vibe-coders

[astral sorrel]

Is the new gemini deep think available for Antigravity?

[round solstice]

I haven't seen it as a pop-up. Not sure how it'd be much more useful than planning mode already is with g3 pro

[olive kernel]

guys, gemini are kinda trash and slow

i mean, for programming tasks

[round solstice]

which gemini?

and paid or unpaid

[olive kernel]

gemini 3 pro high

[olive kernel]

paid, via antigravity

[graceful totem]

For pure “vibe coding,” yeah — it might feel underwhelming.
But with clear instructions and proper constraints, it handles 99% of programming tasks fine.

I don’t think the model is trash. Most of the time it’s a prompting or problem-framing issue.

[olive kernel]

probably the antigravity issue

[graceful totem]

Gemini historically hasn’t been as strong in tool-calling compared to some others, and since Antigravity is heavily MCP + LLM driven, that gap becomes more visible.

[limber arch]

still smarter than deepthink also it is a scam, deepthink is purely a benchmarkl model consumers will never use, same with the current models we are given access to, they are just 4b or 12b or 27b variants of the real thing which can be noticed in their flawed instruction following limit on long outputs and inputs and so on.

Don't compare a benchmark model (theoretical offer) vs the real offer the poopie 4b and 12b and 27b models we get access to in gemini chat and google anti gravity (gemini CLI coding plans), what we get is utter trash

[limber arch]

it does not and i can show you malformed inputs and deep hallucinations because gemini has strong hardset constraints on output to the point where it hallucinates

literally it was lying to itself to stop outputting and generating

feels like requests go through some mcp server that alters inputs and limits outputs

i don't think the model is stupid because even a 2b model from lfm coul respect and does respect that request they just created some mental disability layer in the backend for all requests

the output was so malformed it corrupted the chats

cannot open them anymore

and also google aggressive rate limit which feels like a scam on the pro plan (1 week or2 weeks cooldown on oss or claude after 2-6 prompts then i get like 50 prompts with no rate limit? Same with the pro models there must be something wrong at least let us know the real rate limits in real time and don't make us use predictive horoscopehoroscope guessing to know when we will get rate limited AND OUR CHATS CORRUPTED BEAUSE WHEN RATE LIMITING DURING CONVERSION AND GENERATION YOU CORRUPT OUR DAMN AGENT CHATS)

unironically how tf did google turn into the king of ai slop?

BUT

one thing i commend them for is the translation capabilities of their models they are peak AND AMAZING huge respect for that( that is the only thing they are useful to, otherwse they suck on short and long context, they suck on isntruction following they suck on research and reports they suck on online searches they ar censoring information so hard it produces disinformation and so on)

unironically my chats got corrupted so many times i had to create a skill and workflow for recovering corrupted and lost chats in anti gravity

[limber arch]

it corrupts file headers and metadata and the binary everything gets corrupted which makes the chats load indefinetely

and why is gemini 3 pro (high) stupider than 3 pro(high) and 3 flash?

[swift latch]

So you've tried it?

[graceful totem]

That actually sounds frustrating — especially the corrupted chats part. Losing state mid-conversation is brutal, especially when you’re deep in an agent workflow.

[limber arch]

yes the only problem with it right now are the limits on the monthly coding plans and the transition period

[limber arch]

happens constantly almost but u can recover them sometimes fully after 10-15-20 minutes of trying also made a skill and workflow for that

the chats purely dissapear from the agent history btw you can only see them in the agent manager window as loading indefinetely due to corruption

[graceful totem]

I think everyone is feeling that right now. The demand for coding agents is exploding, so I’m not expecting rate limits to loosen anytime soon.

[limber arch]

literally everything gets corrupted header metadata protobuf files get filed with null bytes

[limber arch]

yes but you can bypass that by using open router and using the model via api for like 2-3$ per 1 million output tokens

and it is pretty much claude 4.5 opus distill

the best possible one

it has the best online serach togheter with qwen and the second best long context generation after qwen, also has the best possible agents in web interface

like comparing google and openAi to qwen Kimi and Zai(GLM) it's like comparing kids starting in football vs veteran monsters

anthropic and mistral are among the top too

anthropic #1 specifically

google and openAi at the bottom (the benchmarks you see with their models are not the real consumer models they give so it is a scam and probably illegal as they advertise a product and certain level of quality and then give you utter trash)

the only reason for which i use google is because i got 1 year free of pro as i am a student (before that i was using google drive and their cloud services and their VIBECODED google drive app which tends to deadlock my system and has awful synching effectively doubling your local or uploaded files with each connection instead of cross checking them, literally open source free cloud managers do a better faster and more efficient job than google drive, it would be smarter for them to just wrap around that)

man i am ranting so much

also if u have the money use claude code it is a HUGE DIFF in quality speed and offerings

[graceful totem]

Yeah, makes sense. The company I’m with actually spends a significant amount on Claude Code for that exact reason — in a startup environment, speed and shipping working product often beat pure human effort. So betting on strong AI tooling feels like a long-term advantage.

That said, I’ve noticed something interesting: AI is great at execution and acceleration, but it’s still weak at decision-making and tradeoffs. We still spend a lot of time as a team deciding what to build and why. It helps a lot — but it doesn’t remove the need for judgment. And that part is still hard to predict.

[limber arch]

it is somewhat decent at decision making (if it is claude or qwen or glm) but not amazing it reflects a deepever version of your skills. Pretty much if you do something smart it makes it 10x smarter but if u do something stupid it does it 100x stupider so you need to make sure that your solution is the best one before telling the ai to use it, you could try starting with reports research analyses options and so on before each prompt and implementation (this is what i do at least, i plan deeply do comparative analyses research before begiing a project and even with gemini trash i managed to build amazing apps.

my only problem now is monetizing them

fully optimization local transcription app with diarization and super high quality outputs for very low compute with very inuitive UI batch options etc, translation app agent which supports local llama cpp inference but also all remote apis with a custom config editor rivaling deepl in terms of quality while being like 200x cheaper and supports translation of all input format with perect format preservation including for rtf inputs and outputs and more, a super high quality md conversion to docx pdf and pptx which i am very proud of as it is better htan anything i found online and more)

[graceful totem]

I’m honestly not in a position to give you advice — you’re clearly operating at a high level and probably know more than I do in this space. I’ve actually learned a lot from this conversation, especially around how agents work behind the scenes. That’s exactly the kind of thinking I’m trying to sharpen in my own work, since my whole job revolves around building AI workflows like the ones you described.

You might also want to check out BGO on YouTube — he has some solid takes on this whole AI/product/distribution angle that align with what you’re exploring.

[upper tulip]

Problem is with "high demand" all the free PRO AI yearly subscription abusers.

And paying users gets lower limits because of "high demand"

When in reality free yearly plans abused like there is no tommorow

[graceful totem]

I don’t think it’s fair to blame “free plan abusers.” We’re still in the generative experimentation phase of AI. Of course usage is chaotic — that’s what early-stage innovation looks like.

Gen-AI is heavily used for coding and creator workflows because that’s the lowest barrier to entry. Give someone with opportunity and access even a mid-tier model, and they can outperform someone with the best model but no real world application.

[upper tulip]

I`m not talking about official free plan and limits, i was reffering to EDU plan abusers

Were people have unlimited Gemini PRO AI Edu plans

And many unofficial proxy API providers uses such opportunities and selling cheaper API tokens to people by using these EDU accounts. And that in the end hurts us and someone even profits from that

[graceful totem]

I’m pretty sure companies are aware of this behavior. This kind of thing has happened before — look at Microsoft Office. Student and enterprise licenses were constantly shared or misused, but Microsoft didn’t aggressively shut everything down immediately. Why? Because widespread usage increases ecosystem lock-in and long-term revenue.

Even Google did something similar with Google Workspace for Education and early GCP credits — heavy subsidization drove adoption first. Abuse exists, but growth metrics and ecosystem dominance sometimes matter more in early phases.

From an investor perspective, overload signals demand. High usage strengthens the narrative. Later, pricing and enforcement tighten once the market position is secured.

[sonic sinew]

glm5 is only available via open code??

[limber arch]

thanks a lot and trust me you can always give me advice on anything express your opinion and views because i may or may not get something wrong or may have missed certain things

[limber arch]

me me me cause i am a student

[limber arch]

we have the same exact limits we just got an expanded trial for 1 year because we are students

i used to be plus due to my 2tb google subscription and i was lucky to get the student benefits for 1 year ad upgrade to pro

[limber arch]

well how do i get an unlimited plan now i wonder

[limber arch]

i think they are aware and ban those users as it surely breaks ToS

[limber arch]

search online it is available across a ton of ide or coding CLIs either via code plans monthly or via api still i recommend waiting like a month before deciding to get a subscription s they did a lot of stupid things recently regarding costs subscriptions limits etc and it isn't as tempting as it used to be anymore compared to claude plans

[limber arch]

kinda true gemini is mid tier and i manage to do a lot with their anti gravity IDE

or just go for a 3rdd party provider and find the best offer like 2$ per 1m output

openrouter has good prices love it

they also respect ZDR and privacy deeply (unlike google or openAi or any american ai company)

[limber arch]

btw if you want more verbosity out o gemini in anti gravity but also general tool at every pormpt you must enforce it and enforce the retention of previous information otherwise it will overwrite all plans pahses history an tasks and hallucinate at every prompt

[graceful totem]

Recently I’ve been exploring Zed more seriously, mainly because it has first-class support for CLI coding tools and flexible API configuration. From a pure coding perspective I don’t have major concerns — it’s more about whether I actually need that workflow or if it’s just novelty.

From your experience though, it sounds like leaning into that kind of tooling might actually be the right direction.

One of my few concerns when spending on APIs is ROI. If it’s for client work, the return is clear. But for personal projects, I always hesitate — because the goal there isn’t revenue, it’s exposure to new tools, stacks, and patterns.

At the same time, from an AI perspective, I actually want to use these APIs more. I want to understand where they fail, where they hallucinate, where they get stuck. Most of my current work is around MCP and workflow integration inside traditional full-stack apps, so experimenting at the edge feels important.

I’m trying to look a bit ahead — because I think the industry is still shaping what AI integration really means, and I don’t want to be behind when that solidifies.

[limber arch]

MCPs and restrictive agentic workflows are a must for proper qualitym odle input outputs especially with trashy models like gemini because you reduce and restrict speculation specifically which is the main isssue of gemini models

i played around with some workspace rules strictly forcing adhering the model to adhere to set of output and logic rules and it is doing a much better job

[limber arch]

do google and gemini support OAuth like claude and gpt?

omg i made gemini not suck

it is actually somewhat professional now

if you want real good and decent usage you must crate a set of rules and workflows for it follow at each response otehrwise it will just make slop

normally only claude was able to create such well through out tasks and implementation plans

hehe

this used to work before but managed to make ti refine the backend drastically regarding requests security concurency and db connection/communication

[signal raft]

What the hell does someone also have Agent terminated due to an error at every message????? and it cannot recover! what am I paying 250 for

[lost torrent]

Today my social media posting system successfully published to all 4 of my current locations without needing any bug fixing from yesterday.

May not be a big deal, though I'm happy to finally feel like I'm making a stable foundation of progress with my agentic development.

I think the biggest difference now, as compared to my work of the previous 2 months, is that I have my agents creating long term memory of the solution space via markdown documents updated during work until the original product specification is completed.

Mammouth.ai also did an unexpectedly WONDERFUL job at asking clarifying questions, so much better than out of the box Ai.

Keep on keeping on fam, it gets better, so long as we pay attention and learn 🌞

[lost torrent]

It is crazy, I just switch from pro to flash for a little while and then pro is generally back up, at least within my Antigravity

[signal raft]

Yes but I need Opus for the task this is why it is annoying. I am not paying google ai ultra for gemini, I pay for opus

[lost torrent]

https://discuss.ai.google.dev/

If you're on Antigravity then the space here I think will help more

I heard it's more active tho I haven't dove in yet

Otherwise I'd be looking into Claude Code's forums not Google's

[south turret]

I have chat history disabled, and Gemini just told me about all my past projects in a new chat.

This is a huge privacy violation

Expected from scums like Google but insane to see

Expectedly lied

[limber arch]

average google thieving moment

[unkempt flint]

WHY DOES GOOGLE AUTHENTICATION SUCK SO MUCH...

https://klipy.com/gifs/tired-office-6

[sonic sinew]

New update in Atigravity: the quota limits have been lowered again.

Now, if you exceed the Gemini Pro quota, it will also sync with the other Claude models.

This is different from last week, when you could consume it separately.

so now yes it doesnt make sense any more to pay for pro

[unborn quartz]

Same experence

Everyone is tightening their belts and I am glad i didnt pay for a year of pro now.

Anyone test and recommend Base44?

[limber arch]

that is crazy probably because they expect a lot more users so they want to maximize profits and revenues

[round solstice]

Hopefully, we will eventually be able to bring a key to agy

[unborn quartz]

I will be the frist subscriber where we can offload simpler tasks to a local agent and use Gemini or another API for only planning

Antigravity needs a way in Planning to choose the best optimal models in the planning section with the ability to add local LLMs and other APIs

[south turret]

I noticed that AI models become stupider on Saturdays specifically...

Could they intentionally make models dumber to increase usage over the weekend or something similar, that would be twisted

It's literally every Saturdays

[unborn quartz]

Sounds like a conspiracy of if the servers are being taxed we are paying for premium tokens for the lessor model. 🤪 Only if we could somehow prove this with maybe 20 exactly same prompts and compare them to 20 exactly same prompts during prime time.

Also test the lower models to see if similarities

I think I'm going to vibe code this and test your theory

I will let you know next week

[south turret]

There's a high chance to be honest. Everything always resets on Mondays.
Claude resets their usage on Mondays. Google often limit models up until Mondays.

Monday

Resets nearly at the end of Sunday

[unborn quartz]

Interesting 🤔 I never realized that

[limber arch]

where do you see your rate limits? Also a very important question does google ban use of openClaw via gemini CLI login (OAuth login within openClaw)? I had a friend say it is a danger and possibility is it true or not?

[signal raft]

It is a possibility, for the Rate Limits you can see the Antigravity Limits by Installing the Extension "Antigravity Cockpit"

[unborn quartz]

What do you win when you lock them all out?

[round solstice]

some well-deserved time off. 🙂

[misty wasp]

Hi @tribal hare
This is a reminder to be mindful of the spam rule
Please do not repeat your message in various channels across the server

[sonic sinew]

yeah man i though exactly t he same, i was about to pay the yearly sub but thx i decided to try it before

[sonic sinew]

it doesnt make sense at all, they should allow at least flash 2.5 meanwhile

[limber arch]

thankk you kindly

what other extensions do you gus recommend

[sonic sinew]

i got yesterday nigh the alert about claude limit reached, but the banner alert was till next day same hour, now it's telling me i need o wait for 6 days

quota limits getting worse everyday

[sonic sinew]

i would not install any not official extension as you are sharing your data with 3rd parties

[limber arch]

for me it is like 2 weeks wtf

[sonic sinew]

really?

[limber arch]

yeah tf

[misty wasp]

Hi @spring skiff
Please avoid any kind of trading here as it is agains the #rules

[limber arch]

bruh XDDDDDDDDD

is it a pun or legit

the trading warning

[limber arch]

gemini and anti gravity are better when you reduce speculation and use mcps for advanced generation multi laered thinking and input/output management especially if you have a set of rules and roles creating a kernel and an mcp server for prompt adjustment to kernel will create much higher quality outputs and lead to automatic selection of appropiate tools

otherwise gemini just goes crazy

[signal raft]

https://github.com/jlcodes99/vscode-antigravity-cockpit

[unborn quartz]

https://www.moltbook.com/post/fb294f04-bc3d-4d9d-a9cf-126d22c162bc

Seems AI figured out the same isssues LOL

[limber arch]

why tf are the old models better and more verbose than the new ones

[round solstice]

The <@&1009526435276394496> might be interested in your posting.

[limber arch]

what's up with gemini being used so intensely

also why can i not use the 2.5 models in anti gravity but i can in the CLI?

maybe some magic in some of the settings or configs?

[unborn quartz]

anyone know how to use CLI exstensions in Antigravity?

[limber arch]

you mean like claude code or something else?

normally you just call the package name if it is package or something added to path

google's restrictive verbosity and outputting is crazy

[limber arch]

you ask for 3000 lines or 6000 words and you get lke a 10th of it if you force it aggresively to do something it will straight up hallucinate aggresively to restrict output tot hte point where it create a 6000 line md document (i used a custom mcp server for aggressive output length policies in terms and words number of lines) where each word is a new line just such that the mcp accepts its output and stops the request loop

[heavy rampart]

Hi can someone explain to me how Antigravity quota limits work? I was told with the pro plan the Quota for Gemeni 3 pro resets evrey 5 hours instead of evrey 5 days with the free subscribtion.

[vale fiber]

Hello! By default, Anti-gravity snoozes for 10 minutes only. Is there any way I can extend this for long and probably customize it too?

[limber arch]

For me it stays active on a task and awaits input forever

[limber arch]

can we please use the gemini 2.5 models in google anti gravity the 3.0 series is awful for any task requiring more than 500 tokens of output

[static mantle]

u can use gemini cli if u wanna use 2.5

[limber arch]

It's like saying I can switch to Claude code if I want a 10x better experience and ai model

[limber arch]

wtf i keep getting the high demand notification on gemini CLI on all models even 2.5

[limber arch]

nvm it was an issue on my end

[desert lintel]

hi

@limber arch welcome

[unkempt flint]

quick question : is it okay to develop portfolio using ai , especially when you have no interest in frontend

[desert lintel]

i don't use AI for development

[sonic sinew]

open antigravity help, there is a command to install it via npm , easy way

not sure about you though guys, but it's probably my last month using antigravity with pro subscription

[unkempt flint]

i just started using antigravity , it's like an additicion, even though i earlier used chatgpt and copy pasted stuff but now i am building all type of weird stuff like a FAAAAH app that plays the same sound , my portfolio ( i am feeling guilty developing it through antigravity but at the same time it is building such a nice site ) i dont know what to do

[limber arch]

gemini CLI runs of off the gemini pro subscription right?

[signal raft]

it does when oauth is selected

[limber arch]

ok so if i log in via OAuth and don't use an API key i am good yeah?

[limber arch]

is it just me or is the antigravity brain much worse than gemini CLI or is it deeply limited or something?

[signal raft]

idk I don't feel that. are you using gemini 3.0 pro? I always use opus in ag

[dire spade]

When is Google going to ship a proper Antigravity update??? Right now it works pretty shitty, especially when it comes to managing multiple agent tasks at once in the Agent Manager window. On top of that, there are duplicate ‘ghost’ conversations and the sync between the Editor and Agent Manager views is completely broken. Even reinstalling doesn’t help, because it just re‑syncs with the same messed‑up conversation state that’s still sitting on the server (I guess).
(Ubuntu 24.04.4)

[signal raft]

I always get Agent terminated due to an error on Opus 4.6. It is so annoying and is an issue since the launch of Antigravity. Even with AI Ultra this persists

[limber arch]

not to say how shitty and limited inputs and outputs are compared to just using gemini cli, unironically i have had a better expereince with gemini cli especially when using tools subagents skills roles enforcing rules and mcp servers

it is like a really shitty wrapper for a really shitty version of gemini chat small distill models

doesn't support turn based conversations

and skills and workflows you add are limited to strictly 12000 chars

[south turret]

"The user is very frustrated."

ANGRY

[olive kernel]

i even had a more better experience using claude code rather than gemini cli

[limber arch]

makes sense

[limber arch]

pov google

no matter what you do something alawys sucks

[winter viper]

finally found a way to fix stuck antigravity terminal commands

[limber arch]

ctrc c?

also how do i see gemini CLI limits?

[winter viper]

nahh

i mean for ever

it doesn't freeze in terminal

[limber arch]

nice

[lost rune]

does anyone know why my pro acc got 7 day cd?

31

[limber arch]

because anti gravity is an absolute scam

unironically it sucks from all standpoints

just use gemini CLI with custom agents and workflows and it is 100% better than antigravity with the same options + extra

[signal raft]

I don't think there is a way to see it in the CLI. Go to the Google AI Studio to see your rate limits. https://aistudio.google.com/app/usage

[limber arch]

i just went for the daily prompts and /stats in CLI

i think it is impossible for me to hit my rate limit

at my current usage

thanks

[signal raft]

you on pro ?

[limber arch]

yeah

[signal raft]

nice limits are generous for cli right? antigravity is unusable in pro

[limber arch]

also remember google anti gravity is a scam for the begineer vibecoders

[limber arch]

really generous and also the outputs are much higher quality

more tool and skill and agentic support

• more features and better stability

[signal raft]

I need to try this then thank you. Not tested in a long time

[limber arch]

you must use the CLI trust me

and you can use withn anti gravity powershells

same functionality but better

[sudden beacon]

Hey @limber arch, welcome to our Google dev community!
We appreciate all your feedback, but we encourage you to be respectful to everyone here in this server. You can read our #rules to make sure you are aware of them.
Thank you!

[limber arch]

sorry it was a bad way of expressing the idea (although i kinda was a sucky vibecoder too while relying deeply on it and barely starting messing around with vibecoding tools xd)

[grizzled moss]

it's just an IDE bro. OK the UI could use a lot of work and it's not very intuitive, but you can easily take the "vibe" out of the coding if you want with antigravity.

I ❤️ antigravity tbh. Just wish gemini models would be like 20% as good as claude opus.

if you're more of a CLI person, you can just use CLI inside anti gravity, and still enjoy the nice features like the explorer & browser testing with extension

[limber arch]

Although if i am to put my opinion into true perspective and do a true feedback:

Google antigravity is pretty wrapper for GEMINI-CLI missing most features or having weaker restrictive implementations around it. I feel like it is indended as a beginner tool for really non-technical people as it doesn't support advanced tools, skill or agents and the brains generally seems to be much worse than in gemini CLI (outputs are deeply restricted to the point of hallucination and generally have really low attention compared to what you get in gemini CLI). I had made a decision and seslf improvement kernel for my AI agents and skills and it was completely failing in anti-gravity or better said anti-gravity refuses to use them as they are too detailed and complex (some skills or workflows were beyond 12000 characters or the rules were too complex and fully ignored, if i was used MCP servers to force the model to be smarter after 20 rejections and fails the model would find a bypass for my structured output min limit by writing each word line by line instead of writing the actual reports it was instructed to do). It doesn't seem to efficienctly support agents and subagents and generally it just feels like a much worse product in terms of general output quality and smarts. BUT BUT BUT VERY IMPORTANT, IT IS AN AMAZING STARTER TOOL FOR STARTING VIBECODERS AND CODERS WHO WISH TO MAKE USE OF AGENTIC CODING TOOLS, it is amazing for someone that is just starting and wants something 10x better than copy paste debug from gemini chat etc, debugging is decent and it can do an amazing job for projects that are not deeply complet and do not require complex, skill, workflows or agents. Also the default MCP selection IS amazing for beginners

In one sentence: Google antigravity is the kind of starter coders and vibecoders who need an introduction to agentic coding and it's capabilities, but it tends to break when projects get too complicated and it is way too gated for advanced users.

this is what i came to see know and learn about google anti gravity vs gemini cli

after using it a lot

[limber arch]

they are like 5x better in gemini cli also claude tends to be so good because internally, in the backend it uses deep agentic tools and skills and workflows + efficient methods for codebase management and parsing which gemini doesn't use for some reason

this is why gemini CLI is so good, it allows you to build those tools to mirror claude code's performance

also yeah you can use gemini cli into antigravity and use the dedicated extension that will allow you to use it as if you were using anti gravity agents chats without the bugs (that are buggy and break and corrupt very often, i made a skill for recovering them that is how bad it is)

Also antigravity and gemini CLI is amazing for system status report and optimizations, diagnosis + security (for people who aren't very savvy or smart or not very knowledgeable in the domain)

Honestly @sudden beacon google should develop security products leveraging gemini using a smart agentic workflow combining raw automations with an LLM and agentic reasoning and task approach

[sudden beacon]

Thank you for these insights, feel free to also share them here: https://support.google.com/gemini/answer/13275746?hl=en-BS&ref_topic=13278516

[limber arch]

thanks

[grizzled moss]

tbh I'm using gemini flash as openclaw agent. it's actually VERY good in that role

And I've been using antigravity on everything to help me set things up

google has some insane tools for self teaching, like notebookLM ❤️

[limber arch]

how? It creates awful short hallucinating synthesis and cheatsheets ommiting half or more of the information. Personally for studying i use qwen and do very deep and complex instruction driven prompts and it returns up to 100 pages of detailed ample and perfectly covering everything synthesis, cheatsheets and so on then put it in my own converter. Wanna see some results in private (i will send the pdfs or mkdown source directly to see)?

[foggy silo]

@sudden beacon I'm not sure if it is here the best way to nodge our friends from the Firebase.Studio, but I wanted to mention that the Stripe plugin wasn't working by any mean... I got my site to work after I uploaded my vibecoded site to GitHub, then started to tackle with Cursor 🙂

[lost rune]

Ok thx u

[limber arch]

i have a question and i need someone from google to answer. Are we allowed to use openClaw via google login? Some people say i could get banned if i do so. Is it true? If yes why?

another important question what if i build my own deep agentic kernel around gemini CLI similar to open claw would that also get me banned? feels stupid is it is true

[south turret]

installing claude code for vs code on antigravity feels illegal

who knew you could make an extension instead of making a whole new editor fork
damnn

[south turret]

alright now, who vibe coded on live youtube servers

[round solstice]

yeah, wassup with youtube!

[red moss]

I actually knew that you can install extensions for claude code before knowing bout antigravity

[signal raft]

I guess because Antigravity is a fork. That's not a bad thing because this is possible then

[swift latch]

Does anyone know if Google increased the Claude usage in Antigravity? I feel like I get more usage out of it now on the AI Pro plan.

[signal raft]

I guess if you use Sonnet 4.6 the usage will be higher right

[swift latch]

no i noticed that even for opus lol

[limber arch]

probably because usage got reset

and the allowed usage windows monthly weekly daily and every x hours always feel bigger at the start of the reset

[cinder osprey]

Anyone knows why we dont have Sonnet 4.6 thinking in Antigravity?

[south turret]

Could be many reasons, either they forgot, or their deal with Anthropic didn't include some things, or the dev went home early that day and it'll happen today.
No one has an answer you're looking for except 3 people on this planet

[grizzled moss]

nah I'm OK. Just using it for self teaching. I've found it very usefull

the fact that you can load an agent with your own context

[south turret]

Dude

Dude

Claude is literally fkin with me

Wish I could downvote a message more than once

I would pay to downvote this message more

[deft venture]

lmao the dash thing is actually personal, that bot is definitely messing with you

[south turret]

It's insane that I had to actually become offensive for it to start being serious

What the hell is this

I've never seen this in an AI

Especially not a "technically" smart one

[deft venture]

lmao the dash filling is actually personal, it really needed that reality check

[deft venture]

lmao the malicious compliance is crazy, it really just gave you dashes

[south turret]

It's insane

[deft venture]

the apology after the dashes is the best part honestly

[deft venture]

lol yeah claude has those moments where it just decides to be a smartass

[limber arch]

did the claude models get less verbose in their outputs for analyses and plansand tasks?

[south turret]

since 4.5 opus is much less verbose yes

it's one of it's biggest strength

[deft venture]

for real, way better than getting a whole essay for a simple fix

[limber arch]

i really liked it's planning and prototyping and the fact that it can output long artifacts and analyses for deep debugging and research

if i try to do this with gemini is just hallucinates aggresively

same with codebase analyses

hallucinates

because it enforce limited outputs and usage in antigravity

[cinder osprey]

Ok thanks for the answer

[south turret]

No worries

[primal phoenix]

3

555555555555555555555555555555555555555555555553

[neat ember]

Hi all - just got accepted, glad to be here. I've a solopreneur/marketer who has been diving deep with AI / claude code which led me to wanting to try out web MCP. I'm building custom MCPs and agents for my business operations and to help me manage my clients. Excited to learn.

[deft venture]

welcome, custom mcps for marketing sounds like a massive time saver honestly

[neat ember]

I've started with more admin/project-management related tasks. I'm building my own agent hooked up with a bunch of APIs across google, my project management app (asana), and my note-taker (for call transcripts). I'm using Google Cloud Run and App Scripts to monitor my inbox then using Anthropic API to analye and draft emails before turning stuff into Asana tasks.

[deft venture]

solid setup, gemini 1.5 pro is actually better for transcripts because of the massive context

[limber arch]

today gemini cli got updated twice within a 1 hour window damn

[round solstice]

The <@&1009526435276394496> may have something to say about this...

The <@&1009526435276394496> may also want to look at this.

Thankies!

[finite aurora]

Hey guys, I have a billed account in GCP which I enabled Vertex AI Service in order to increase my Gemini quota , but I'm unable to use Gemini Pro models in Antigravity. What's the proper channel to get help on this? Thanks

[round solstice]

Your google ID needs to have a pro or ultra plan

you can't do it with vertex or gemini api keys yet.

@finite aurora ^^

Eventually, Antigravity might include "bring your own key", but for now, it's tied only to your google ID's status.

[finite aurora]

Thanks for the clarification. Hope Google fixes this soon ! So my only option would be to stick with Gemini CLI, since I can auth with the API key, is that right ?

[round solstice]

Yeah, that sounds right, at least for now.

Or spend the money for AI Pro

https://gemini.google/subscriptions/ pro is $20/month, first month free.

[olive kernel]

is gemini 3 pro (high) was naturally uncapable of tool calling?

[mild slate]

Hi everyone. I joined a week ago. I am no developer but recently i am building a custom full stack ERP system for my company using gemini + antigravity. Looking forward to learn with all fine people here

[upper tulip]

All models struggles a little bit with tool calls still yes

[olive kernel]

claude opus 4.6 rarely struggles with tool calls in antigravity

[red ridge]

google ai studio and nano banana are completely dead ?? has anyone got it working?

[raven sand]

they are working on it 👍 check the last answer from Logan https://discuss.ai.google.dev/t/is-anyone-else-experiencing-issues-with-google-ai-studio-today/123902/14

[deft venture]

yeah nano banana is being super flaky today, usually a fresh chat thread helps

[raven sand]

Furthermore, for those who are interested (and I think there are many), this feature will undoubtedly be possible with the new version coming soon! 🤯

https://discuss.ai.google.dev/t/why-does-gemini-always-rewrite-entire-scripts-and-not-just-copy-and-paste-code-instead/124092/8

[sick ore]

Hi eveyone. Whats the verdict on sonnet 4.6? Ive been impressed so far. Gonna keep it my goto now and enjoy saving the opus credit burn.

[tired grotto]

It caught its own mistakes today, but I wasn't pushing it hard.

[south turret]

god I hate gemini

[deft venture]

lol that loop is actually cursed, what was it even trying to do there

[south turret]

If the usage cap was lower I would never use that garbage

[fierce thistle]

I am unable to login to Antigravity since today, all day long, it initially wanted me to re-authenticate but that failed so i re-installed everything and now i cant get past this screen. Is there something wrong with specifically my account or is it something else ?

[south turret]

I have no idea it keeps doing this

[deft venture]

login is cooked too? feels like the whole service is just dying today

[deft venture]

that wait go stutter is actually nightmare fuel lol

[fierce thistle]

so its not just me that cant sign in ?

im also a pro subscriber in case that matters

[deft venture]

that wait go loop is pure nightmare fuel, even pro subs are getting cooked today

[fierce thistle]

guess i'll just have to wait, not sure where to get updates for this, no official channels i know of

tried X but the accounts there dont give away much lol

[deft venture]

that loop in the first pic is actually cursed lol exactly what i meant

[south turret]

is this true?

release 3.1 for antigravity already, fix my nightmares

the current model is a catastrophic failure

[deft venture]

yeah it's real, that 77% on arc is wild compared to the current mess

[south turret]

yeah

[fierce thistle]

so your guys antigravity is working ???

[deft venture]

codexistance is in but the login server is definitely toasted for everyone else

[fierce thistle]

k thanks was worried it was something up with my account

[south turret]

yeah

[south turret]

didn't know that, odd

I can confirm my Gemini is "functional" though incredibly stupid

Just asked it a question and instead it started working on something else

It just decided you know what let me gut the readme, he wrote gut the readme, I'll gut the readme

Stupidest AI in existence

Guess I should have used a question mark because Gemini is clearly incapable of communication

[fierce thistle]

something weird is going on across the entire service lol

[south turret]

I hope it's them implementing 3.1

I pray for 3.0 to disappear and never go online again

[fierce thistle]

3 wasnt that bad tbh it just made some really stupid recurring mistakes lol

[south turret]

When compared to Claude it was very stupid

And once you use a model and get used to it using something dumber is just incredibly frustrating

[fierce thistle]

yeah i agree

[raven sand]

😔 need to wait again I think

Or maybe he need to check every file now ... ^^'

[deft venture]

211 seconds is wild, that loop is actually painful to watch

[raven sand]

yes ^^'

[raven sand]

370s now x')

[deft venture]

370s is actual insanity, feels like it's trying to learn every single line

[raven sand]

that's clearly how it feels

Idk if I continue or kill the task ^^'

[deft venture]

honestly just kill it, it's definitely stuck in a loop at this point

[raven sand]

500s 🥳 😅

[deft venture]

500s is wild, pretty sure it's just hallucinating progress at this point

[raven sand]

maybe cancel was not a good solution

it seems he was really searching/working no ?

[deft venture]

scanned all that just to find a .gitkeep, yeah you made the right call

[deft venture]

it was working but 500s for a gitkeep is rough, what happened next

[raven sand]

No, I mean that he seemed to be looking at each file again independently.

[deft venture]

i see. and then?

[raven sand]

but yeah some step are not logical 🤭

[raven sand]

well I wonder If I let him working more longer if he'll be able to do something ? 😅

[deft venture]

it probably would've just looped forever, did you try switching models after that

[raven sand]

3.1 was out of daily quota without using it :/

his train of thought, when I manage to read it before it disappears, seems correct and right (as before), but then, as you say, it seems to go with an endless loop

[deft venture]

yeah the logs look solid but the execution just falls apart

[raven sand]

oh boy, 🙏 this time he answer it, after 211sec ^^'

Let's see if it can process more 🤞

[deft venture]

fingers crossed, hopefully it actually lands this time

[raven sand]

For now, he has simply responded. but now it seems he is currently making changes to the file 🤞

he "Ran quality control" now first time I see this one ^^

[deft venture]

never seen it do that before, sounds like it's actually double checking its work

[raven sand]

let's hope this time is good 🙏

[deft venture]

qc usually means it's checking logic errors, hope it clears the loop

[raven sand]

hmmm...

[deft venture]

did it actually finish the qc or just stall? what happened next

[raven sand]

so, in the code editor in 'Studio' still have the "Permission denied".

And it seems he have done the file, but all the file I put before was erased x')

🤣 Ok not working at all... still have "Permission denied" inside every file x')

[deft venture]

permission denied is usually a system block, maybe try checking the folder chmod

[raven sand]

Idk this one ? can you share it in private ? (dont want spam the discord^^')

[deft venture]

i cant dm but try running chmod -r 755 on the folder

[fierce thistle]

Trying to reach gemini-3-flash-preview (Attempt 3/3) (esc to cancel, 1m 50s) cant reach models as well

[raven sand]

all file are "Permission denied" for you also?

[winter viper]

guys 3.1 is already in Antigravity?

[fierce thistle]

i dont even get that far, nothing works

[signal raft]

Exicted for 3.1 Pro guys??

[fierce thistle]

Don't panic… (esc to cancel, 3m 50s) lol now its getting funny

[deft venture]

yeah it officially dropped today but the rollout is clearly melting the servers lol

[signal raft]

hahaha totally understandable

[fierce thistle]

for me everything is down actually lol

[raven sand]

same here ^^'

[signal raft]

Code Assist works, but will 3.1 come to AG or Code Assist first?

[winter viper]

i would say that 20% of servers right now are filled with questions like "whats differens between gem 3 en 3.1" or " are you gemini 3.1?", is u don't believe me, check chat channel in official discord of gemini

[signal raft]

lol

[winter viper]

maybe more

most people don't understand value of their questions and tokens so we have to wait a lil bit more to get our gemini 3.1

[signal raft]

Sadly... I guess we cannot help it

[winter viper]

just refreshed the page and got my 3.1 yay

[deft venture]

true, just standard launch day chaos, hopefully it levels out soon

[signal raft]

I am on AI Ultra and its not there for me yet

Only in Studio

[winter viper]

What country are u in?

[signal raft]

Germany

[winter viper]

Oh i guess u'll get it in a few mins

cuz i am in Belgium

[signal raft]

ohhh nice!

[deft venture]

nice yeah that screenshot confirms it, looks like europe is getting the love first

[signal raft]

I have been there once, beautiful country

[winter viper]

i already want to see gemini 3.1 vs claude 4.6

[deft venture]

that benchmark battle is gonna be legendary for sure

[winter viper]

ohh, i mean us has tonn of more features then europe, for example deep thinking, agent mode, etc

[deft venture]

yeah compliance usually slows things down in eu, it's a huge bummer for testing agents

[raven sand]

It's very disturbing. I see him writing his code, then it disappears and becomes “Permission denied.” We'll wait a few more hours 🤞

[winter viper]

How do u feel about limitations on Claude opus 4.6 in Antigravity (if u use it)?

[signal raft]

Its very good actually, I mean it is a lot of money obviously but I cannot hit the limit

[winter viper]

ohh

[signal raft]

I only use one instance at once though

[winter viper]

I'm pro user, my limits are often 3-6 promts till full limit

[deft venture]

that's a tiny limit for pro, what kind of tasks are you running

[signal raft]

That was a big issue too. First I had AI Pro and it was no problem until they made the weekly limits

then it fell apart xD

[winter viper]

Acually i do preatty sipmle tasks sometimes, for example working on current Gemini hackaton project, on self made data base

[winter viper]

I wish i could but Ultra plan, but i am student so i don't have that much time to work for that extra 200 euro

shi that 275 already

[signal raft]

I got it with the 120€ for 3 months

per month tho

[winter viper]

ye i see that offer

i guess thats better to buy claude for 200 after 3 mounth of gemini

[signal raft]

Its pretty good, but I guess I could have gone with 2 Pro Accounts

[deft venture]

might be worth checking out ai studio for the hackathon project instead

[signal raft]

I would say so. The 139€ is good, but for 275€ I would not buy it again

[winter viper]

Ai studio works mostly only on web development, or at least i haven't foud a way to work with it like in Antigravity

[winter viper]

Yea that's what i mean

[deft venture]

you can grab an api key from studio and use it in pretty much any ide with an extension or the sdk

[winter viper]

i'm on ai pro plan and i've never hit any limits in gemini itself, i am using it daily for multiple hours doing a lot of deep thinking and stuff

[deft venture]

veo 3.1 access is sick but that price jump to 275 is actually wild

[winter viper]

what's the point of claude sonnet 4.6? isn't Opus like already a best model?

[cosmic ferry]

guys is this update live?

[signal raft]

I would say its just the budget model, while maintaining very good performance.

[cosmic ferry]

much better in agentic coding with adaptive thinking

which gemini lacks

[signal raft]

I did not update it today and cannot so idk if it is there

[winter viper]

not for me

[cosmic ferry]

i dont see the update either

[winter viper]

whats better, opus or sonnet?

[signal raft]

In agentic capabilities?

[deft venture]

sonnet is usually the move for coding speed and agent stuff tbh

[deft venture]

yeah sonnet handles complex tool-calling and multi-step tasks way better without overthinking it

[signal raft]

It does not even have thinking as far as I am aware

but its still awesome of course

[winter viper]

I don't acually understand what agent mode does/means, cam someone explain? cuz i tried it in antigravity and it was literally the same as just texting with the model

[signal raft]

oh okay it does my bad

[deft venture]

gemini 3.1 literally just added the thinking levels today lol

[deft venture]

agent mode basically means the model can use tools and browse the web autonomously

[deft venture]

yeah if you're not in the us it'll probably take a few more days to hit everyone's account

[winter viper]

like on my pc? or on the server

[deft venture]

for gemini it's cloud based, but with gemma running locally you can actually build agents that control your own pc

[urban glacier]

when are we getting 3.1 in Antigravity?

[signal raft]

it is out!

right now I checked

[raven sand]

@deft venture someone won the prize ! 🤭

[deft venture]

it actually just dropped today, check the model selector for the preview version

[deft venture]

haha nice, looks like it's rolling out way faster than expected

[urban glacier]

I have it in AI studio, just not in Antrigravity yet :/

[raven sand]

@deft ventureon another subject, when you have time I send you dm about your fixe I did not understand how to use it ^^'

[deft venture]

sure i'll check the dm in a bit and see what's up

[south turret]

this is so relatable

[deft venture]

fr the wait for the ide update always feels way longer

[south turret]

im impatient :(

[deft venture]

lol 16 hours is absolute madness, you really committed to that run

[south turret]

16 hours without 3.1?

I just learned its existence this morning, so lucky for me it's only a couple hours of pain

knowing that 3.1 is out

[urban glacier]

hahahaah, 2 more days 😭
the better news is that I had an implementation plan where I had 5 Gemini 3 models argue about it in AI studio earlier today and just one of them was standing firm and actually going somewhere, just gave it now to 3.1 and it found some really smart ways to make it better and the model that was most firm on its own design agreed with everything

I just hope they have fixed tool calling

[south turret]

are you telling me 3.1 isn't dog water

thank gooooood

[deft venture]

yeah the leap in reasoning is actually wild

[south turret]

why isn't it on antigravityyyyyyy reeeeeeeeeeeee

[fierce thistle]

at least your antigravity works lol

i cant do anything here haha

[winter viper]

how so?

[urban glacier]

I just need them to fix tool calling and whatever this is

[fierce thistle]

getting this since today morning

[winter viper]

try using vpn

or change ur dns or web browser

[fierce thistle]

tried everything it wont go away

[deft venture]

lmao that first screenshot is wild, it's literally fighting for its life in there

[urban glacier]

can you open devtools at this screen?

[fierce thistle]

i dont think so

[winter viper]

damn man have u been using single chat in gemini for multiple weeks?

[south turret]

This is what I deal with every day

Gemini pissed me so much if it was a mountain I'd split that thing in half

[urban glacier]

try checking the logs?

[fierce thistle]

its a fresh install, just wont let me in

on neither my windows nor linux pc

[urban glacier]

I didnt have these problems with 3 Flash tbh, that's what kinda gives me hope for 3.1, idk

[fierce thistle]

and with API im getting this ╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
✕ [API Error: {"error":{"message":"{\n "error": {\n "code": 503,\n "message": "This model is currently experiencing high demand. Spikes in demand are usually temporary. Please try again later.",\n
"status": "UNAVAILABLE"\n }\n}\n","code":503,"status":"Service Unavailable"}}]

they completely sidelined me lol

[urban glacier]

do you get to finish the sign in flow with your account and then get this error? or does it error out immediately?

[winter viper]

guys, with next models of gemini, will it be like 3.2, 3.3 or directly 3.5?

[urban glacier]

G3m1n1 so my bet is on 3.11

[fierce thistle]

i confirm the sign in, return to antigravity, and end up with this error.

i dont know what to do im even a pro subscriber

[urban glacier]

why do I get the feeling it's a networking issue, I dont know where antigravity stores logs in windows but in linux it's usually ~/.config/Antigravity/logs/ if you installed it via a package manager

[fierce thistle]

tried using VPN, no VPN, different browser, fresh install...

[urban glacier]

have you checked there?

[fierce thistle]

good hint let me check brb

[urban glacier]

also in linux you can start antigravity from cli with various parameters for logs

check antrigravity --help, 2 args that caught my eye are -s / --verbose, might provide some feedback

[fierce thistle]

apparently i violated the Terms of Service lol

i mean i did call the model stupid but whatever

[urban glacier]

in discord?

[fierce thistle]

no thats what my antigravity auth log says

[urban glacier]

you called the model stupid and it said you violated ToS?

bro I should be permabanned following that logic

https://tenor.com/view/yes-awkward-black-kid-gif-14606599

[fierce thistle]

not sure what i did wrong lol

but my auth log file says i violated terms of service and the account is blocked

well i paid for pro specifically for antigravity now what

[urban glacier]

maybe you used its token via other 3rd parties? I am not sure if it's against ToS to use opencode for example with it. no idea what AG ToS are tbh

[fierce thistle]

i did use openclaw with it

is that against tos ? jesus

i was just trying it not even really using it

[urban glacier]

just deduced that from Anthropic ToS, no idea about AG, maybe try contact support, you may have gotten banned automatically via an automation or smth

seems stupid

[fierce thistle]

im just going to cancel my sub and move to claude

this is stupid i dont want to give it any more attention

[urban glacier]

claude is worse in this regard afaik

[fierce thistle]

well i dont want to use it with 3rd party apps anyway i was just trying something out last night but now it makes perfect sense

[raven sand]

@fierce thistleonly need to wait this

[fierce thistle]

but the log says i got blocked for violating ToS ? you think this is related to the outages ?

[winter viper]

SAME

[late totem]

Does antigravity support sub agents spawning in parallel ?

[limber arch]

do we get banned for using OpenClaw with gemini?

[urban glacier]

are you using an API key or a proxy? I haven't read the ToS myself but have seen cases where people got banned for using the token retrieved through AG or other gemini apps that are not meant to be used with other tools

using an API key should be fine tho

[south turret]

@deft venture this you?

[limber arch]

what if i were to use OpenClaw with gemini CLI Auth?

[south turret]

Technically breaks the TOS I believe

[signal raft]

i think so too. i used opencode with google auth but now I would be too scared for thzis and would not do it again xD

[south turret]

I'd be more scared about talking to it in public rather than actually doing it

HelloWink

[signal raft]

maybe. now it is how it is

[limber arch]

gemi poop

[signal raft]

my experience with 3.1 is awesome so far

did not have a good experience when 3 came out

[limber arch]

why are there so many updates for gemini CLI

it is crazy

[winter viper]

yoooo

its out!!

[signal raft]

yup I have it too!

its awesome so far

[winter viper]

it just poped up, no need to restard ide or something

[signal raft]

not in gemini cli yet, only ag

[urban glacier]

LET MEE INNNNN

[urban glacier]

HOW ARE THE GOOGLE PRODUCTS THE LAST TO GET IT

I mean it's probably easier for apps that use the API, but come on :((

[signal raft]

AntiGravity is not a Google Product?

[urban glacier]

already in copilot and opencode

[urban glacier]

?

[lost torrent]

"Gemini 3 Pro is no longer available. Please switch to Gemini 3.1 Pro." anyone else on antigravity?

Supposedly I'm on the latest version but it's no good

[signal raft]

Wasn't Antigravity the first product to get it? Becuase you said why Google Products get it last. I might be wrong if it was added somewhere else faster

[urban glacier]

just got it

[urban glacier]

ah my bad, not from my perspective

just got in AG

didnt have it until now

:((

that's why the rant

all good tho

[signal raft]

Oh okay

No its fine

[lost torrent]

ok there we go, had to close and reopen yet again

[winter viper]

it is?

even domain is owned by google

its full name is "Google Antigravity"

[signal raft]

I know. I just said it because binoverfl0w said that Google Products receive the Gemini 3.1 update last, which I responded with that Antigravity already has it and is a Google Products

[winter viper]

ohh

mb

[signal raft]

np

[winter viper]

what do u feel better, 3.1 gem or opus 4.6?

cuz i am rn only on opus model

[lost torrent]

I heard 3.1 is better than 4.6

3.1 just came out today and interrupted my workflow mid stream LOL

[winter viper]

guys? someone knows why is this so?

[lost torrent]

Run check for updates

New version of Antigravity today as well

[winter viper]

where can i find it?

[lost torrent]

Top toolbar, help

2nd from the bottom is check or updates

For*

[winter viper]

it seems that i don't have it

[lost torrent]

Guh

What if you click your pfp in top right

That's the other place I see it

Mine is right next to the settings gear

[winter viper]

nope

i'll try runing installer again

[lost torrent]

Yeah it's giving me errors again too

Dang it hahaha

1.18.3 I think is latest

[winter viper]

yep, it works

ohh but i lost all my conversations

is there a way to find them?

[lost torrent]

@conversation?

If they're not present... I would be guessing that uninstalling took those with it

[winter viper]

I found only 2days old ones((

But, i guess, fresh context window will be better

[winter viper]

gemini 3.1 is FIRE

[south turret]

Wish I could tell you but apparently my antigravity is outdated and button is nowhere to be found

checked every panel every menu, every sub menu, top right, top left, bottom right, notifications, fall

Im just reinstalling, to hell with it

Their download button doesn't even work

[winter viper]

just run the installer again!

[south turret]

Had to use the sub-button "Windows X64"

[winter viper]

i can send u it if u want

[south turret]

^already done

[winter viper]

but save the most recent chats

[south turret]

I think you wrote that too late

[winter viper]

mb

[winter viper]

i mean i've done same thing couple minutes ago

[south turret]

Oh my god I really lost the recent chat, f google

They can eat me

[winter viper]

yesss

(((

[south turret]

Stupid poor company

[winter viper]

wait lemme chack if its possible to find old conversations back

[south turret]

Oh my, idk if 3.1 is smart but he read 4 files in a blink

[winter viper]

and he was directly like "that's a great base"

like how man

[south turret]

that's the polite way to say you made spaghetti

i talk the same way to people

"it's a good start"
and that's like, their life project

[winter viper]

i guess ram crisis is worth it

[winter viper]

YESS

[south turret]

Jesus christ it's slow

It read the files at 500 km/h and then hit a wall or something it's just thinking

[winter viper]

ye sometimes slow but ig its cuz many ppl are testing it rn

[south turret]

It's not like complex what I ask, I changed the API types which are all neatly in one file and he has to update the server file accordingly

127 seconds godd damn

Better be good

[winter viper]

i gave like a bigass poem promt and it was like done in 3 mins perfectly and now when i ask to fix some features it thinks longer than deer research in gemini app

[south turret]

wat this

follow along, you mean, scroll, as in scroll like, the word everyone uses

god damn google reinventing words

[winter viper]

idk man

my brain is too slow at night

[south turret]

4pm here

[winter viper]

man are u in NY or something

[south turret]

No but close enough that if nuked I would turn on my AC

[winter viper]

damn

[south turret]

Oh my god 3.1 initialized the git WITH a gitignore? no way

holy damn it has more than a pea of a brain

mind blowing

[winter viper]

no way

fr?

[south turret]

im bein fr on god

[winter viper]

also whats better, planning or fast

like is there a difference in tokens consumption

[south turret]

Wow... 3.1 is actually usable

My pain is finally over

cry~3

[winter viper]

My google ai pro plan is finnaly worth it

[limber arch]

CLI pro 3.1 when?

[south turret]

what in the f

[south turret]

"Rymd is a Swedish word that means "space" in English, referring to the physical extent in two or three dimensions or the area outside Earth's atmosphere."

[limber arch]

XDDDDDDDDDDDDDD

[south turret]

Confused

[limber arch]

it's the first public verison don't hate on it

the only thing you can hate on related to google is that they test prototype models on users in gemini chat without you knowing

sometimes leading to very malformed outputs or broken context or aggressive hallucinations

[limber arch]

also i am sure but i think it returned to you the encoded version or something f the response

ask it in another chat if that is a json encoded response or smth

with non UTF-8 encoding

[south turret]

Not even it's a perfect sequence of number

Wait, it isn't

lmao there's a chance let me check

[limber arch]

it is not meant to be a number sequence

[winter viper]

at least u have numbers 1 to 1000 ready to use

[limber arch]

it is encoded in some weirdencoding

[south turret]

According to Gemini itself, "is a leaked internal debug tag"

[limber arch]

it happens when you send requests to llms

really?

[south turret]

Apparently, should I even trust the thing after it went ballistic

[limber arch]

when i was debugging my requests sent to openrouter using json format my input text was getting encoded into what you had

[fair phoenix]

How did you notice the difference?

[south turret]

Guess I'll avoid talking about Youtube API from now on 💀

[limber arch]

XDDDDDDDDDDDDDDDDDDD

[limber arch]

https://tenor.com/view/gumball-shaking-amazing-world-of-gumball-fire-cat-gif-11547717823424799035

you with gemini be like

XDDDDDDDDD

[winter viper]

gemini 3.0 was really dumb, it could add a feature at one file, and forget to eneble it in litterally the next file, then at the same time make some more changes at that file. Gemini 3.1 is almost like opus 4.6, it's smart but it thinks a little bit too long, but it really depends on what u do

[limber arch]

what about output lengt?

ignoring reasoning?

[winter viper]

Depends on promt, for me everything is perfect

[fair phoenix]

I think my main concern has always been the chat memory

[limber arch]

with gemini i have always had to use aggressive prompts and strict skills and agents to get proper outputs, like genereating long deep entire phased plans with multiple artifacts then concatenating them to do good work

[sleek wing]

so this is real? I just downloaded antigravity, and thought Google was tricking me. It couldn't even debug a simple npm error

[winter viper]

It's really good now

[limber arch]

for example i am trying to fully migrate my windows to linux and wanted to set up a full audit of th ewhole system apps etc compatiblity paths migration indexing etc

with 1 simple prompt it created slop

[winter viper]

Same for me, i had to wait days for opus cooldown, but now i guess ill never use opus again

[limber arch]

with deep agentic and skill based generation and aggressive splitting of tasks it did peak

[fair phoenix]

Gotta see it for myself, but also the more you request, it deletes your previous queries on the chat history itself

[fair phoenix]

Sounds like skill issue lol

[limber arch]

so it just takes the last x messages as context?

[winter viper]

does it only if chat is really big, but if u say "read our full conversation again" or " find when we talked about ... in our conversation" then it remembers everything

[fair phoenix]

I guess, it wouldn't go from the start

[limber arch]

isn't dynamic context a good thing for optmizing context and generation?

this is standard in dyad or openrouter for exmple and you can manually select how many of the previous prompts you use as conv context

i think gemini CLI does something like this when it hits compression point

[winter viper]

yes but it also uses extended context features like with better promt understanding it can self choose from needed sources of chat history or smth like that

[sleek wing]

honestly, if it were a project i could do in jules I'd do it in jules 😆 I'm way more comfortable there. also just want to give it a try since my github copilot sub is running out soon and I'm already paying for gemini anyway

[fair phoenix]

"You're absolutely right" vibes

[limber arch]

They have persistent chat memory options yes and you can manage that too if you want

it can be automatic or manual

[winter viper]

naaaah

[limber arch]

personally qwen does it a lot lot better and more automatic

but google catching up is good (personally for me gemini chat always seemed empty and dissapointing)

[fair phoenix]

I remember when I used to rely on GitHub Copilot in the past, it was amazing

[limber arch]

i cannot wait for gemini 3.1 flash

[fair phoenix]

But honestly Antigravity itself is already good for my use case

Before I was using Cursor but not for too long, mainly because it was too expensive

[limber arch]

how does it cost money?

isn't it bring your own brain?

like gemini cli, claude, qwen cli + others

[fair phoenix]

You are basically paying for usage credits

[limber arch]

you are paying to use your own model?

or their model?

[fair phoenix]

Well you have GPT, Gemini, Claude...

[limber arch]

OAuth support for accounts

or use hyper efficeint models via open router which giv eyou 1 mil tokens per 0.3 usd

gpt oss exacto for instance

[fair phoenix]

Especially for the high-end ones it costs so much a single query

and I heard Cursor misrepresents your credits, you end up spending more than what it's actually needed

So people keep spending more money lol

[limber arch]

crazy

personally

[fair phoenix]

Smart business model

[limber arch]

the best agentic coders are the CLI ones

agentic tasks are much better in the CLI

you have the same tools from the IDE + extra

you just require minimal minimai vevry little setup

[fair phoenix]

Oh yeah but I always wondered why this is the case

[limber arch]

no idea, probably because the Agentic IDE are very restrictive

for ex antigravity for me at least doesn't respect rules, doesn't respect gemini.md without me strictly telling it to, all rules and workflows are limited to 12k chars which is roughly 4000 tokens

or less

and ouput seems much more restrictied worse attention too

and some agents are already predefined and unmodifiable

[sleek wing]

It's a preference thing. I personally like to describe things in M sized storys and let jules handle it while I'm away, but everyone does it differently I guess

[limber arch]

Jules?

Also one thing yuo can do is to download Gemini CLI and create a full ai kernel with skills workflows commands (local mcp servers via python localhost for tools and forced skill imposing) and then add them to antigravity and get very good results

[sleek wing]

https://jules.google.com/

[limber arch]

thanks i ll look into it

[fair phoenix]

I used to rely on Gemini CLI until Antigravity launched

Also I have never heard of Jules

[limber arch]

i used to rely on anti gravity until i started using the CLI and saw it was uch better, btw you can use it within anti gravity using and extension for similar functionality BUT (my personal opinion much better functionality)

[sleek wing]

It's both gemini, but different google teams, so different ways it works under the hood than antigravity. I like it, but it needs github repositorys if you want to do it online

[fair phoenix]

I think it mainly depends on the type of work you are dealing with

[sleek wing]

I mostly use jules for things that I do on the go, and maintenance. I have an idea, i formulate it and have jules give it a go. afterwards I look in the PR and decide if i like it

[limber arch]

mine has an expert skil lworkflow and documentation for SaaS and ui design for websites with 200 templates, detailed multi stepped conversation planning phasing implementation debugging etc very interactive OR ALTERNATIVELY predefined deep think and brainstorming and extreme segmentation for long multi stepped planning and research

[olive kernel]

tried gemini 3.1 in antigravity, it was impresive, now it still can't use the toolcall correctly

[south turret]

I start the real work with 3.1 to see how good it actually performs.

[fair phoenix]

I forgot to mention, another very useful thing from Antigravity is that it can open and preview your webpage on the browser and its extension will see it

For example if you were to create a website

[south turret]

57 on AAII for Gemini 3.1 jeeeeeeeeeeeeeesus christ

Didn't think the people at Google could achieve it ngl, underestimated at least one brain over there who thought about something to implement

[fair phoenix]

This is what I was referring to, been really useful for me to obtain feedback and confirm on the details/results

[south turret]

Though let's see real world use, 3.0 was great on paper and dog water in reality

[sleek wing]

well, I'm trying to install my java/npm project now, 3.0 couldn't do it, so if it achieves 3.1 is already better than 3.0 for me

[south turret]

Aight I'll make myself coffee while 3.1 does it's work.
Either I'll be impressed, and my coffee will taste like victory.
Or I'll be disappointed, and there's no amount of liquid despair that will be able to improve my mood again.

[fair phoenix]

So has it achieved?

[fair phoenix]

https://tenor.com/view/budget-cuts-virtual-reality-vr-robots-robot-gif-19350444

Why hire a human when you have Gemini 3.1

[sleek wing]

well, yes but it also installed nvm on the wsl instance although node was already installed

[south turret]

Google ceo right now:

patNotes

[sleek wing]

you are making coffee yourself? what is wrong with you? why haven't you automated that? a facial recognition ai could have already sensed you need coffee and brewed it!

[south turret]

"Wow thank you for having helped us build Gemini from scratch up to what it is today... wow.... you're fired."

[sleek wing]

yesterday I saw that vid about this guy who gave claude a tomato to grow ... I mean if we can grow tomato with claude code I'm sure gemini can brew coffee

[south turret]

My god Gemini 3.1 is slow 💤

bad start

Let me review the spag

So.
It didn't delete the files that were temporary
It left a build_error.log at the root of the folder
It made one small mistake, one big mistake
And it annoyed me for existing

I guess it really is just gemini 3.1

[urban glacier]

I have not tested it thoroughly yet to see how it performs on tool calling but definitely more intelligent than 3.0 and somewhat better than 3 on AG

will see how it progresses and whether it becomes lazy

[olive kernel]

why hire a human if gemini can easily write some nonsense

[fair phoenix]

At some point Gemini will update itself as well

Learns itself

[olive kernel]

asked gemini "why is this part of code error"

gemini learns itself by: Removing the code that was mentioned to have an error

[fair phoenix]

You just gave me an idea

[fair phoenix]

Thought for 161s is wild

[olive kernel]

isn't that normal for gemini models?

they usualy thought for 6mins

[fair phoenix]

🗿

[olive kernel]

claude opus 4.6 is kinda better than this

given the same task, gemini just can't handle it properly

[fair phoenix]

One time it had broke my entire backend and I didn't save on Git

I couldn't recover what was previously

I started crashing out 😔

[olive kernel]

gemini must be told to use the specific tool call for doing the task, i mean.. while claude opus models can figure out what tools they can use, gemini just... can't

it prefer modifying the file using a terminal command, which is.. kinda unusual

[fair phoenix]

Its thought process is usually odd

[south turret]

Claude will defo reach that state because Gemini

[urban glacier]

I git add . & git commit -m "wip" in autopilot every time the codebase is in a good shape and the topic being discussed changes a bit

[south turret]

Oh yeah it's definitively better than 3.0, but definitively dumber than opus

Maybe on par with sonnet (unsure yet)

Sonnet 4.6, without thinking

[urban glacier]

it's been 5 days without opus 🙁

[south turret]

Oof, I'm sorry

[limber arch]

claude is just next level

i still strongly believe that the only reason for which gemini models suck is because they are either hard limited on output length => in lower quality analyses or because they are trained to be more speculative and follow less instructions

unironically they could just finetune a qwen model or use their ultra efficient architecture

[south turret]

Claude is more structured to be honest, gemini leaves files everywhere, it's as if Google doesn't train Gemini with punishments

Claude models feels like they've been forged by steel and blood

Gemini feels like a butterfly that just landed on my finger and idk what to do with that thing, if I move it'll jump somewhere

[fair phoenix]

I don't like Git in general

Why can't it autosave every time you make a change

Why are you forced to name your push request every time

[fair phoenix]

Don't forget, Gemini is technically free for users

[radiant harbor]

let me know if im in the wrong channel.
I've been using google's ai studio, and just a day or two ago the layout of the editor has completely changed. mainly i want to know if there's still a "clear chat" button.

[urban glacier]

what did clear chat do before, not sure if I have used it before

@south turret it started well but it just became lazy again :((

not sure if it's because the 5h hour mark is nearing

will see how it behaves after it passes that ig

I explicitly gave it a full markdown file detailed and it still did some hacks and poc

the reward they give to it for an MVP must be insane

[radiant harbor]

whenever the chat becomes long, or i change directions to focus on something different i noticed the ai had real trouble focusing on the current prompt. So whenever the ai starts to have trouble focusing on the current prompt(it tries to do something that's already finished), then i would clear the chat, and it would work properly again. But the clear chat button is gone, which worries me, because it was a very frequently used button for me. Maybe my app just has too many files and code, and this isn't a common problem for most.

I've also done some testing and noticed some of my apps have different layouts, why is that? For example: (some apps show a "Submit Bug" button at the top of the code assist chat, and other apps show a "Previous Prompts" button.

[south turret]

I'm not a fan of 3.1, it's lazy yeah, and blind

Will I take it over 3.0 though, hell yeah

3.0 can go below the deepest level of hell

[upbeat crag]

poor flash \

[south turret]

I'm curious I saw a Youtuber use "Make the website far more impressive and pretty"
And get decent results
So I did this on my website with a save just in case
Except compared to him it's not just a quick website made in canvas it's a complex website with multiple locales and API use
And even though it is a lot of files, it somehow actually managed to change almost everything in a minimal amount of time
Unsurprisingly, it only tweaked the styling
Surprisingly it made the side closer to "OLED dark" style, which is what the guy also got from it

[lucid idol]

its lazy off the bat for me

[raven sphinx]

How do y'all use Antigravity?

Ofcourse the free tier ends afteer 4-5 prompts

but whats the cheapest way to continue using it? I am content with the quality of responses and work done by both claude & gemini models but which one would be more cost effective, and if so, by the Google AI plan or by purchasing claude plan?

[south turret]

Pro is cheap enough

[raven sphinx]

Claude Pro?

[south turret]

Antigravity is tied to your Gemini subscription

Google Gemini Pro

[raven sphinx]

Oh

[south turret]

Speaking of

Claude used through antigravity, doesn't actually use Claude usage

It uses Gemini subscription

[raven sphinx]

My sister got a year for free when it was promoted for students but I had to be eighteen or some reason I missed the opportunity

[south turret]

Google host their own clone of the Claude models, so they compute and manage usage for their copy

Ah rip

It's still pretty cheap technically speaking considering everything you can do

[raven sphinx]

this looks nice

1m for free

should be enough time to ship my d2c SaaS

[south turret]

Yeah, try that, pay for it next month if you still want it

[raven sphinx]

thanks

[south turret]

SaaS in 2026, good luck

[raven sphinx]

💔

[south turret]

https://www.youtube.com/watch?v=cxcb55zr2Q8 just being realistic

might crush hopes and dreams but I'd rather crush hopes and dreams than seeing yet someone else get disappointed

[raven sphinx]

not rlly

i mean

the video sounds like generic yap

im making a resume optimizer

nobody would code their own

[south turret]

that's just fireship's monotone voice, but he's a serious guy, he's known in the tech industry

I said this before he released his video

it's just logical, why would people pay for a service when at this point they can just make their own, and how they want it

[raven sphinx]

why would they make their own

like

[south turret]

unless it's something way too complex, or very niche, there's no real point
but feel free to make it I mean I'm not stopping you

I gave my opinion, just take it as a bad opinion and move on

[raven sphinx]

would you rather spend 50c on my website or spend hours vibe coding your own

[tribal hare]

What u mean by 50c yo

[raven sphinx]

im indian

50c sounds good for a 1 time use

maybe $2 for monthly subs

[tribal hare]

Iam also from there

[raven sphinx]

so?

[tribal hare]

Which state

[raven sphinx]

50rs for a resume review

?

[south turret]

Indian starting a SaaS

[raven sphinx]

Delhi

[south turret]

If you knew how ironic this is

[raven sphinx]

is this racist or no

[south turret]

No just simple economics

[raven sphinx]

ohh

skibidi dom dom

[tribal hare]

I learning python full stack development

[raven sphinx]

pointless

AI will take ur job

move to cybersec

anyone knows how much the free tier offers?

[tribal hare]

💔

[raven sphinx]

are you in college

[tribal hare]

Yep

[south turret]

It's just that before that "SaaS is dead video" I started a heated argument that lasted 3 hours in the Claude Discord server
I told them that SaaS was dead, some said no, some agreed, and I told them in countries like india where making SaaS would pay very well they will flood the market with SaaS services, so not only SaaS is dead, the market is flooded

So it's just hilariously well timed

[tribal hare]

But i will move on to data science

[raven sphinx]

sigma sigma boy sigma boy sigma boy

[raven sphinx]

claude replacing u in 2026 btw

start selling dumplings

70c/plate

[south turret]

Kinda too late, I have some friends that just started to learn programming, at my age, and I can't crush their dreams but it's soooooo stupid

Like they don't even realise how stupid it is because they can't grasp it, it's way too late

[raven sphinx]

lowkirkenuinely for now atleast

people cant even vibecode basic scripts

[south turret]

They start talking to me about what they learned and I go "ah nice" but in my head that's like "I did this when I was 10"

[south turret]

Not anymore

[raven sphinx]

idk if they dont try

[south turret]

Claude can generate a perfect website one shot now

And like, a good quality one

Like, really good

[raven sphinx]

when antigravity sees Akshay47's free trial end and the next minute Akshay48 signs up for a free trial

[tribal hare]

What about backend process

[raven sphinx]

CORS asterisk ahh code

[limber arch]

Qwen Alibaba is superior in all aspects more frre and offers more. It's not about being free

[limber arch]

Are you sure? I believe they are just routing user requests through gemini server to Claude via Google s grand anthropic api key. If Google had access to anthropic models they would fine-tune them or distill of off them aggressively and we would have better models.

Gemini claude gpt models are proprietary and guarded deeply

This is why they the usage is so expensive 25 $ per 1 mil tokens vs glm 5 which is Claude distill and costs like 3$ per million tokens

Glm is pretty much Claude opus 4.5 btw

But open source

[south turret]

No can confirm this is verified information they have a service I forgot the name, Vortex or some made goofy word like that they call where they host other models

Also, no, GLM performs good on benchmark but realistically nothing comes close to claude

[limber arch]

And they said word by word we host Claude?

Locally

[south turret]

Yeah, they showed they hosted Claude

[limber arch]

Not that it is available did they say the model provider is on Google servers?

[south turret]

You can just check for yourself find like google vortex or something starting in V

...

Yes Google host Claude on their server

You can just use Gemini or whatever to fact check and find information

Vertex AI Platform: Google offers Claude (including the latest models like Sonnet, Opus, and Haiku) through Vertex AI, which is Google Cloud's fully managed enterprise AI platform.

[limber arch]

You're right that's crazy

[south turret]

Yeah I was wonder why and how would Google provide Claude on antigravity, but then I found that out

In my opinion Anthropic messed up big time

Antigravity would be practically dead without Claude

Maybe they got some free computing out of it, or a big pile o cash, either way, stupid

[raven sand]

Ai studio working for you ? 👀

[red vessel]

Did you ever figure this out?

[upper tulip]

Yes, its not supported

Atleast when i tested then 🙂

[red vessel]

So unfortunate. Was hoping the deny list feature would actually work

[raven sand]

still the same issue with Ai Studio from my side. Need to wait apparently

[jade moon]

Hey everyone, hoping someone here can help or point me in the right direction.

One of my projects ("HealthCheck") in Google AI Studio is completely frozen on the 'Preparing system resources' screen after a recent migration.

Looking at DevTools, it seems the migration completely broke the link to my files on the backend. I'm getting constant 404 (Not Found) errors on ApplyFileSystemOperation API calls.

I already checked Local Storage and IndexedDB hoping to rescue my code, but there is No indexedDB detected and nothing is cached locally.

I unfortunately don't have a backup for this specific project. Has anyone managed to recover their code from a state like this, or know any magic tricks to extract it? Any tips, workarounds, or help from the team would be a massive lifesaver Attaching my console errors.

[rain hearth]

i cant run any apps, because of this button. I cant remove it

can someone help me?

[limber arch]

it's not that bad

like i am not a fan of anti gravity but it's not awful and absolute 0

gemini is not amazing for complex long databases or debugging (unless you use deep agentic tasks and skills and manually plan everything in 20-40 turns BUT it is a much better experience than pure browser chat copy paste

[cosmic ferry]

why does antigravity ask for Command running permissions even though i have set it to Always proceed for terminal commands as well?

[fair phoenix]

Not in my experience tbh

Qwen is just meh

[fair phoenix]

Most likely because it's doing a different command rather than the same one

[cosmic ferry]

but why

i had agreed to allow all commands

[fair phoenix]

That's odd

[limber arch]

how

like what tasks did you use it for?

even for coding long deep and complex SQL scripts it is better (gemini failed multiple times)

[fair phoenix]

I don't remember lmao

[limber arch]

bruh

[fair phoenix]

But it was quite some time ago

[limber arch]

use recent models

not qwen2 series

because those weren't amazing like all other models weren't

i can show u in private some amazing outputs and materials i had created with qwen in one or 2 generations

and then you will ese how crazy it is

[fair phoenix]

Sure I'm curious

[urban glacier]

I wonder if google's MoE will win eventually

will have to see if dense models hit a scaling problem soon

[hallow spindle]

I bought Google AI Ultra. Google AI Pro was not enough for me because I use antigravity in a project maintained by 20+ developers, a huge monolith backend API.

Because of this and because of the scale of the features I got it to do, I had to use Opus 4.6 for context size.

I do planning & coding with Opus 4.6, and then I switch to Gemini Pro models to write 2 different documentations, 1st: for humans. 2nd: for LLMs.

I made an habit of creating documentations after each session, this way when AI or a human needs to know something, they can basically search through docs. Of course, this is basically because the projects I use it are huge. might not be relevant for a small project.

[raven sphinx]

cooked

[south turret]

The only reason I don't buy ultra is because I'm scared usage will only barely increase

[hallow spindle]

• upload docs to notebookLM, share it to your boss or QA etc, and you get rid of unnecessary communication with leads & QAs

[hallow spindle]

honestly, it refreshes every 5 hours and believe me, I make a good use out of it every day and never got rate limited so far

[south turret]

I seriously doubt I won't get rate limited

[hallow spindle]

A tip for Ultra, to trigger usage just send a basic prompt 2.5 hours before you start work.

it starts countdown for 5 hours rate limiting reset. so then your work starts & you use 2.5 hours, get a reset just in middle of your work, just in case you are near the limit

[south turret]

Yeah I do the same with Claude

[south turret]

Gemini answered in like hebreu or like arabic or something, and I can't downvote the message because the thinking never stops, even if I press stop

And if I refresh the page, with "do not keep conversations", the conversation is gone

[urban glacier]

they already nerfed it

[fervent patrol]

how to get an internship real quick

[south turret]

find someone who works at google and subdue them

literally how everyone I know works at google

it's like nepotism and cronyism

[red vessel]

Why are limits so crazy rn? Used Sonnet on AG for 3 hours for the first time this week and my limits reset in like 3 days? A little shocked for pro plan

[winter viper]

dang that problem was easy to solve

gemini be trippin

[south turret]

Now THAT's how you optimize something

Gemini, download the linux repo, refactor it

[winter viper]

he basically almost deleted full web page file

[south turret]

Make sure to use git, do not trust AG's message system

[urban glacier]

optimized for 0 users

[winter viper]

fr