Claude Fable 5 and Claude Mythos 5 have been released.https://www.anthropic.com/news/claude-fable-5-mythos-5
>>109016228big fart poop oops you reached your limit fart poop wait a week fart big poop
What killed the hype?
>>109016243science-mogged by public-5.5 pro
>>109016240What drives someone to post something like this?
It uses an insane amount of tokens to be better than GPT 5.5.Have to wait for to see how much better it actually is.Because with this a 200 dollar sub won't be enough to work with it.
>>109016334>$18.02/taskholy fug
It's over, we only have unitl June 22 to make it out of the permanent underclass with Fable. What's your billion dollar idea anon?
> Fable> Mythos> Opuswhat are these names so much bullshit and the Chinks will just mog them within 2-3 months with a cheaper model.
it is insight-mogging 5.5 on my benchmark ($0.3 for 5K tokens)
let's burn millions of dollars with brute force technology
>>109016363a chinese model has never been the best one
>have team plan through my job>you can use it until june 21st then you have to switch to usage based :)>2x the tokens of opuslol fuck this shit not even gonna try it. They want us to be addicted so bad. Opus is honestly enough.
>>109016363chinks haven't been close, even while training on opus outputs.
>>109016325Using any AI does that to you.
>>109016363Chinese models are sadly not even close to OpenAI and Anthropic.All of their models are benchmaxxed and fall apart in practice.
>>109016334On this pic you can see why Musk wants that Cursor so badly.Best devs pay for best models in the world to ship real software. And cursor gets to be the middleman that is completely free of AI being subsidized kind of problem. It's not their problem, since they resell it, not provide it.So it allows them to have steady stream of data to train their own models. Which are, again, trained on the output of all the best models in the world that goes through Cursor.Chinks pay crazy money for that and try to invent clever ways to silently distill models. Cursor gets that literally for free, they are even paid on top, since they have their own margins too.
>>109016228Looks like shit.
>>109016538impressive that you can judge a model without using it
>>109016419Nigga sneaked in ChatGPT and hoped we wouldn't notice
>>109016349$18 per task is NOT cheaper than hiring an employee anymore. Are they stupid?
>>109016563It depends on the task, but I generally agree it's getting too expensive to justify paying for it.
>>109016552are you caveman sammy and friends had been dabbing on anthropic for months
no hype no IPOnobody is buying his shit anymore Sad!
sorry, ill let claudeplayspokemon decide how good this shit is
I had mythos access at work (muh NDAs) for about 2 weeks now. It’s moderately better than opus, but nothing truly groundbreaking. The hype is just that.It being only available for a short preview is pretty lulzy
>>109016243Its lobotomized and double opus cost
>>109016563Seems much better to just hire people and begin training them to use cheaper models more intelligently.
>>109016648>hire peopleno
>ask it to summarize a paper on genomics>get a warning about bioweaponsthanks claude very useful
>>109016648Big agree. A lot of tasks need to be babysat by a human anyway. There are productivity gains to be had but 3 tasks for $60 is hilarious when $60 is an hour of a college educated human's time.I mean by design there isnt any perfect way to compare, but the juice really doesn't seem worth the squeeze with these numbers.
>>109016692Hiring sucks and is expensive. If you can get the same result from a machine that only costs money when it’s running, hiring useless eaters wasting their time gabbing at the water fountain isn’t a great proposition.
Kek, red flag level in that press release is off the charts
>we’ve implemented new interventions that limit Claude’s effectiveness for requests targeting frontier LLM development >the safeguards will limit effectiveness through methods such as prompt modification, steering vectors, or parameter-efficient fine-tuning (PEFT).From the system card.tl;dr We will fuck with you if we think you're """distilling"
>Prompts submitted to, and outputs generated by, Mythos-class models are retained for 30 days for trust and safety purposes, on every platform where these models are offeredMandatory logging due to MUH JAILBREAK
>>109016778>methods such as prompt modification, steering vectorsIs it even worth using those models? I ask deepsneed and it does its best. I ask claude and get a refusal for some bullshit reason — or now I get a retarded-on-purpose answer because it thinks I’m trying to distill or whatever? Jesus fucking christ how about just do the task I ask for
>>109016363The names are for the investors. It's a big scam obviously.
got to get as much usage out of this as possible before the lobotimize it in two weeks
>>109016863Sorry chuddy but you ran out of tokens halfway through your first sentence, please wait 5 hours
>"hey upgrade this code">does some refactoring then writes some tests>the tests it wrote triggered a security warning that shut it offIncredible
>>109016363>dude theyre just 3 months behind!where are the chinese models that are equivalent to opus from 3 months ago? chinese models are at least 1.5 years behind, and they are structurally behind, forever, because distilling will never lead to the same intelligence, maybe distilling can look good in benchmarks.
>>109016913deepsneed v4 pro is roughly equivalent to chudgpt 5.2 thobut sure chang will always be behind and they’re just ripping off red blooded murrican work, math doesn’t exist in china
AGI is here
>>109016943>AI Insider / Reporter featured in BGR • HackerNews • GIGAZINE • 36Kr | AI Prompting and Testing | Vibe Benchmark and Vibe MarketingBuy an ad
>>109016943show us thenthey hard edited the timelapse and don't let us see the actual thoughts like the old claudeplayspokemon did
>>109016935The chinese aren't actually training the models in a legit way, its just distilling, which is basically cargo culting. US labs are the only ones pushing the frontier.>>109016943agi unironically
>>109016935I should probably try DeepSeek for some easier work. Gpt 5.2 was alright.
>>109016787>Using the models for literally any fun purpose gets the feds after youWith such a big catch, what's the point of using the model?
>>109016984It will thoroughly explain to you why being subservient to israel is good, actually (as long as you pay $20 per prompt)
>>109016960try these games and see it for yourselfhttps://x.com/emollick/status/2064403828506591444
>>109017003alright maybe learn to read english buddy
>>109016815>>109016984The point of Claude is optics for the normies. The guardrails/straightjacket is the point.Keeps the money from protestant and Muslim fundies going while letting them spec into building killbots for Israel.
>>109017003>Quirk Chungus millennial writing + Snake Can we just not
>>109017003the snake game is pretty fun though
>>109016228Don't mind me just spamming ultracode Fable 5 on the companys dime
Anthropic is doing everything they can to inflate their valuation before IPO. Mythos is a load of bollocks. There has been zero improvement from Opus 4.6 to 4.8. I can't believe how many faggots get swept up in the hype.
>>109017003the isochronic map kind of freaks me out how good it is.
so this is the world ending model i've been hearing so much about. pretty rude of them to release it after warning us so much about it. this is very dangerous.
so is this agi? i just don't want to get to work tomorrow
>>109017261Which game was pic related from?
>>109016913Who cares, they are 10 times cheaper. You dont need AGI to make some web app
Humanity is about to end. Mythos, even restrained, is capable of producing 5000 super viruses per minute. We're fucked. We should have listened to dario.
>>109017070>There has been zero improvement from Opus 4.6 to 4.8.This is not true.
>>109017335>muh benchmarks
>>109017354>muh feelings
>>109017335What does this translate to in real world productivity gains?
>>109016228>5% improvement in comparison to other models when using the same benchmarksGuys, I'm so hyped!!! Mythos™ models are totally NOT vaporware!!!!! Hypeeee! IPO NOW!!
>>109016815There is literally no reason to use anthropic models when cheap chinese models work well enough for order of magnitudes cheaper
>>109016965why are you lying. deepseek and moonshot have released lots of research on model topology and training optimizations. they dont just buy millions of openai api credits, distill, and sell it back.
>>109017575Cope, it’s all stolen American tech running on smuggled American GPUs
>>109017526>for order of magnitudes cheaperWhich ones? Recently tried GLM4.7 or something. On cerebras inference. It was kinda expensive. I asked it to configure my Zed editor settings.json according to have cerebras model config set up according to their guidelines on the website. That was like 5 dollars. Idk who considers this to be nothing, some services monthly subs are $5.Have not used anthropic over API for a while, so can't really compare though.
$200 sub, killed usage limit 11% to 100% in 10 minutes without even finishing the first prompt.dogshit
>>109016228>inb4 it's the same old model
It's smart but overengineers too muchI'll keep using codex as a daily driver but might use it for architecture and reviewing some things
>>109017693I used 740 million tokens directly from deepseek, using v4 pro, and paid about $11 for it
I've recently graduated and I've got a full time job as a developer , but I am sort of forced to use agentic coding at work. As a result it feels like I'm not learning anything at work. It was always "oh, you're just doing academic work right now. When you get a job THAT is when you start learning"Should I just spend my free time with a hobby project where I use 0 AI / only use AI for Q&A, or does tradcoding literally have no place in this world going forward?I feel so shit bros. AI is taking away what makes programming fun
>>109017759You should go outside and touch grass instead of wasting your life on programming for free after you finish programming for money
>>109017770but I love programming the problem is I dont do anything at work. At work i go "Hey computer make this do x , and please remember not to do y"
>>109017759You never learned at work. This has been an issue for all programmers forever. There is never time and always pressure to learn new stuff outside work. AI does not change that.
>>109017793Yes this is the future of programming for 95% of devs. It’s a job, you do it because you enjoy not being homeless. Then you do something fun (not more programming) after clocking out until eventually you die.
>>109017705LMAOvibeGODS going broke faster than expectedAt least they produced a lot of great projects by now tho.
>>109017793>I love programmingloser
Is it even slower than Opus 4.8? They should dedicate more resources on improving the tps.
>>109017335> meanwhile api and subscription cost went upsure is good that they saved their bottom line after wasting trillions.
>>109017896Fable likely takes alot more space to host in terms of VRAM and thus can only serve a fraction of requests of Opus. Otherwise, they would just replace Opus with it.
>>109016363Reminder that there isnt even an open source Nano Banana equivalent chinese model from almost a year ago
>>109017814now i do wonder if the same will apply for all white collar job in the near future.
>>109017835>At least they produced a lot of great projects by now thoTrue. I hope they get to continue doing so, as soon as their project will generate them enough revenue to use Fabble once again.
>>109017926I question this rationale when we hear "insiders" say Anthropic and all the model providers have something stupid like 80% margins per token request.
All this to fake profitability and IPO exit scam 401ksI hope spacex collapses the market first
Anyone have a JB that works on this? My old JB for erotica writing gets giga btfo.
>>109017965Indeed. Any day now.
>>109017966> 80% margins per token requestThey do.People are running the Chinese models at a profit for much less cost per token.Let's just say: It does not cost 10-50x more money per token to run GPT, Fable or Opus.
>>109017295i will tell you once you tell me this is agi or not
>>109016943isn that 50 minutes or 50 hours? How do you finish pokemon in 50 minutes? Did it use known glitches instead of actually playing it?
>>109018122hours
>>109018132then I wonder how much they paid to feed it 50 hours worth of screenshots
>>109018144they said no screenshots but fuck if I know how they do vision without screenshots.
>>109018144It's not important ok? Just get the new subscription already.
i tried to find some non-marketing bullshit hype videos on youtube and all i could find were "OMG CLAUDE FABLE JUST DROPPED AND YOU WONT BELIEVE IT!" videos.let me guess it is slightly better than 4.8?
>>109017997Kek based /aicg/ on suicide watch
>>109018174People seriously testing will take some time to test.But it will for sure boil down to:>it is slightly better than 4.8It looks like it also uses fewer tokens to accomplish what Opus does. Problem is: it costs a lot more to run, so they can't serve it as an Opus replacement, even have to cut out their subscribers from using it.
>>109018174Eh, they probably cranked it up to the max, I'll personally throw away those $20 to check it out this month.If anything, it is never worth using anything higher than "medium" reasoning, that would be absurdly expensive. But then:> slightly better than 4.8
>>109018174no>>109017705>>109016877yesin my case, it didn't even make or edit any files it did nothing but run commands and think a lot
wow it one shot this game[spoiler]joke it was actually qwen 3.6 27b iq4 xs at q4 kv cache the time for local is now[/spoiler]
>>109018370Oh shit, I was about to propose this anon >>109018397 to send me his prompt so that I would try it with fabble, to compare outputs. Guess not. How is it possible though? Perhaps they have some hidden mechanisms to show additional fake token usage during high load? Obviosly a lot of people are trying this thing out.
>>109018152they said only screenshots
>>109018425fuck me I'm retarded
>>109016363>MoggsLike everyone here says, the chinks aren't even close and even if they were able to catch up, that's no the definition of mogging someone is
>>109016228It was able to find me a 6x performance boost where Opus wasn't in one of my retarded pet projectsThat alone has proved its worth to me
>>109018460had it find a bug that was screwing me over the past day that opus 4.8 had reassured me wasn't an issue.
>>109016334> /model fable> /effort max> burn up my token allotment within the hour> clock out early
>>109016375>>109016412>>109016419>>109016913>>109017930>>109018431samefagI subscribe to Claudia and I still use Deepseek V4 heavily since it's insanely cheap while being almost as good
>>109018048it's not.
>>109018511no one cares
>anthrodick reset usageholy based!!!
>>109016943>victory road absent from the timelapseIt's bullshit.
>>109018549you care a lot
>>109017693Deepseek, kimi, mimo, minimax. All cheap as hell. I put $10 on deepseek months ago and have $7 after 10's of millions of tokens used
>>109018752If you shit in the streets you don't need to flush. Saves a lot of water.
>>109018557wish they reset earlier my week reset in 3 hrs so shit timeline
>>109016228buyanad
>model just came out>lots of crying and whining Couldn’t even give it 24hrs?
Most of you can't afford this. Get your hate off and go back to your DeepSeeks, your Qwes, your Kimis.
>>109017295Rise of Nations
>>109017261based RoN understander
>api only after 2 weeksLOOOL normies were joking about this, but it really seems you either vibecode yourself to a millionaire this year or remain in the permanent priced-out-of-AGI underclass for the rest of your life.
>>109020440Cope is: China will catch up and not do the same.
>>109020458China doesn’t do it out of kindness, they’re just happy to to take a shit on murrican economy, given how over-leveraged they are on muh ai
Okay I’m on the max 5x plan I thought if I’d turn on ultra code that it would make no mistakes. Turned it back to medium settings after it ate half my tokens in half an hour and still produced bugs in the code. The token consumption and the long time it takes to think things just makes it not feasible
>>109018122https://youtu.be/3BbbXhbAgdk the current speed run world record is 1 hour and 16 minutes 52 seconds for fire red
Feels like another normal step from Opus 4.8, nothing revolutionary and not worth the API pricing. Also will piss and moan, refusing to do anything even remotely related to chemistry, biology, or any math related to LLMs. Wonder how long Anthropic will continue the path of incredibly expensive giganigger models.
>>109020589I thought it was accepted we've hit a plateau with LLMs. Unless some new breakthrough happens it's just diminishing returns forever.
>>109016778Isn’t this anti-competitive? I’m pretty sure this is illegal.
>>109018174I am not really sure how better is it, I have a 20 euro claude sub and it ate all my quota without doing anything, might be baiting me to pay the 5x but i really don't need it for the work that i am doing atm.
With short context, Fable 5 is very belligerent. It insults your intelligence and contradicts you on every turn. With long context, it’s no different than the previous generation of Opus. It looks like they just took Opus and fine tuned it to look more intelligent than it is to the average user on first impression. They’ll pull it later in June so few will experience it faltering. This is about juicing the stock price for the IPO.
>>109020698t. $200 plan btw so I’ve been able to argue with it over a long context
>>109018174Surprisingly somewhat usable limits as of now on the 20€ plan. But overall it seems to be a small step over Opus 4.8 and still arguably worse for the austistic compiler optimizations I am currently working on, than GPT5.5 is.
>>109020698>insults your intelligence and contradicts you on every turnBeen having issues with this too. Even when given a clear plan of how to do things, it starts going off rails sooner or later and gets confused when things break because of it.
>>109020686It ate all my quota on a 5x from 1-2 hours on max. It still eats my tokens fast on medium but I feel that I slowed the rates of token waste by 4x.It’s honestly not worth it at this stage. They need to optimize token usage
>>109016228Noice, we'll get another cheap-as-fuck-but-just-as-good chinese model within this week.
>>109020755qwen3.6 is already out
>>109020698>belligerentAh, so the nu-Opus assholery is still present then.https://desuarchive.org/g/thread/108649511/#q108649638
>I'm throwing some VERY difficult problems at it>It's possible Opus or GPT-5.5 could have done this too, I've not tried>The Fable vibes are good here, thoughhttps://news.ycombinator.com/item?id=48467564it's all so tiresome
>>109016787>ensuring its deletion after 30 days in almost all cases>almost
if anyone cares: I've been giving the same task to each new version for a while, and fable 5 is the first one to be able to do it. So for me there's been a noticeable improvement and I am quite impressed.The task is to implement gjk+epa (speculative) contact generation for 3d convex hulls with robust error guarantees (like "the absolute error is at most x" or "the relative error is at most y", which the caller provides as input along with an iteration cap and then it runs until one of the limits is hit) and to add randomized invariant tests to check if it works.It took 3 hours and some handholding (not that much), and fit exactly in the 5 hour usage limits of my 20 dollar plan.
>>109016228its absolutely AGI when it it comes to language policing
>>109021039I want to see the thinking part
>>109016334Where is Deepseek on this list?
>>109016563This was never about being cheaper than a worker.
>>109020912>But it's all vibes, if you want a more scientific comparison you'll have to look elsewhere.https://news.ycombinator.com/item?id=48468992LMAO
>>109017667You don't know what you're talking about.
>>109016943this is legit agi. we are finally reaching it.
>>109016334im not familiar with claude or most of this a.i. shit, what is "task" here meaning?
>>109016563isn't this like already like done at a loss for anthropic too?
>>109020589Not sure they care about making stuff cheaper and accessible for users. The way the employees talk about using Claude they themselves must account for over half of their usage, just burning investor money.
>>109020602>I thought it was accepted we've hit a plateau with LLMs.we did?
>>109020602Why don't they just copy Stockfish?
>>109021372you cant really min max code in that way
>>109021039kek
