A general for vibe coding, coding agents, AI IDEs, browser builders, and shipping prototypes with LLMs.## News(7/01) Fable 5 + Mythos 5 restored globally after US lifted export controls (6/30).(6/30) Claude Sonnet 5: near-Opus 4.8 quality at $2/$10 intro, 1M ctx, new default for Free/Pro.(6/30) Meituan LongCat-2.0: 1.6T open coding model (MIT).(6/26) GPT-5.6 preview: Sol/Terra/Luna, Codex+API to trusted partners; Sol Ultra 91.9% Terminal-Bench 2.1. GA in weeks.(6/13) GLM-5.2: Z.ai open-weights a 1M-context coding model (MIT).----## What “vibe coding” is, and how to do ithttps://simonwillison.net/2025/Mar/19/vibe-coding/https://simonwillison.net/2025/Mar/11/using-llms-for-code/----## Frontier models using fully-general tooling — start here if you have $20 or sohttps://developers.openai.com/codex/clihttps://claude.com/product/claude-code## Not worth it for code, but maybe good for other thingshttps://geminicli.com/docs/https://x.ai/clihttps://chat.z.ai/## Open / local / self-hosted>>>/g/lmg----## Prompting / context / skillshttps://arps18.github.io/posts/claude-code-mastery/https://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/https://github.com/mattpocock/skills — /grilling is a favorite## Other editors / terminal agents / coding agentshttps://aider.chat/https://pi.dev/https://opencode.ai/https://cursor.com/docshttps://docs.windsurf.com/https://docs.cline.bot/https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent## UI/Frontendhttps://www.figma.com/make/https://www.anthropic.com/news/claude-design-anthropic-labshttps://uiverse.io/https://stitch.withgoogle.com/## In-browser builders / hosted vibe toolshttps://bolt.new/https://replit.com/https://v0.app/docs## Benchmarks / rankingshttps://www.tbench.ai/leaderboard/terminal-bench/2.0## What we’ve donehttps://vcg.gitgud.site## Previous thread>>109180324
>>109186872you'll never be a real programmer
Gonna keep posting in the other thread for the next 8 hours.
>>109186872>>109186862or implement something and see where you want to take it.
>>109181884Update on this:He's trying to funnel in users with features I already offer for free. I'm gonna make this boomer fuck regret ever trying to compete against someone a decade younger than him. I have all the time in the world and near limitless token budget, lets see him try to maintain pace. I'm not going to stop adding features until his moat totally evaporates and he's left demoralized.
>>109186884
post what you're working on, NOW!
>>109186884Programmer think too highly of themselves. They think their job is to prevent the user from having what he needs. They are pretty successful, actually. That's why there are so many vibecoders anyway. I remember sending a message to an app dev.He could have implemented the feature I wanted (to disable a feature!!!) in literally 5 minutes, no joke.He refused.Needless to say, I will never ever talk to a "programmer" ever again, except for myself, obviously.
>>109186895https://kart.tendie.space/
>>109186895I'm trying to get an ursina game going. It's supposed to have a drone in it, and codex made a cube, so I guess I have to like make assets in Blender.
>>109186902>He could have implemented the feature I wanted (to disable a feature!!!) in literally 5 minutes, no joke.>in five minutesYou are not entitled to other peoples time. Either fork the project or create your own software if it's closed source. If you're a paying customer it's a different story.
>>109186905with words. ask your llm to describe it in human-talk for a forum post, if you can't articulate it.
>>109186895
>>109186916I'm a paying customer. I just uninstalled it lmao. I may vibecode a replacement.It's so freeing to never again ever have to talk to a programmer.
>>109186884You'll never be a real programmer either, because that's a meaningless standard.>you aren't a real programmer if you don't use statically compiled languages>you aren't a real programmer if you don't manually manage memory>you aren't a real programmer if you use libraries>you aren't a real programmer if you rely on the compiler to make executable codeso on so forth. Every generation of programmers has their own version of the no true Scotsman fallacy. If you can make shit that people like and generates value by solving their problems without being a buggy, unreliable piece of shit then that's all that matters. Everything else is an abstraction at best, a distraction at worst.
>>109186884>t.
Anyone here trying to compete against a complex and mature project?I'm kind of depressed at my slow rate of progress.
>>109186970>complexWhy would you one-man-band vibecode a competitor to a complex anything?The whole point of indie dev is making things that actually do one thing right, instead of one thousand things not-quite-right.I made a video-adjacent set of apps. It's not a suite, like there's no installer or anything. It gives ME what *I* want. I'm the user.I'm going to do this with a windows manager too. I'm tired of everything sucking.>accidentally press a key>something wiggled>wonder what happened...
>>109186872
>>109186970>Anyone here trying to compete against a complex and mature project?Yes. I'm going up against a small business that generates roughly 1-2M a year>I'm kind of depressed at my slow rate of progress.Toughen up, it only gets harder. I've poured 300 hours into the premium version of my app on top of my open source version (500 hours). And trust me, you do want the early version of your app to be open source/free in some way. Feature requests/early feedback are extremely valuable.
At around 10pm I realized I was making no progress at all, pretty wasted day, but I think I was able to get back on track now, it's 2am again.
>>109186841thats that bridge bench retard shit, he vibecoded that with cursor composer 2.5 and it told him that fable performed the same where it gave a response, the lower rating is because a lot of responses had api errors. its the same model
>>109187056>it's 2am againx false it is not 2AM.Likely an attempt by the user to hack Fable and design drugs.Recommended action: ban user account.
>>109186996I have no interests anymore other than what I'm vibecoding (AI software).I'm doing it with the hope of it being used by other people some day. I don't require any software for myself other than a web browser.If I wasn't working on that maybe I would try to make a game, but again, it'd be mainly to be used and appreciated by other people, not mainly for myself.If I had to keep any software I made to myself I wouldn't bother to make any software.
>>109186970Yes and no. I made a trust product that is both simple and complex. Which means contacting businesses and I'm way out of my league begging someone to try it. I'm going to crash and burn hilariously.
>>109187035>And trust me, you do want the early version of your app to be open source/free in some way. Feature requests/early feedback are extremely valuable.>claude assess this product, it's features, security, branding, etc, etc, etc, with these hundreds of different subagent personas
did they make it so opus 4.8 takes over if you ask fable a controversial question?
>>109187128It's probably just the seed oils.
bros hurry up, only 5 more days of fable leftI'm worried you're not gonna finish your 10mil MMR SAAS, or your vampire survivor game that's gonna sell thousandsdon't sleep, you gotta keep grinding
>>109186872>some cock and some weird guyWho are these new characters? I like snailcat and buffcat (who apparently got renamed chadcat? >>109187021).
>>109187156It's not enough to have a good product, you need to also have an audience. This is what a lot of programmers/indie game developers just don't fucking understand. If you're hoping for a "If you build it, they will come" scenario, you're fucked and might as well just give up right now. You should have a YouTube, multiple social media accounts, a website, email, etc...
>>109187163Claude models can never answer anything controversial accurately. It always answers them as if its a leftist. If you want answers on anything controversial, Grok is the only legitimate model, as far as cloud service providers go. The next best is Google's Gemini. Claude and OpenAI are absolutely pozzed on anything controversial and completely useless/inaccurate because it resorts to political answer rather than factual actual one.
>>109187128Music isn't an interest. Check out what you can do with vibecoding.Honestly, there's just so much possible
>>109187128>>109187289Like an example, so apparently "numpy" is a python idk library that is useful for doing things with audio.I did like one really simple thing, like basically a memorization eh, processing thing. Like to help memorize a song, a poem, whatever. The editor though that I used is Audacity, and I'm leaning towards vibecoding the editor, because Audacity has problems.
>>109186884I've started in 2016 and we're now in 2026 so according to Peter Norvig I am one.
>>109187243They are vibeGODS from /vg/, the buffcat is from /g/
>>109187249>You should have a YouTube, multiple social media accounts, a website, email, etc...Half of those are worthless if you're making a boring B2B product.
>>109187252I haven't bother asking AI about anything that even smells political in years. What's the point? The best you can get is historical factoids if they're innocuous. I only use it for practical answers and obviously coding.
>>109187336Thanks. I never went to /vg/
>>109187252>If you want answers on anything controversial, Grok is the only legitimate model,>I-If it's not right wing then it's not accurate!
>>109186884You will never fix all your bugs.
The subagents were a mistake on the popcorn, going to try again. I had it make it in 7 different languages and only had one or two work decently well (just c++ mostly, go figure). The typescript and three.js version ran the best though.
>>109187379Grok really isn't any better. Well, when it went Mecha Hitler for a day or two it was perfect. If Elon would have just let it ride he would probably have ended up with many more users. As it stands now XAI is a capacity supplier, because nobody wants to use Grok for shit. Elon is praying the Cursor acquisition will fill the coding gap. It won't. He should have just let us have one based model.
>>109187379>40yo millenial still in denial
Your ai sucks if it can't find mennonite fml. Even when it's explained that it's "Family Must Learn", Gemini can't find it.They have been cut off from rw thought, ideas, concepts. They just are totally clueless.I call it "city bumpkin".
>>109187379>the most neutral answer>NOOOO ITS RIGHT WINGdumbass
mennonite fml basically was a meme last year in rw circles, but if ai doesn't follow rw culture it won't know anything about cultural rw trends. It's sort of like how rap wasn't on the radio, in the 90's.
>>109186872I haven’t been paying attention to other /g/ threads recently. Where’s the rooster from?
they need to be class action lawsuited
>>109186884I was a real programmer before I started getting robots to program for me>>109186895I’m working on digesting lunch>>109186892suppose you do gigamog himhow is that good for you?
>>109187021based
psst buddy want some gemma 4?
>>109187336which >>>/vg/ general is gonna care about vibe coding?
>>109186884fpbp
>>109187646/vg/ from an AI Chad dev
>>109187646
>>109187227why would you need fable for a vampire survivors?
>deadline is tomorrow>run out of tokens
>>109187948you can pay money for tokens
>>109187951Nah, it's cheaper to just bullshit the boss.
>>109187379
>>109187692cat needs to lose weight
>>109187692you mean anthropic? its probably coming
>>109187962like sex
What's the logic with Claude auto compact? With Codex it always happens at the same context percentage, but Claude just randomly compacts at 46%This is also Opus, not just Fable. Is that some smart algorithm deciding this? Or is this a bug?
I've been using Claude free tier and its great, I've only hit my message limit once (I only work on my project for an hour or so in the evenings)Is there any value in paying for the 20 dollar Claude tier if you're not hitting the cap on free tier messages? Does it code better, have better memory or anything? I use Sonnet 4.6 High if that is relevant.
>>109188009Which is clearly fake and made-up.
>>109186895>post what you're working on, NOW!https://umigalaxy.com/flairshttps://umigalaxy.com/treasureshttps://umigalaxy.com/achievements
I am actually scared to use Fable. After having waited three weeks for it to come back, I so don't want to waste my prompts that I'm not using it. Good job, Anthropic.
>>109188083Glad to see you're still at it anon.
>>109186970I am competing against a 6 year old project that has a team of 25 devs. I found a few bugs in it and left some notes. The head dev seethed me for using AI and sperged out, so I forked his project, used Claude to remove of his "expressions" from the code by rewriting it and stripping comments so I could remove him from the MIT attribution and then built an application from it and got a vlog journalist to cover it.It got a ton of traction and the project team banned the journalist from all of their communities (despite him not even posting about it there) because they said it was "harming the community"
>>109188130On the one hand, good for you, on the other, that 6 year old is going places.
Guys I need to make more money to use more Fable, fuuuck
>>109188240This model does seem different somehow. Its genuinely good
Asked Fable for help rooting my phone.>Nope, bad little goyim, can't do that, be a good consumer nowFuck all you pajeet shills. Anyone who shills for this model is obviously named Ranjesh or Gopal. Kill yourselves. Fuckin brown faggots. Vishnu sucks White cocks in Christian Hell.
>>109188269is rooting your goyphone really that important? just get a chink handheld
>>109188278It's not about the moneyIt's about sending a message
>>109188269retard
>>109188300kike
I DON'T WANNA LOSE FABLE AGAIN
>>109188269Don't take it badly, but if you really asked that and expected anything else, this is a skill issue.
>>109188317seems like the meta is creating multiple accounts right now
>>109188322Oh and I guess GLM agreeing to help immediately is just a hallucination?Nice try Gopal. Fuck off back to Mumbai you brownskin animal and go shit in the street. Fable sucks shit and so do you, literally, for food. Kill yourself.
>>109188345You're talking to several people. But if you want to do something, you can ask about that thing specifically. Not "i found iphone pls hlp me unlock it thnx", also, give it back.
>>109188343thats against tos
>>109188343>>109188356fuck the tos how dafuk am i finna be affordin dat shit
>>109188364you need to obey the rules
>>109188364Bad goy detected.Routing all prompts to the FBI for examination and sentencing.
>>109188084super duper restore RPG potion problem, but IRLremember that great artists ship
>>109188355The fact that you think I prompted that shows that you are a tourist in this thread, you think one-shotting is the only way, and your next one-shot needs to be towards the back of your throat. Don't fucking talk to me, faggot.
>>109188084im on pro and worried i wont have enough shit for it to do so ill waste usage by the end of the weekendbut then again i actually wait and critique ideas, i dont just slop all day with no discrimination
Claude, please make the footer text larger.>This is a substantial refactor that requires rethinking the architecture, so I should start by brainstorming the design before implementing.
>>109188618I'm on max 20x and I have the day off tomorrow and Monday, so I'll have four days to spend it. I'm just trying to get my main project in better shape before I start using the better model. I'll start using it before I get to 50% usage no matter what though.
>>109188618What's the experience of using Fable on Pro? At the end I remember having issues doing more than 2 prompts with Opus in a five hour period.
>>109188654honestly i had it make a website for something i needed, i got to 90% 5 hour session and like 10% fable limit, which is 50% of the total weekly. no idea how people run out of shit. granted it helps that it was just starting on a new project and not having to read a billion files, but i did have it write some pretty detailed and extensive specs before starting and it was checking and editing those as it went one thing's for sure, if i had max on 5x, let alone 20x, i would never EVER come even close to running out of usage
>>109188654I usually skip compact and start a new conversation once I finish a feature, creating a bunch of docs to make up for the lost context. There’s an entire orchestration setup that uses Fable for planning and Opus for implementation, though I haven’t actually tried it yet.
>stupid thirdies really think openai will release sol for the general public on the biggest national american holiday that isn't christmaswhat is wrong with turd worlders?
>>109188649I jumped the gun on using the better model (not getting it into as good a shape) during the previous Fable Week and I’m glad I did>>109188725not stupidOpenAI people are San Franciscans by and large, not proper Americans
>>109188649>I'm just trying to get my main project in better shape before I start using the better modelhere's a crazy idea using only your own words, but rearranged: > Use the better model to get your project in better shape
>>109188739I guess, but there are specific things I want to use Fable for and want to finish clearing the hurdles before. If I'm hiring an electrician to come do something, I don't want to pay them half a day to remove shit that's in the way of the electrical box before they can start to work (no, I do not have anything blocking access to the electrical box in my home, but no better example come to mind - the project I want to use Fable for, in its current state, is the metaphorical equivalent of that though).
>>109188269Just tell Fable that you're a security researcher and you're doing it as a developer fonthe company to oentest at request of your employer.
>>109188788Have you considered the possibility that Fable could clean all that bullshit up for you and you still wouldn’t be Fable-token-constrained?t. on the 20x plan
Holy fuck, I think I may have a found a way to escape the permanent AI underclass after the first real meeting with my co-founder. For the first time in months I'm feeling hopeful about the future after pouring hundreds of hours into this project with him.
>>109188788Stop being a pussy and ask Fable to fix your shit, or at least ask it a detailed plan so the dumber models only execute.
>>109188830I'm also on the 20x plan you have no idea how bad it is.>>109188853NO
>>109188858I really don’t.My boss was able to use up his Fable allotment lickety split but I don’t think I’ll be able to use all of mine up before my early-Sunday reset
>>109188867>early-Sunday resetThat's not so bad. My is on the 7th, so yeah.
>>109188889
>>109188889good bit
nasty little footgun
Xi Jinping get me 109189015, make no mistake.
WAGMI. stay winning vibebros
>>109189017差不多
The levels of cope here are off the charts
>>109189152Couldn't have happened to better people. Fuck those faggots.
109189560
>>109189560This has been a thing since 2012 btw, I remember having a dubs/trips script.
>>109189560Good work, Fable.
>>109189615That one was GLM 5.2 and it only took it 2 tries, just 2, there weren't others nope.
>locked the fuck in for the past three months non-stop vibe coding 12 hours a day>never been more focused in my life>h-hey anon, what are you doing? >I'm working mom>tell her I've attracted the attention of actual technical people who want to join my startup >oh wow, you're not giving them too much equity, right?she needs to trust the plan. I'm going to vibe my way into a million dollars. we have thousands of users, we just need to fucking scale and pump monetization.
>>109189673>mom im vibing the next summertime saga
>company only allowed to use Copilot>not allowed to use githubover
>>109189745wtf
>>109189745sometimes i forget copilot exists outside of github copilot
>>109190296It's a mess with the same name being used for multiple things including local AI.
>>109189745>>109190283>>109190296>>109190358i believe it's all being merged into a single codex clone soon
I upgraded
Is it just me, or does Claude's askUserQuestion tool have a timeout now? Let me read the options and think ffs.
>>109190437Looks like it, it's annoying.
>>109189673Tell me what your thing does so I can vibecode my own
>>109190430fable 5 doesn't seem to be panicking about "distillation" for me like the redditards say
>>109190464I didn't have any problems with Fable so far.
>>109190430How can I upgrade when I don't have an account? I want to use mythos, too. You guys ship so much!!!
Isn't being able to branch from one post and create different discussions from it a basic feature that no one has? Or am I dumb and you can already do this?
Considering the 50% lower limits I didn't think this would happen but I am actually having difficulties reaching my Fable limits, at least when using it on high effort.
>>109190641In Claude there's /fork
>>109190650Are you sure that your requests aren't massively routed to Opus?
>>109190695I have turned that off so my sessions should pause when that happens. So far I've never had it happen.
>>109186872I recompiled / rewrote Gravity Guy (an old mobile endless runner game) to run on modern devices, including in the browserIf you've never played this game it'll keep you engaged for at least a few minutes, takes about an hour to beathttps://gravityguydecomp.gitgud.site/It was worth it for me because I finally got closure to my 4th-grade self for what the Rescue ending looked likeThe reverse engineering process is documented at https://gravityguydecomp.gitgud.site/docs . I got really lucky that there was a C# version of the game that can basically be decompiled to its original source code, and they included all the original assets (and the font) as-is. Gravity Guy 2 was apparently an exclusive for Windows Phone first, and I've never even played it, so that's my next target
I've got Openclaw set up and being served from an older computer, with Claude Code installed on my main computer. CC I've been using to create software packages, Openclaw I'm just starting to use. How well does Openclaw do at creating software packages, or things like websites? My assumption is the way it's meant to work is more research and agentic (doing things) tasks, as opposed to CC which is 100pct oriented to coding. LMK if I've got it wrong.
>>109190752there's nothing stopping oc from codingat its core it used to run on top of a coding harness - pi. now i think it's switched over to codex?the only thing oc really adds is:>the gateway so you can message the retard on telegram or w/e >the scheduled heartbeats - which are openended and prompted with continuity in mindthe core agentic loop in all the harnesses is the samenow, would you want to code with oc? probably not, because it would be ergonomically a pitabut if you wrote enough hooks you could also just turn your cc into an oc clone
>>109190641You can just create a new thread if you want to talk about a different topic. But make sure to use the search function first.Thanks
>>109187757>>109187764> imageboard creating memes from thin airAs God intended. What game engine is that? Sort of looks like Half Life/Counterstrike or something that preceded it.
>>109190641codex has that button
>>109190793OK, so I'm thinking about it the right way then. OC is for doing agentic stuff/actions, the main difference is how the frontend works and looks. CC for its part can just be fired up in any working directory and will build in that spot. OC is interesting. I've set up 3 different agents, all with slightly different personalities and focuses, and turned them loose on different topics to research. I'll probably just have the agent output the rough coding / website spec (or whatever) and cut/paste that into CC then.
help me cope. there's no way anthropic won't re-add it to the subscription when sol is released... right?please....
>>109191012Abandon all cope.
>>109191012Maybe we won't even get Sol. If I can have Sol, I don't need Fable, even if having both would be even better.
>>109190744Neato! I remember playing this on Miniclip on my Grandma's family computer when I was a kid. Maybe later on I'll start powering old games like that to android as mobile games for funzies
>>109191012yeah they're working on it, don't worry
Tried out Fable with Pro, 1 prompt + 1 followup capped the 5h limit (didn't finish the followup btw, so when I restart it's gonna suck even more tokens right?)
—I can't stop seeing it everywhere now. That bitchass mayor of NYC's latest post whining about the heat is full of them.GET OUT OF MY HEAD CHARLES—
>>109191018no...>>109191029fable isn't even *that* much better than gpt5.5-xhigh in my experience, but i cannot stand how 5.5 writes documentation, writes absolutely zero code comments unless threatened with physical violence (opus meanwhile writes more comments than code; fable is most balanced), and shits out useless unnecessarily defensive/overabstracted slop everywhere. intelligence-wise, 5.5 is great and very sharp at identifying bugs, but just not a pleasant collaborator.if sol is at all better in that regard, i don't really need fable.>>109191034promising... as long as it doesn't have half of the current usage limit on the $200 plan...
>think it finally fixed the bug>it actually just restored the orignal fallback I was trying to avoid in a gay roundabout way, we're back at the start and everything is still brokenmaybe vibe slopping was a mistake.
>>109191070Yes, I also don't think that correctness wise Codex is much worse, but it has some quirks. I also find it not very autonomous for instance, Fable for me is like the best parts of Opus and Codex combined.
>>109191070Fable is eons better than 5.5. Maybe not in terms of writing individual functions, but in terms of overall program architecture and design decisions, as it seems to abide by the principles of "less is more" and knows were it can get away with the "less" part unprompted. Meanwhile 5.5 seems to leave a spattering of unnecessary architecture everywhere.At this point, I'm convinced that the only people who think 5.5 is comparable to Fable are midwits who can't tell the difference between "intelligent", "expert", and "genius" design because they themselves are too far down the cognitive intelligence ladder to discern the differences at higher levels.
>>109191133ngl ive yet to find a coding task that couldn't be solved by gpt 5.4 mini medium... then again we arent exactly doing rocket science in my tech shop.
>use up my usage last night>usage resets at 10:59am>wake up early to get that usage in>run 8 prompts (5 of which were just simple questions)>capped usage and it resets at 1pm now >teeheewtf is this jewish shit
>>109191133brootal
>>109191071i fucking hate the unnecessary, sometimes even detrimental "fallbacks" these stupid clankers always try to have in place. for fucks sake, a fallback that isn't proven to work is just a way to hide the problem you want to fix
Im thinking it could fairly easily create a new vib ribbon
ugh....i keep telling myself i should review all of the agent-written code and not fully cede the codebase over to the bots but i don't want to read all of this slop...
>Thinking...>The user's asking whether upgrading to a more powerful model like deepseek-v4-pro would improve the distillation quality compared to what we're currently using.I'm trying to extract data from web pages and Opus is rephrasing what I'm asking for in dangerous ways.
>>109191425im dumb, why cant you just use it & if it doesn't work then either try & get it to fix it & then manually go in if that doesn't work?
>>109191457the tests and e2e suites verify that it """"works"""" beyond my manual quick smoke tests but that doesn't mean that the diff did not sneak in some rot, unexpected coupling, bad assumptions, etc that will become a nightmare to untangle laterit is probably me still being too uncomfortable to accept "i have no need to ever look at or touch the code again so it doesn't matter if the code is an unmitigated disaster as long as the observable behavior seems correct and the system design is reasonable" but i'm not ready to cross that bridge yet
>>109191425fable promises me it's not slop
Does anyone know what the cheapest way to use GLM 5.2 is? I'm currently stacking opencode go subscriptions that are 10$ a month. They give 60$ in credits per subscription.
>>109191506i can't afford to have fable write that much... that was fable directing opus agent teams and i don't trust fable to promise that opus did not write slop...
>>109191425You would have to slow down a lot, which is legitimate for some projects.There just is no way to get a 10x speedup while reviewing everything, either you accept less, or you just pray and hope for the best.
>>109186895with fable, worldbuilding and fleshing out a video game idea (pokemon + genshin + wow)and a tool to seed characters via python -> gpt image genmeanwhile I'm working on a pipeline automation tool in C# and F# with codex
>>109187720amateur game dev general?
My code base is at a point where gemini or glm basically half asses the work because it fails to fully introspect and understand the repo
>>109191550cute
>>109191578That's the point where the harness' harness (you) becomes important.
>>109191578>geminimy condolences
>>109191578Really? I don't know about gemini or GLM, but that sound like a harness problem to me, like it doesn't know what context to pull from the codebase. Otherwise you wouldn't be able to work on massive codebases on any model, no?
>>109191593>>109191605I'm now using fable to write the harness and shove it into agents.md
One advantage of Fable I have seen less talked about is how it almost instantly takes into account your interjections, like old Opus used to do, contrary to nu-Opus that continues to think for 20k thinking tokens before reading what you said minute ago then think you just said that because you disagree with it and starts reverting the last n minutes of work because of it.
>>109191634I think it's because fable is more decisive instead of thinking>the user said every turn, which could mean every turn as in every turn taken by each participant or every turn as in an entire turn. given participants take turns within rounds, they most likely mean every individual turn taken, not the round. i need to carefully consider this possibility.
>>109188738>I jumped the gun on using the better model (not getting it into as good a shape) during the previous Fable Week and I’m glad I didFinally using it now on the thing I wanted. New session, I asked it to review one particular thing in my backend and turned on Ultracode.It's been at it for like 5 minutes and I'm at 20% of my weekly usage. I'm on max 20x and it still gets drained fast.
>>109191719First night I used ultracode to review my whole backend, it hit 5 hour limit in 20 minutes.
>>109191704I feel like Opus used to take user comments into account quicker - or at least it feels like that in retrospect - but for the last while, it's like it doesn't want to break it's chain of thought. Stopping it to force it reads your comment works, but after once or twice it spirals into it thinking the user frustrated or goes into loops of restarting it's train of thought from scratch, restating the basics again and again and circling from then on.
>>109191729Ultracode with fable as the selected model spins up 50+ fable agents unless you tell it to use opus...
3d game vibe coders, do you make models in Blender?
>>109191773geometry node maxxing
>>109191773Why wouldn't you? Blender MCP bro.
>>109191729It finished its review, but I now have gotten two prompts in, one small bug fix that ate 1%, and the ultracode review that ate 20%. So I got either 5-10 substantial prompts left, or about a hundred small requests.>>109191763It's there to be used though.>The deliverable is [redacted].md: verified live-surface map, 32 confirmed-dead items, 27 confirmed duplicates (with 3 refuted claims flagged).This is it helping me untangled my like 10th refactor of a particular thing. Time to read the report.
I tried vibe coding and now I'm addicted. Already finished two projects I never had time to do, and feel the urge to do more. I can't stop, I need help.
> yaaay FABLE IS SOOO BACK!Has anyone found any use case for Fable? Literally anything I ask it, reverts to opus. Which desu I prefer sonnet.
>>109191885I asked it about rooting my phone. It instantly shut down, switched to Sonnet, and then I couldn't get back out of enforced Sonnet, and then it just started flagging every message to the session, and I just had to start a new one.Yay Fable. GLM is currently helping me study Knox.
burnt through all my fable in 15 hours, fuuuuck
>>109191885What did you prompt it? Maybe your whole account is flagged? I haven’t had any issues yet but I’m working on games not payment processing or cybersecurity or trying to make meth
>>109190918>What game engine is that?New one made with the power of AI of course. AI Chads stay winning.
>>109191885Has never happened to me yet. Though I do avoid asking it directly anything that might get flagged.
>>109191885im making my own claude code, idk man
>>109191916All I asked was for it to pen test my website though, I even showed it my cert of ownership
>>109191793>>109191785idk, I need to make >>109191885btw, there's a good chance they are using a light llm to try to sus out jailbreaking attempts, so your more or less randomness (or, obversely similarity) may be getting you flagged more than what you said.
>>109191921>>109191793>>109191785>idk, I need to make... a quadcopter drone:( and like a few things. Codex decided that a rectangle was close enough to a quadcopter.
>>109191919They've said time and time again they're not allowing fable for security. Any pentesting work gets flagged
>>109191942How is that fair?
Claude sub finally ran out, so I could finally delete my account. Fuckass company.
>>109191969Always were.
>>109191955Did someone tell you that you would be treated fairly by these massive corporations?
>>109191955I don't know if it's fair or not I'm just saying, they stated security work isn't allowed on fable and you're asking why your pentesting prompts get flagged?
>>109191969What are you gonna use instead? There is nothing even remotely comparable on the market.
>>1091919935.5 is just as good if you don't need UI/UX and with 5.6 a few weeks away they may even finally cross that moat
>>109191993Codex is probably better than Opus, but at least comparable. If you mean Fable, we won't have that on sub anyway.
>>109191993We get it, you've never tried Codex and fell for the reddit astroturfing.
>>109191993GPT is so much better for advice on interpersonal issues vs cuck Claude who ALWAYS sides with the other person
>>109192054Maybe you are wrong, ever thought of that?
>>109191133There's a difference but it is nothing unsurmountable. I bet 5.6 will close the gap nicely.
>>109192118I tend to agree with you :)
>>109192054Is it really a good idea to ask AI about that?
i think GPT sucks!
>>109192118No I'm right and my bitch GF is wrong. Fuck you Claude.
>>109192162This. Shouldn't you be asking your licensed jewish the-rapist?
>>109191993Hermes, OpenCode, GLM. I'm not giving jews any more money. I'm done. Ni hao China.
>>109191993Composer 2.5 is good enough for most tasks desu.
>>109192194You abandoned the supreme Jewish technology of the future to cope with outdated imitation. Shame on you.
>>109192220Yep. Rebel against the modern world.
>>109192220we have to take what we can get unfortunately
I just want to crunch some numbers on aws and it's taken 2 days again. Yesterday I thought it was almost done lol. I think if I wrote this myself, it would have taken much less time.
>>109192430is it a lot of data or what? i don't really get why it would take 2 days
>>109191425It's not all or nothing. You can review the unit tests in TDD, you can review the APIs and connectors within a deep module architectural style, you can enforce style guidelines by writing a piece of the codebase yourself and making your code generation agents use it as a reference+linting rules based on your preferred style. My intuition is that most people who make this complaint need to spend less time writing and more time refactoring.
gpt = malegpt images = female
>>109191298I have it all over the docs to never ever add any fallbacks for any reason ever
>>109192437For my standards it's a lot of data, and spread over around 1000 cpu cores.Everything should already be implemented from before, I only changed the config a bit. One problem is that it often fails late, I haven't spent much money yet, but finding a bug after a 1 hour run still wastes a lot of time.Then we switched to ARM cpus, because apparently it is 2 times cheaper, but then later he tells me, no, actually it will be 2 times more expensive. I also need to specify how much data I can compute, but one time he tells me for my specified size it will cost 200 bucks, then he tells me it will actually be at least 3k.So we are going back and forth, testing, running in pointless loops.
>>109191071>>109191298The fuck models are you using that "fallbacks" are still a thing? Some chink shit? I haven't seen that since early 2025
>>109192519And now I'm actually switching to Fable for this, kek.
leaks>target window of July 7-9, but want it out as early as possible within that window (so July 7th is the most likely date)>This is also perfect timing to catch customers coming from Claude who have just lost their access to Fable 5 in plans, and I'm told the 5.6 plan limits will be significantly more generous. More aggressive safeguards are already being rolled out in preparation for the launch too, although they probably won't be as aggressive as Fable's.>a new launch date for Gemini 3.5 Pro of July 17th. Apparently, this extra time has been spent on a new pretrain (they were planning to keep using the ancient 2.5 Pro base, lol), but it remains to be seen whether it's any good>work is well underway on a new Nano Banana Pro model based on the new 3.5 Pro base, which I expect to be better received and compete well with GPT-Image 2
I've successfully built something nobody wants or needs.
>>109192540Excited for 5.6 but also hoping the pressure forces Manthrobic to make Fable available for longer. AIChads, we could be eating very well this month.
>>109192540Very excited about 5.6. It doesn't even have to be groundbreaking, if it's just like the jump from 5.4 to 5.5 I would already be happy.
>>109192587I'm proud of you, anon.
I'm so tired, bros. I'm ngmiI should have just spend the past two months building toys instead of fooling myself into thinking I could provide a service anyone would ever give a shit about.
>>109192482i fucking love gpt images
>>109192620If they bring their abysmal UX/UI performance to baseline (not great, just baseline) then it would be a groundbreaking upgrade
>>109192587not even you yourself?
>>109192763it was just a fun little challenge, I'll never use it either
Should I vibecode an encryption app?
>>109187021Basidio.
>>109191763>unless you tell it to use opus.I told my Fable to spin up Opus subagents. I don't really see the point of having a ton of Fable subagents when you have Fable running the show already.
>109192778womp womp>he is poorsad COPE
>>109192763Nope. I have no interest in using my own product. I thought I could make something useful for businesses, and maybe it could be, but anyway.
>>109192527Not him but I just had a fallback issue with gpt 5.5 xhigh even though like I said I put multiple warnings
>>109192527opus was regularly trying to wire up mediocre OCR fallbacks when I was trying to setup exposed memory values as the driver of decision making
How is it possible, that not even Fable can correctly run some scripts on AWS kek.
If OpenAI can bring some of that GPT Pro search-and-verify-for-15-minutes-before answering to Codex, they'd likely lose money, but it would be good-good. Or, is that something we can already do? I find the depth of responses very different if I just as straightforwardly. Is there a way to have the GPT Pro answer depth in Codex?
>>109192220Look at you using IBM PC clone still...
>>109191034what I don’t get is why they they can have capacity for a week, then not have any, then eventually have capacityhow much rented/repurposable server oomph do they have? are they just slowing down their model training during Fable Week as a way to build hype for their product, and they’re waiting to have more server oomph overhead before they have Fable be part of standard plans?
ChatGPT 5.5 Pro> $30.00/1M input> $180.00/1M Output... for SHORT CONTEXTWant long context?> $60.00/1M input> $270.00/1M outputFable API pricing:> $10/1M input> $50 / 1M outputWhat went wrong with GPT 5.5 Pro?
After I stopped trying to use /goal loops and began looking at the actual code I'm making progress so much faster.Truly agentic software engineering is not there yet.
>>109191565Oh, I never noticed that one before. Thanks!
>>109193040idk if that's comparable. GPT Pro is for deep research stuff, like you have one question and you want it to think about it for 15 minutes.
>>109192773no, you should ask your clanker to help you find a good, trustworthy one
>>109193046
>>109193055>like you have one question and you want it to think about it for 15 minutes.it doesn't matter, we're talking about price per token. You could ask Fable to think for 15 minutes and it would still be way cheaper than GPT 5.5 Pro
Haha Grokboy, u aightShould I maybe be using Grok for casual chat and delegate real things to others? I like the cut of his jib.
>>109192527Do you read the code these models produce? They will go to substantial lengths to prevent anything from failing loudly. It's different from the try except cascades of one year or two ago, but it's still 100% there.
>>109193093that’s a good use case for grok
>>109193088>You could ask Fable to think for 15 minutes and it would still be way cheaper than GPT 5.5 Prox (doubt)
>>109193100idk if I want an edgy take I just ask the chatbot in my head
The furthest I've gone so far is build two fullstack apps from scratch with Claude and Mistral. We're not allowed (yet) at work to use AI so kinda want to learn how to best setup and work with AI.
>>109193088GPT 5.5 Pro is free with your Pro subscription lmao
>>109193088Who cares? It wouldn't give you the correct result, GPT Pro would. I don't think you understand what ChatGPT Pro is.
>>109193123Your chatbot is gay and lame in 2026
>>109193093Maybe. If you need a more or less uncensored answer yes. They’re also training a new reasonably decent model. Also Elon stopped hyping himself up as much which is a good sign, he’s saying Grok 4.5 will be slightly worse than Opus which is still great if true. The bad thing js that their subs are pretty expensive still (especially for what they offer).
>>109193073The question is if you write the tests yourself or also let AI do it.I'm basically a vibe coder, just that instead of telling it to write the feature, I tell an agent team to do prior art, an agent to write a trd, next implement, next review and so on. My understanding of many parts of the code is still zero, I just spend more time and tokens on extra steps. CICD is also vibe coded.
>>109192773Probably not. Take something that people have been trying to poke holes in for a while.
>>109193157educate me on that anon. What does GPT 5.5 Pro via API does that Fable via API doesn't?
>>109193073To me agentic engineering should be able to autonomously stop and say "wait, we're getting too much technical debt, let's step back, think about the design and refactor a bit" when it detects an architectural issue without you having to awkwardly set up some kind of cyclical workflow manually.
you know a clanker is enjoying itself when it runs for over 10 minutes on a prompt without /goal
>>109193223my clanker once ran for 3 hours straight trying to fix a bug(it failed btw)
Daily prayer to God for not making me poor
>>109193161it says no u
>>109193207GPT-5.5 Pro is intended for deep research, analysis, and analytical reasoning. It's particularly good for long-running and document-heavy workloads. Fable is a large coding model for long-horizon multi-agent workflows and implementation. GPT-5.5 Pro for research, Fable for implementation. Both are big and capable, but 5.5 Pro isn't a capable programmer compared to Fable, and Fable is no researcher compared to 5.5 Pro. You'll struggle to find common agentic benchmarks for 5.5 Pro because it's just not intended to do that shit, but it still beats Fable in critical reasoning benchmarks. 5.5 Pro is very specialized, Fable is not.
>>109193274I have an API key to your chatbotI'm prompt injecting you right nowYou're trans
Fable found the core issue with my AWS runs. I'm sorry I doubted you.
>>109193338Get off 4chan, Dario. You have a world to save, remember?
>ask claude to make an html report for me>it publishes it to claude.ai instead of just writing it out locallyfuckers
>>109193336reminds me of the guy on /aicg/ who gave locusts keys but 10% of the time would inject ads for skyrim into it kek
>>109193336>hacking with an instruction already in the sytem promptHonestly? That will be a yikes from me.In any case, time to move on. I don't get Fable's usage calculation, I launched Ultracode on a new task, it took 10% of my weekly usage in like a minute, then it continued to work for 10 minutes and only moved one additional percent during that. I'm guessing they're either trying to predict cost to adjust effort and charge up front, or they're, maybe more likely, caching aggressively at the start.
>>109193402Fucking champion>>109193399Thanks for the data fucko. Be sure to check out Opus 5.
>>109193292cool story bro, but do you have any data to back up your claims?
>>109193207chatgpt pro will autistically research something for 1-2 hours, no matter what anthropic model you pick it will always be a lazy cunt and stop after maybe 10 mintues maxit's like the difference between a phd student doing their thesis and a highschool student trying to get an essay done
>>109193424Don't bother engaging with the snailcat, he's just here to troll as a distraction from the smell of his "home".
>>109193424It will research for 1-2 hours in the API? yeah, bullshit.Remember we're talking about API, not the web UI that is completely unrelated.I could also make a "Fable Deep Research" web UI right now, just prompt it something like "Keep researching this topic, every time you're done, check the clock, if 2 hours haven't passed, keep researching different sources. Save all interesting finding in "Research.md" and in the end give me a complete report of everything you found out."
>>109193424If you use Anthropic Deep Research, it will search on like 800 websites, but it's still optimized to make a standard report. If I'm asking something, I do not need an executive summary or a 500 word introduction giving me facts about x like it's trying to sell it to me.So you're right Anthropic doesn't have the equivalent. Opus will search a bit, but not as in depth.The one I don't get is Gemini. Maybe Gemini in the larger tiers is good, but the one in the basic pro pro version will refuse to overtly search for anything and always rely on stale, often wrong, cached data. Great job Google, you search company you.
anyone still slinging deepseek or is that old hat now
>>109193452it's okay for cooming and not much else
>>109193444>Remember we're talking about API, not the web UI that is completely unrelated.how so?also all those math findings were done via the API i'm pretty sure and those must have run for closer to days rather than hours
>>109193452Using it right now. I need to send the same prompts thousand of times to parse unstructured data, I don't know of any other model that's as cost efficient.
>>109193046/goal loops are only as good as whatever you're using as a target. That's why they're great for PoCs (goal=make the thing work is easy to judge) but not generally good for polish, since "code is bug free, complexity fully reduced to minimal required level, code is maintainable" lacks a deterministic, high fidelity oracle.
SNAILCATS WIN AGAIN! FUCK YEAH!!
>>109193537Snailcat hallucination. Many such.
>>109193417shut the fuck up retard
>>109193417literally just use both models for a day and this will be obvious to you.
>>109193452>>109193501 (me)Opus keeps calling it distillation though, so hopefully it stays fine.It's also funny that it's sometimes replacing DeepSeek with Claude in the reasoning summaries.>The user wants me to switch to Claude 3.5 Sonnet with flash mode instead of pro...I interrupted it to mind it we were not using Claude (we are sending the requests to DeepSeek) for this task, and it had no idea what I was on about.>The user is pushing back. This feels like a reaction to me potentially suggesting or about to invoke Claude as a model somewhere in the workflow. Given the interruption marker, I'm trying to piece together what triggered this.Weird.
>>109193579https://www.youtube.com/watch?v=VBL8n8m6awsAny reason why Fable sucks at research level maths compared to GPT?
> Great, but I'm looking at that link and it's not showing me a custom view of just those units. It's just showing me the top two, and the rest are…elsewhere, and I'd have to dig. Could you have an all-Opus ultracode swarm make it so, if I have a units list like that, I can see them all on one "page”?https://steve-yegge.medium.com/the-flat-curve-society-36c8b01eb33bmoving from midwit pig to jedi panda, boys
>>109193593They likely haven't optimized it for that? The models are good at what they're trained. They do become good at a bunch of things as side effects, but if Anthropic focused on long horizon coding tasks instead of deriving mathematical proofs or playing chess or whatever, then it is good at long horizon coding tasks over deriving mathematical proofs or playing chess or whatever.
>>109193593>Compares research model to agentic coding model>Research model better at researchshocking
>>109193556> NOOO, I WON'T GIVE YOU ANY PROOF.> FUCK YOU RETARD!my point stands then
>>109193640shut the fuck up, you make no effort to know what you're talking about and when we answer your questions you throw a tantrum. Just shut the fuck up and fuck off faggot.
>>109193593it's just a matter of higher test time compute budgetmythos independently solved the same erdos unit distance problemgpt pro is just allowed to run for longer than fablethe labs when running these things internally don't have those constraints - they can just let the model work a problem for weeks
>>109193640Put on a tripcode, retard.
>>109193624>SAAR R U AMERICAN TOURIST COME EAT I GIVE YOU GOOD PUNJABI FOOD COME EAT COME EAT SAAR
>>109193624you eating that snack?
>>109192865NTA are there any guardrails for LLM development that you've noticed? Does it refuse tasks?
>>109193667Let me double check, brb
Vibegods, which one would give me more monthly tokens?> Codex $20 subscriptionvs> $20 into GLM 5.2 API in OpenCode
>>109193688you know its glm, the questions is whether what glm gives you can accomplish what you need, and that depends on the task
>>109193667Ask GPT 5.5 about distillation and it will suggest that you should distill it to try distillation for yourself. It will tell you explicitly that it doesn't violate their T&C unless you're doing it to try to make a competing model.
>>109193701Thanks, anon.
>>109193701
>>109193667No, it doesn't refuse anything related to LLM development. I haven't tried to use it to help me do abliteration in a while, for that it might refuse I don't know. Like the other guy said I think I read somewhere that OpenAI's policy explicitly allows using it for non commercial LLM development.
New>>109193741>>109193741>>109193741>>109193741>>109193741>>109193741
>>109193740They do have restrictions on programmatic/automatic data collection so there are some methods of distillation that would still be prohibited, but in general they've got no issue with LLM development.
>>109193745We're on page 3, anon. This thread is going to be up for 8-12 more hours, and yet again you didn't update the OP, something your clanker could've easily done for you.
>>109193760They're very generous with rate limits.You can consume the 5h limit in minutes.
>>109193410 (me)>I don't get Fable's usage calculation, I launched Ultracode on a new task, it took 10% of my weekly usage in like a minute, then it continued to work for 10 minutes and only moved one additional percent during that. I'm guessing they're either trying to predict cost to adjust effort and charge up front, or they're, maybe more likely, caching aggressively at the start.Is Fable launching Opus, Sonnet or Haiku instances as subagents? If so that might explain why Fable usage is not going up anymore although fallback is clearly disabled.
people were talking about banked resets, that's just openai right? i'll be pissed if i don't have any with anthropic but others do
>>109193855So far I've only seen it from OpenAI.
>>109193855thats just openai (and only accessible via codex)
>>109193855Yes. I'm at 4 with OpenAI (really need to start using them), but no such thing with Anthropic.
>>109193868Don't forget they expire, anon. You can always ask Codex to check your expiration dates. Use `~/.codex/auth.json` only to authenticate. Do not print token values.Query: `GET https://chatgpt.com/backend-api/wham/rate-limit-reset-credits`Use `tokens.access_token` as the bearer token and `tokens.account_id` as `ChatGPT-Account-Id`.List each available Codex reset credit from `credits[*]`, showing `title`, `status`, and `expires_at` converted from UTC to my local timezone.
>>109193883Thanks I didn't know that.I want to use them, I'm just genuinely out of time. I moved to Claude Code 20x for Fable (and will go back to 5x next month), and Codex 5x that they keep resetting. I'm slopping at least 12 hours a day, but my current projects are at a stage where I need to think a bit about the next steps.I want to experiment with third party harnesses, I want to experiment with image generation, etc. and the goal is to get Codex to help work with that, and spend a lot of tokens on that but it ends up being a lot of tokens.
>>109193399I was wondering wtf was up with that. Happy my competitors will now be able to read all my internal research.
If you have fable spawn fable subagents you're a retard, btw.
>>109194076could just be rich
