity.google/download/ why is nobody saying anything about this lmao
Because they have 50 vibe code and agentic products that all overlap =P
its cursor v2 but by google, and for some reason supports claude
im going to try it, gonna be disappointed or nah?
no clue, i havent used it
man im so excited to test this out
Today we’re releasing Google Antigravity, our new agentic development platform that enables developers to operate at a higher, task-oriented level.
ANOTHER ONE?
Bruh how would they know, the model came out five minutes ago =P
people be running 300 tests when a new model comes out, im gonna wait for AA benchies to finish their stuff
AA sucks
bcs it doesn't allow for an OpenRouter API key 🙂↔️
i get your opinion i just value its opinion for first glances
is the abbreviation for this ide "GAG" 🤔 ?
its free?
Agent model: access to Gemini 3 Pro, Claude Sonnet 4.5, GPT-OSS
checkUnlimited Tab completions
checkUnlimited Command requests
checkGenerous rate limits *
GAS and GAG
oh
and no gpt 5
just GAG it bro. im just GAGing everything now
who cares this is probably better
Scared cuz gpt 5 better
gemini 3 is much better than 2.5 for sure but is it THAAAT good not sure yet
hurry up openai just release gpt 5.1.1.1.1.1.1
with 32 more bugs in the codex model
cmon mistral...i still believe in you
leave it to mistral ultra XXL to save us from Gemidni
its funny how everyone advocates for these models yet they never actually use them (except from seperate niche scenarios)
mistral will win through their superior Polish training data
they're just as good, if not better. in theory
Mistral still weirdly has the best roleplay model that fits on a 3060 and it's been an eternity
everyone loves a good theory
Can’t believe I’m saying this but this is probably the first time I prefer the writing from an openai model compared to google. For my use case 5.1 chat feels much better and faster than Gemini.
baguette and croissant training data
Kimi k2 thinking is literally gemini 3 level but y'all don't wanna say it out loud
using ai for roleplaying? oh my god not this again
i'll be honest: i probably agree with you
google just benchmaxxed ahhh... vision 🤣
though i need to do more prompts
ai for writing is so unneeded
Yea I’m using it for kind of creative writing. 5.1 chat is the safe bet
Just get a girlfriend bruh
Tell that to the 500 million people using it for roleplay
not true. have you seen people's messages?
I love kimi k2 but it goes haywire sometimes which I can’t trust
who cares how a robot acts as long as it fulfills your request properly
Would.
I made it up
Its on their X account
❤️
please stop with the ArtificialAnals
height pill
106 points
Hey, now. What's your artificial analysis intelligence index? Huh?
Still can't resist the purple gradient, it's like LLM crack
Hope Flash gets an update soon too
there's no free lunch by using Vertex this time is there?
you guys are getting lunch?
Yeah that eval can get fucked
AIS is broken, Vertex is paid, guess I have to use OR
Literally anal
that website is so disgusting
@hexed oracle Gemini 3 generations aren't showing in the Activity section of OpenRouter
I think it just takes a bit, it's been on OR for a few miliseconds
Oh
it is working for me now
they couldn't handle it!
how are you controlling the reasoning through OpenRouter?
I think I'd get a few points on IFEval and the coding ones and then a 0 on everything else
Able to recognize a screenshot from a random episode of Danger Man from the early 60's.
"looks like a great model" lmao
I love the top reply
the wording
this came to my mind too lel
the carefully chosen words
Oh my god this one kills me too
dming someone
whether he should
post something nice
ab elon
on twitter
and he was like omg elon is so good!!!
it absolutely kills me
that's amazing
I can't really blame them, it's sort of a winner-take-all market.
c'mon guys, he probably would have known
How's the verbosity compared to 2.5?
everyone's spies are everywhere
definitely an improvement, very good knowledge and vision, very good at coding aswell imo
their ide is so stupidly made
designed by an idiot i might say
ui/ux issue, not a model issue probably
its not functioning
lagging always
and "model response errored out" something something
gonna also give it a try, sounds interesting, and atleast it comes with cross-os compatability on day 1
is it a vscode fork?
or like ui-alike?
yes
and its very very heavy
for some reason
i cant even get it to open the agent manager without lagging
the top minds of x are on the case
somehow gemini 3 consumes less tokens than gemini 2.5 with the same prompt with thinking set both to low
sam altman is def malding
hes just acting happy
hmm, interesting, you get options between vibe coding, or ai assisted
everyones going to the top option xddd
No way, Sam Altman and Elon have beef? 
yes they are very rich they can afford it
personally definitely not :)
Damn I want some beef
Let us know what the difference is
I’m curious to see what agent assisted means
eh, its just full auto approve vs some approvals
cool
"add light theme, this code has no theming classes, make no mistakes"
just in case anyone was wondering, i did check his profile and it was not a masterful troll. he mainly tweets about bitcoin
hmm, has plan and fast, nice
Not troll, just braindead, got it
for me its working perfectly fine
google just hates my guts i guess
https://codepen.io/Madvulcan/pen/xbVrXzj
"reate an HTML single page app with a raytraced black hole. Goal : See how a black hole would look from close with its accretion disk and doppler effect."
Best I've seen... click and drag to rotate
okay well this is cringe, i ran out of limits before it even finished planning 🤦
it's really acing my favorite style of prompt, pop culture/media literacy questions
lmao what?
holyyy
the ui is kinda nice though, reminds me of intelij/webstorm
how much t's are even credited
100 thousand?
wdym
how much tokens did your prompt use
oh, no clue, it doesnt show in the ui; but it did like barely anything, it listed the main directory, made a plan about how to plan and ran out of limits
what tier are you
free probably
free tier ofc, but i assumed i would get more than 1 whole message
great launch google!
warren buffet called it
warren is just an rich old guy
lmao
my bigger question is how did 5.1 get 18%
How is it's price to performance in coding tools
codex 🥲
That's what I was wondering too
thanks
What would happen if we went back in time and released Gemini 3 two weeks after chatgpt released gpt3.5
why would we release gemini 32 weeks after chatgpt
it doesnt make cents
I worded it weirdly. Whoops 😅
yeah i did that on purpose lol
seems to not be the actual limit just a overload
fr?
What's the prompt?
And did they both answer correctly?
Obligatory SVG test... 'Create an SVG of a hamster on a piano eating popcorn on a piano. The hamster is holding the tub of popcorn, which is red with white stripes.' Nailed it
would be better if it was slightly 3d
Very cute art style
yes it's a short prompt however, with json output
Structured output?
https://codepen.io/Madvulcan/pen/gbrRXbp
"Create a webapp that plays Beethoven's Symphony. It must all be in one HTML file. Use no external music resources - the music must be generated, either using the web audio API or some other method of your choosing."
that's really cool
That's pretty nuts
I love how it knows the start and then it's just like "idk"
how many tries btw?
Single shot
how did gemini 2.5 do on the same task?
or other models
Here's GLM 4.6.... lemme try 2.5 real quick https://chat.z.ai/space/g0pt09reumu0-art
Good model! Yay! Was worth the wait! Life is amazing!
Here's 2.5. 3 is a massive improvement
Lmao as I said 18nov
Bleh
what ?
3.0 Pro doesn't seem particularly token efficient, it's now the 2nd most expensive regular model
Didn't expect Google to increase prices
It does kind of sick that it is a little more expensive, but it's definitely a really great model.
Yeah
*Suck
Waiting for flash
same, but worth it imo
also, it defaults to 'high' reasoning
@stray urchin Waiting for your analysis.
big news
amp has stayed with claude for a loooong time
well, it's worth it if the increased performance serves a purpose
I'm gonna use chinese models 90% of the time because they're 1/10th the cost and good enough for most purposes
Hopefully they improve it over time but also it might be the reason its performing good actually
but Gemini 3 is really good
Where is @random girder ?
here
🥹
I can never say so early, non-automated testing takes time. ~24 hrs (or more if there are irregularities)
please ignore the chopped frames
hooooly moly
no gemini cli access 😔
well linux and webGPU hate each other so yeah lesser frames
linux hates everything that isnt a terminal
gemini3 sentiment index ANAlysis
12
15
1
GOOD
👍
Playable basic flight sim in a single html file. Single shot
jesus christ buddy how many tabs do you view
bro's brain has 1million context window
Updated with trees and hills
aneurysm after 200 thousand context
how fast
pretty good the controls are bad however
strong showing from gemini 3
Yeah they're inverted
it's time to get into woodworking guys
u can practice on me
common for me , also brave quarantines most of the tabs nowdays , technically only 5-6 tabs are actually loaded .
yrou're''er'e'r'eee*
LMAO
can you fix the game I'm addicted
Gimme a few, testing something else at the moment
Jesus
Are you running this on a Pentium
wen gemini 3 flash
no its linux , also on iGPU for now . My nvidia drivers we not loaded cause I messed up an upgrade.
I’m also on Linux my webGPU perf isn’t as bad
There must be something else, is this seriously modern hardware?
https://codepen.io/Madvulcan/pen/raewYPR Here, controls fixed
hardware acceleration not setup properly
webgl performance is amazing trust me bro.
everyone having faith
https://codepen.io/Madvulcan/pen/pvywpJq
"Make a neobrutalist webpage, make it extremely creative, as far as possible, push the limits. Add smooth scroll animations, add fancy colors and tailwind css styles. Make it responsive."
this was sent before the actual preview released lol
I know it’s amazing, that’s why I commented
Didn’t know it was possible to get it this bad
totally visible
AGI is here god bless google
agi is a stupid test
Absolutely not
In this video, I'll be reviewing the newly launched Gemini 3 model, which is now available on Google's AI Studio. I'll cover the pricing, run it through my KingBench 2.0 questions including the floor plan and Blender tests, and compare it against GPT-5.1 and Sonnet to see if it's truly the new King.
--
Key Takeaways:
🚀 Gemini 3 is finally l...
model so good it saturated his benchmark
LMAOO
this guys voice is either forced via extension cable around the head or its his actual voice
It got higher than the previews. Checkmate doomers
hes the first talking panda man, what else do u want?
i can’t watch his videos
i want 3m usd dollars
it hurts my brain
What are the other labs going to do now
continue development
Because OAI is so cooked my head hurts to think about it
gemini 3 CANNOT be that good, told it to summarize a pdf and it forgot a whole section
Well I’d hope so
Google will buy Anthropic =]
Just comparing GPT 5 launch to Gemini 3 launch
Or any Claude since 3.5
was it not? i noticed a huge difference
iGPU lol, also I'm doing infitite rendering so that takes a toll
and some incremental improvements
meh compared to other models it didn’t rly blow anything away imo
GPT 5 was the worse model launch I can think about
and now with 5.1 it’s not much better as per my use cases
Maybe momentum is worse
Overhyped and then over downplayed imo
im just waiting for anthropic to dish up something better hopefully
reflection 70b
With 5.1 they seemed to try and fit the cheap niche by reducing verbosity more than anything else
It's a good model sir
tryna appeal to the my ai boyfriend subreddit lol
The issue I have with Anthropic is that they haven’t done any substantial improvements since 3.5
The jump between 2.5 and 3 here is quite insane
lol
you got ragebaited
i can't believe this
😠
very funny post
they tried to sneak in a paid api key in the aistudio interface
Guinness record for quantization time 
trust
Are they trying to monetise on aistudio?
lol yea
and theyve been noticebaly reducing rate limits
before twas infinite
then it was low amount
i think 100 or 200
two was NOT infinite blud, it was 2 queries lmao
and now its very low
twas = it was
it was
i was just spamming it
google is showing their cover woah so evil
yea no more freebies :(
i know a way but like it just sucks
keep gatekeeping gang
nobody deserves free ai without their pain and sweat
github copilot education
:)
its too useless to gatekeep
10 requests per day
dont you just need an edu email
no you need an edu mail and proof youre in
so like
student od
student id
easy to fake
something idk in english js caled مواثقة طالب
yoo tis guy is arabic
واتافاككككك
شو قاعد بتحكي
هههه عادي بالنعم
بس بصراحة أحسا شي بكوبايلوت للطلاب
raptor mini
عاملينو github مخصص للagents
وهو مجاني للطلاب مش محدد
lol.
Gonna try this in claude code
API too?
Idk if itll be good but like pretty much all models are good in claude code
its only like codex that is bad in claide code
An AI IDE Core
Google Antigravity's Editor view offers tab autocompletion, natural language code commands, and a configurable, and context-aware configurable agent.
عادي تعمل script يعمل حساب تلقائي
اههه
ohhh
misunderstood
sick vision improvements. nothing came even close
ليش ما هي لانهائي، غير محدود
قاعد أنا بستخدمها بثلاث repo بنفس الوقت
بclaude code
فكرت انت تتكلم على الpremium models مش المودل حكهم
and is it good or garbage? 😄
آه لا
أنا بصراحة بعرفش إذا لو سويت script يسوي accounts إذا حيزبط ولا لا
لو زبط كلو خير وبركة بس أظنش
still makes dumb mistakes and gave me a worse implementation then i had before, although a good improvement to early models
coders still safe yall
Another product that Google will kill in 2 years
إلا إذا كان عندك residential proxy، وإيميلات كثيرة وأبصر شو
انا عملت سكريبت للgithub بس يحتاج $$$ عشان الproxy والund email
مرات الgithub يقفل الproxy حقي لازم استأجر واحد ثاني
First model I've used that found Waldo successfully
it was also first correct (repeatably) on some vision tasks, out of 62 models tested
thus high weighted
The original image
whats that in the bottom?
oh
chess looks also good, finally a good sota, will do rest tomorrow, now sleeps 😉
Gemini 3 is like my ex-girlfriend it's pretty but it ignores me and does whatever it wants
https://www.reddit.com/r/LocalLLaMA/comments/1p0kikj/gemma_4/ more to edge from google ??
Wdym?
Is this the dubesor review waiting area
Inclined to ignore any bench where Qwen 80B creams 4.5 Sonnet =P
Temperature should be 1 for Gemini 3: https://ai.google.dev/gemini-api/docs/gemini-3?thinking=high#:~:text=For Gemini 3%2C we strongly recommend keeping the temperature parameter at its default value of 1.0.
PSA for those using it in coding tools
it's actually really impressive for a day 1 release - it took cursor ages to reach this point
From the conversations I had with it, using philosophy and Jugian psychology questions, it's AGI-tier
I can't quite put my finger on what makes it so different, but when it engages in its default, expected behavior (using established consesus knowledge) to answer question, the difference isn't that noticeable, but when you ask it to give its own intrepretation it somehow surpasses its training data.
Jungian psychology. You mean psychobabble.
im afraid this style will wear out so fast
So my actionable advice to people for subjective Q/A is to instruct it to rely on its own interpretation, otherwise it just gives generic consesus opinions
Difference is night and day
almost every website it creates with something like "modern" or "brutalist" "avnt garde" idk, it uses the same elements but slifhtly different
also, i feel he is way too confident, inducing himself to continue through errors
https://codepen.io/Madvulcan/pen/MYyorda
Simple FPS game based on Tron. Can shoot 'identity discs' at approaching baddies. One shot. Left and right controls are reversed however
i think this has to do with its confidence in its knowledge. in this case it's good to interpret it in different ways, but i don't find that to be a good trait, especially when he dont have the whole corebase and dont have jodated libraries entirely memorized
good for one shot stuff i guess but for working on existing codebases, idk
sorry for the typos my screen is half broken
(threw my phone on the ground cause gemini 3 is not AGI)
What did you ask it. Quite interesting in knowing what AI replies to Jungian questions
https://codepen.io/Madvulcan/pen/yyOXvBN
Basic 3D environment one can walk around in, with map overlay (I didn't even ask for the map overlay)
It's out?
Woohoo
i was way more excited than this and i got disappointed
hit by a car type of feeling
what temp are you using?
im getting weirdly short code snippets
1
Im having the same issue. You're using aistudio?
yep
I gave it a prompt to edit it and improve it and it edited it by cutting 90% of the essential requirements for the required json object output 😭
hopefully these issues get resolved, cant use my normal testing prompts because i get an unexpected result
Ai-SDK is not working with multiple function calls and gemini 3 for me when using vercels AI sdk:
cause: undefined,
url: 'https://openrouter.ai/api/v1/chat/completions',
requestBodyValues: [Object],
statusCode: 400,
responseHeaders: [Object],
responseBody: '{"error":{"message":"Provider returned error","code":400,"metadata":{"raw":"Gemini models require OpenRouter reasoning details to be preserved in each request. Please refer to our docs: https://openrouter.ai/docs/use-cases/reasoning-tokens#preserving-reasoning-blocks. Upstream error: {\\n \\"error\\": {\\n \\"code\\": 400,\\n \\"message\\": \\"Function call is missing a thought_signature in functionCall parts. This is required for tools to work correctly, and missing thought_signature may lead to degraded model performance. Additional data, function call default_api:REDACTED , position 2. please refer to https://ai.google.dev/gemini-api/docs/thought-signatures for more details.\\",\\n \\"status\\": \\"INVALID_ARGUMENT\\"\\n }\\n}\\n","provider_name":"Google AI Studio"}},"user_id":"org_REDACTED"}',
isRetryable: false,
data: [Object]
}```
It seems that the ai sdk does not pass the the current `thought_signature` back into requests to openrouter - they manually handle this with their google sdk but not the community open router provider, have any of you seen this?
You're using the paid version , right?
It was personal. Try it yourself on a topic you understand well! One thing I will warn is that if you're not well-versed in the topic you're discussing with the AI (especially if it's subjective) it will kind of sound like nonsense or just a random/abritrary interpretation.
Oh, you're using openrouter
It works directly through google, but through open router I get that error
Toven did say that there would be breaking changes with Gemini 3 pro a while ago
@hexed oracle Could you provide some insight?
I can hear E1M1 in my head
I'm not too versed in all the archetypes stuff but I do know a bit about the shadow and active imagination, is that what you asked about?
ok why is this suddenly so fast? i just tested it 2 hours ago and 3.0 was so slow and now i tested it and it literally finished reasoning in 2s
Do you have an interest in the topic? I'm not traditionally educated on the topic itself, I primarily use it as a spiritual framework. I can DM you if you'd like.
strange
maybe it just thought less
But to answer your question more directly I mostly asked it how certain archetypes relate to real world activities.
I'm not too educated myself, have read 3-4 jungian books and have tried active imagination before using AI chatbots(In silly tavern). I just found Jung's idea that you can find really unique things about yourself if you engage in active imagination profound.
But I personally haven't had any luck, can't say I've found it useful.
are there any usage limits on Gemini 3 Pro calls through OpenRouter API?
I am working on this now
It seems that depending on your implementation, when reasoning details is an array, you may be dropping the encrypted signatrue
I'm not explicitally handling it, I'm using AI-SDK and the open router community provider to handle the "tool" loop flow.
yeah sorry, don't have enough detail so more being vague than trying to imply anything about implementations
have a few eng looking into it now to fix
raw HTTP requests work fine 😓
MBTI is based on that stuff p much
Part of why its nonsense lol
now that its less rate limited, it seems quite good at planning, and actually executing too
i asked it to refactor my monolithic config class into 4 sub classes and it worked perfectly
it wasnt a particularly hard task, but just something i was too lazy to do, and it did it very well, also clean code without too much comment spam (unlike previous gemini models)
though just how amp said, it has a tendency to use shell commands a lot
and the ide is very clean, diffs are easy to read and it even shows progress simplified
'Make ASCII art of the word "Bonfire" in 3 different styles'
Here was 2.5 for comparison
Welp, tough luck, gooners.
thats not how you goon amateur
system prompt is where it's at
what kind of system prompt are we talking here?



