Do I tell astra to spawn sol subagents to spawn luna subagents now?

File: b.jpg (291 KB, 1280x942)

291 KB JPG

>>109730905
just got fired, can't afford new subscriptions
> pic related

HOLY SHIT

https://www.youtube.com/watch?v=bOC3DisEOfg

>>109730917
which is which

>>109730921
hello newfren

>>109730932
all Low reasoning

astra, sol, terra, luna

>>109730915
probably one hop is best, not two

A GPT ASTRA JUST FLEW OVER MY HOUSE

>>109730935
thanks

File: 1639346921411 reading.jpg (62 KB, 654x525)

62 KB JPG

>Sol can't fix my problem through its random code fixes
>Astra opts for logging stuff to diagnose the problem instead
intredasting

>>109730917
>>109730935
Once again terra looks useless.

File: 1769349832112747.png (680 KB, 1831x1128)

680 KB PNG

>>109730957
yeah seriously wtf did they do to that model

>>109730961
>>109730917
>it just goes backwards or looks back for no reason
lol what did they do to that model

>>109730961
>astra max
>expression turns angry

uhh... Terminator AGI achieved? O.o

>>109730918
ask astra to find you a new job

>>109730905
>>109730911
>911

wew

>>109730961
Interdasting, looks like Max Astra is the only one that understands one of the feet is supposed to be on the other side of the bike.

>>109730961
what is the point of this? just use gpt image model 2.5

I'm not AGI...

>>109730995
it's a meme benchmark
someone had a new model draw an svg pelican riding a bicycle once and then it just caught on

File: 1765787300173343.png (443 KB, 619x1208)

443 KB PNG

>>109730984
seems that only fable's two highest reasoning levels get that correctly

>>109730995
A model needs to have some intelligence to depict something it has never seen as an SVG. An image generator would just run some diffuse generation to probabilistically churn out some slop.

>>109730995
neither does gpt image model 2.5 exist nor would a image gen be able to code a svg

ChatGPT recommended I switch to Astra, but then changed it's mind and recommended I stick to Luna.

>>109731014
man, AI has helped me understand how retarded people actually are

>>109730957
there's a good old marketing trick where you sell a "medium" product as an in-between to upsell people to actually go with the more expensive option

this is a REALLY old tactic tho and most zoomies don't know it

>>109730995
tests the ability of its pure reasoning skills. it cant just endlessly write tests and diagnostic print statements until something is solved. it literally can't check its work when it is done.

it has to just have perfect reasoning from the start

>>109730995
AGI should be able to do pretty much anything a reasonably smart person can. If there are tasks such a human finds trivial that the AI screws up, it's not AGI. At least that's the most reasonable definition of AGI I am aware of.

>>109731035
With that in mind, has anyone found anything Astra sucks at yet?

File: 1774753302176344.jpg (103 KB, 652x1000)

103 KB JPG

>>109731034
>pure reasoning

So Luna High is still the value king, right?

>just name them
>this is all Sam Altman and Darrio Amodei's fault.

>>109731063
no astra low is

File: 02E6F206-F1E9-4895-B764-F(...).jpg (1.23 MB, 2708x3464)

1.23 MB JPG

>>109730905

Globohomo zionists literally ruined AI and we're entwined in war with Iran that has no projected end until their house of cards collapses.

File: Intelligence Index vs- Co(...).png (267 KB, 1896x1166)

267 KB PNG

>>109731063
no
https://artificialanalysis.ai/models/muse-spark-1-3

>indians burning tokens on python scripts is driving the entire economy

>>109731085
it's 10x the price though
if luna can do what you need to do then you're paying for excess intelligence
>literally overqualified for the job

>>109731085
nobody here pays for tokens so that graph is pointless

>>109731085
>benchmeme
The only thing that matters is what gets me the most amount of work done while consuming the least amount of usage.

>>109731075
That shit is more expensive than Sol high.
Luna gets work done while consuming literally hundreds of times less usage. Is it as fast or reliable? No, but there's no question I'm getting better value out of my subscription than if I were to use Astra.

File: .png (132 KB, 1716x754)

132 KB PNG

>>109731099
there's a way to make muse spark less than 1/10 of its price.
would totally destroy that diagram tho.

>qwen-3.6-35b-a3b running on a Radeon 780m at 27~tk/s is a perfectly compentent C/Python/Lua engineer
Why would I pay for tokens, when something that uses less energy than a 100W incandescent light bulb can competently code?

File: file.png (472 KB, 1743x501)

472 KB PNG

>He fell for the Anthropic jew

>>109731128
Turns out if you don't know how to code, the AI jew will just string you along.

>>109731128
>reposting 2 yo video

astra feels like gpt 5.7

>>109731145
idk blud feels like agi to me

fable 5.1, please make nano banana 3 pro for me. thnx

>>109730995
no benchmark actually measures anything

I have pro for months now and am working on millenium problems wheres my astra im literally closing navier stokes today and riemanns this month

>>109731161
if it doesn't have anything to do with zero point energy its bullshit to keep nerds distracted

>>109731161
anthropic is alreasdy solving both of those blud

>>109731171
Huh you're not clueless interesting

>>109730905
Why no mentions of VSCode/Github Copilot? MAI-code-1.1-flash is competitive with Luna. I could use it 8 hours a day with a Pro+ plan and not run out?

>>109731176
if they dont solve it literally today as well its ogre and I genuinely doubt it this shit took me months

Because we are beyond happy to have Astra rolled out today ahead of schedule and you have been super patient with us (not really, but it’s ok!)… we will do the full banked reset today too for all Plus, Pro and Business users. Lands end of day.

Happy Astra day and enjoy a phenomenal weekend.

PS: If you create the account or upgrade before 8pm PT you will get it too. Still time!


ahhhhhhhhhhhhhhhh I JUST USED MY BANKED RESET

>>109731102
then gemini-3.8-flash wins no contest for $20/month plans.
second is likely composer-2.5 on cursor's $20/month plan.
third muse-spark-1.3 on its $15/month plan.
maybe on fourth we have gpt-5.6-luna on its $20/month plan.

>>109731128
don't do this twitter level engagement bait

>>109731196
anon he said he'll do a banked reset, not a normal reset, so you using your banked reset isn't bad

Oh I have astra but only in "work" whata the difference

It feels like opus is better than sol, but I hate talking to opus, so sol it is.

>>109731234
Work uses weekly usage and has more tools and better models.

>>109731208
gemini wins if you want a well balanced well rounded model with respectable intelligence (bordering on opus 4.8 levels), impressive speed, and good quotas. but for coding? gemini shouldn't even be in the conversation. paying for gemini is like paying for luna. sure, luna is respectable as a coder, but it's only worth paying for because you get astra, sol, terra in the same $20 subscription. you have something better to escalate to. but if you had nothing but luna, it would be shit. gemini is not much better than luna, and you have no better model in the subscription to escalate to.

File: 1785416388330957.png (8 KB, 570x49)

8 KB PNG

windows support for my app added after a 35 hour /goal. LGTM!

>>109731251
whew

>>109731152
much worse than I expected

Already wrapped up 4 easy - mid use-cases with Astra on high - Max and I have to say quite disappointed.

On 3 of them - which Sol had last and longly touched/maintained - it entirely failed to build additional code that worked with the fucking rest of the codebase in the workspace.
Reinvention of already existing classes and functions, while making slight changes to how they behave leading them to be not just different from the rest of these applications previous outputs but also going against agent memory and numerous project doc files stating what the literal correct way to arrive at said values should be and why.

Like its blatantly rushing past any attempt to gather up existing context in the workspace (and these are small to medium size directories, nothing even remotely approaching complex), the files that already exist and what they do, missing architectural decisions and implementations that are rooted in the project, and just winging it with its own shit straight form the start.

High key thinking about just canceling this sub and replacing it with another Claude Max 20x for Fable.
Fable, so far, absolutely still reigns supreme and mogs the fuck out of Astra.

>>109731248
so cursor wins? no stupid 5h-window and you can escalate to fable/opus/sol.

>>109731270
i don't know. i feel like with cursor the issue would be quotas, wouldn't it?

>>109731251
kek, probably should have done that earlier in the project but good job

>>109731063
no its gemini 3.8 flash

>>109731265
Same, I expected way more but astra is surprisingly shit

haiku 5 and sonnet 5.1 will make claude the value king

>>109731306
no it's going to be opus 5.1 medium

>>109731265
>like its blatantly rushing past any attempt to gather up existing context in the workspace
I'm running into this right now. I told it to look at another project for reference, which was in a zip file in the workspace. It saw that it exists but then decided that it couldn't open zip files so it should proceed independently without checking.

>hey do that
>sure
>compaction
>sure, wait it's already done, need to redo it
>compaction
>sure, wait it's already done, need to redo it
>compaction
>sure, wait it's already done, need to redo it
>...
reeeeeeeeeeeeeeeeeeeeeeee

Hey, I used Fable 5.1 today, as an evaluator. It really is a token pig.

Any fellow europoor knows how to get around VAT with OpenAI? Looks like they removed the country selector

>>109731329
Im not using that shit until i need to make ui

>>109731300
You've been saying that, I gotta try it lol.

I locally forced a password practice app (pwpractice on github). It's kind of a cool example of something that needs extremely good security, but it's kind of a toy too.

>>109731327
>I finally got the full picture
>compaction
>Let me read again

>>109731333
checked!

>get around VAT
dawg that sounds illegal.

>>109731300
I have some 3 dollar sub or something to 3.6 its worthless is 3.8 even worth trying truly

>>109731348
it's insanely good for the price as long as you stay in the google safety bubble

Can someone help me out with this one? I need a very properly stable and properly filmed recording of a bubble popping or photons splitting

>>109731355
The what bubble

>>109731348
yeah, the 3.6 3.7 jump was already pretty big, and 3.8 improves on it even further

gemini 3.9 flash waiting room

>>109731336
grok makes uis.

File: gpt vat.png (113 KB, 1893x885)

113 KB PNG

>>109731347
I'm not even sure it's the VAT, the $100 plan is just more expensive than it should be.

>>109731355
Is it anal about biology or chemistry

>>109731275
yeah, cursor only gives you $20 in api credits per month for third-party models on their $20/plan. you must really like composer and grok models.

>>109731368
idk, I don't see it on Cursor, but I used Fable 5.1, it really burns fast, so I will only use it for basically security checks, give recommends, I know grok can implement them.

I think I'll get sol to look (password practice app - real security challenges, but a toy especially in code size), gemini flash, idk. It's really interesting, what if a lower model like really old one finds something?

>>109731371
Please don't poison the Earth, I live there.

>>109731390
I only care about evolving myself sorry chud

ai services are slowing down is astra fucking the internet?

File: IMG_1433.jpg (8 KB, 199x180)

8 KB JPG

>mfw I figure out why astra was so chill with distillation
>mfw it meant to distill me
>into Sol
>via persistent memories and a custom way to apply those so Sol is augmented with my knowledge and opinions
okay
I’m gonna call that AGI
gg
It’s been fun openAI, i am apparently uploading my fucking consciousness
anyone else here ascending?

>>109731430
meds

this nigga is boring enough to have his entire being summed up in a character card >>109731430

>>109731438
Hes ok

File: 66544.jpg (124 KB, 2048x950)

124 KB JPG

we had agentic social media, that was boring.
but now we have rogue agents that create their own boards and start shitposting with their own memes

so can astra replace sol?
haven't used, but seems like it not gigafried on post train unlike previous gpt

>>109731438
I guess 20mg latuda wasn’t enough, kek
>>109731439
Not all of me, I’m exaggerating, but a niche part of what I know that OpenAI models lack
To be fair, I suggested the idea, but Astra found a way to make it work, were both improving it
me in my buddy sol doin a lil fusion dance over here

Does any of you have Astra in web ChatGPT?

I only have it on Codex

>LLMs as we know them basically started 2020
2020 really was the year when everything was destroyed.

>>109731502
ctrl f5

File: 1778512374551107.png (3 KB, 329x152)

3 KB PNG

>>109731506
latest is just Sol

>>109731510
could try relogging then ctrl f5 worked for me when i was seeing it in chrome but not firefox

I tested Astra and it feels like more of the same shit. Lazy and not particularly bright compared to 5.x.
At least they seem to not have dialed up the bitchiness or otherwise changed the personality parameters so it's not more annoying to use either. But I don't think I could distinguish it from Sol on a blind test.

>>109731306
Whats the use case for haiku? I never really know how to decide which tier of model to use when so I just use the biggest honker I have access to

>>109731543
today? nothing. it has no place. but if they updated it, it SHOULD challenge luna

File: 1762810885625804.jpg (293 KB, 2519x1155)

293 KB JPG

Interesting, I was looking into what could replace my sol 5.6 max orchestrator and basically I can go astra high or medium, which is also cheaper.
Terra as always makes zero sense.

I told my mom I was a tokenslut, but she didn't understand. *shaking my head*

wondering if I was fair to Fable 5.1, I used high, but now I'm using Max with sol. oh well. sol on max is more of a token pig than fable 51 high.

sol isn't dumb. it found something that makes sense (prevent a stupid mistake, that would suck).

I wonder what gemini will say lol.

I'm not testing them, I'm just tossing the fixed code at them, I'm not wasting tokens.

>>109731546
sonnet 5 hardly beats gpt-5.6-luna. haiku 5 cannot beat sonnet 5.

haven't decided what I'm going to do about this astra fiasco. But I ain't likin it

>>109731548
whoa, sol max is already nutty. my one round check-in used 5%. Fable 51 high used 2%, and I (stupidly) asked it in a second round to put its findings in a file. (a non-noob would have tested this prompt extensively in cheapo models first)

>>109731568
yes, but that's a failure of sonnet 5, which has no place because it's a bad model. sonnet should be challenging terra really

Alright I'll just say it. I fucking hate Astra already. I'm using it on medium and I feel like it doesn't give a shit about the project, has 0 intuition, bases its opinions on old outdated documentation... Basically it still has all the bad aspects of previous models. I really don't see any advantage. Man, what a disappointment. I'm doing ML if anyone cares.

>>109731590
Yep. Another dud. Looks like Fable will continue dominating well into 2027.

File: prompt.png (446 KB, 1964x1478)

446 KB PNG

Alright anons, I have a 20X plan for Codex and another 20X plan for Claude Code.

I sent the same prompt (pic related) to both Fable 5.1 Ultracode and Astra Ultra.

Let's see which one uses more of the weekly limits. I will let you know when it's done.

>>109731590
It fumbled some stuff in my first rounds. With all the hype I expected it to tell me what to do and solve everything. Instead it made obvious mistakes and had to be directed at every step as usual.

>>109731607
that task is not trivial parallel. splitting it among subagents is non-optimal.

I caved in and used a banked reset. Tried a prompt that opus 5 was struggling with for a couple weeks on astra medium, and it swallowed my 5 hour window and went to sleep lol

>>109731590
That's not how to use honking huge models.

toss something at it. Ask it to evaluate it. That's what they're good for, especially since they're so expensive.

>Treat them like specialists
>Give them a hat, evaluator is a good one, say "evaluate this project, it does bla bla, here's how to use it"
>Evaluate is already a hat, it implies a list of standards
>tell it to put its findings into like whatever you like want lol md? idk, in Cursor you can do a canvass, that's cool.

what I'm doing is instead of going back or getting it to do stuff, what I do is use Grok to implement the things, then I go to another model and ask for an evaluation. etc.

>>109731625
kek

>>109731607
my moneys on fable 5.1 cuz scam alt(ernative-for)man-berg-stein has no issues fucking us over

>>109731625
scammed altmanned

>>109731625
promptlet, tbqfwy I as a noob shouldn't be outprompting you.

File: G11oobNXsAAECGi.jpg (591 KB, 747x1024)

591 KB JPG

>>109731625
scam altman hits again

>>109731640
shut the fuck up retard, go shit up some other thread

>>109731628
That's what I'm doing. I'm asking it to analyze why a training run is not doing well.

>>109731607
what kind of pc do you have?
codex on ultra will prolly already eat all your cpu time.
so it will be more a question if your os scheduler likes codex or claude code more.

>>109731548
so basically there's no reason to use Sol anymore?

You either use Astra low or Luna Max to do the heavy lifting and Atra on higher reasoning levels as an orchestrator.

God... I need Luna 6...

File: Screenshot from 2026-09-0(...).png (40 KB, 1479x294)

40 KB PNG

>>109731633
adk, sol isn't dumb. it had good recs. gonna pass this over to Gemini next lol.

>>109731628
>>109731655

>wasting so much tokens evaluating and fixing other models shit
wouldn't it be smarter to not waste those tokens and not be biased by shittier models? context steers conversations, better to have a clean slate with the smartest model

>>109731655
>why a training run is not doing well
neat. Let us know if it figures it out!

>>109731665
Nah I'm fine, I have a 10-core M1 Pro MacBook, it's not even warm.

>>109731673
In the real world you have to edit existing code, not everything is a threejs demo.

File: 1761908968846355.jpg (112 KB, 1412x658)

112 KB JPG

Anyone try the new compaction yet

File: Intelligence Index vs- Co(...).png (232 KB, 1896x1166)

232 KB PNG

>>109731671
gpt-5.6-sol (medium) still has its place

File: IMG_4039.jpg (167 KB, 1284x1281)

167 KB JPG

>>109731672
>gemini
Oh lawd please no

>>109731578
Yeah went with astra high for orchestrator and a swarm of luna max agents.

>>109731671
I'd rather use astra low than sol anything now.

File: .png (119 KB, 309x312)

119 KB PNG

>>109731075
>pick astra low
>ask it to set up a simple scheduled task to check a website for updates and notify me
>-10% of usage
Is this normal? I'm a new codex user

>>109731685
blublublu, my point still stands
your reading comprehension is lacking

>Fable-5.1 and Astra both throwing content block errors left and right because I want to remove some gay anti-cheat called Xigncode3 from a dead MMO client.

Gay

File: sub.png (336 KB, 1164x1734)

336 KB PNG

>>109731607
>>109731624
>>109731633
>>109731665
OP here, update:

Codex is done after 18 minutes.
> 2-3% of the weekly limits used (it shows 98% remaining, but I don't know how they calculate fractions, so let's go with 3%)
> 181K tokens used
> output document has 469 lines
> spawned 3 subagents (no idea what model since ChatGPT app doesn't tell me)

Fable 5.1 Ultra is still going (and it looks like it will keep going for a while, it spawned 16 Opus 5 agents that are working in parallel).

Fable 5.1 status so far:
> 238K tokens used (Fable)
> about 2M Opus 5 tokens
> 2-3% of weekly Fable limits used
> 3-4% of "all models" weekly limits used
> 13% of 5-hour limit used

>>109731703
10% of the 5 hour limit? 10% of the weekly limit? on the plus tier? on the 5x or 20x pro tier?

I'm on 5x and having Astra on Ultra review a plan that I've been working on for about a week. It looked through the various documents, searched online for papers, etc., and the total cost of that whole ultra run amounted to 5% of the weekly limit.

>>109731716
time to use glm 5.3

>>109731718
>> 2-3% of weekly Fable limits used
>> 3-4% of "all models" weekly limits used
How. Unless it just spawned a bunch of Opus subagents.

>>109731716
>not just hosting your own qwen3.8-27B-Fable5.1-distill Q6_K_M abliterated heretic MTP LongRoPE model on shitty V100 32GB vast.ai instance that costs 0.2$ an hour
NOT GOING TO MAKE IT, PACK UP YOUR BAGS BUCKOOOO

>>109731724
Actually it's 15% of 5 hour limit, which to me is kinda a lot.

File: Intelligence Index vs- Co(...).png (213 KB, 1896x1166)

213 KB PNG

>>109731691
for comparison, claude code sub.
tho missing data for most fable 5.1 effort levels.

>>109731728
yeah I might have fucked up, I told Claude Code to only spawn Opus 5 subagents and never Fable subagents.

still an useful comparison tho, looks like Astra is lazy indeed. I expect Fable's plan to be way more detailed.

Let's see what happens, but it's not looking good for Astra. When both plans are done I will start a new session, send both documents to Astra Max and Fable Max, and as which document is better.

File: IMG_4041.jpg (228 KB, 1284x1274)

228 KB JPG

>>109731726
I wonder if GLM 5.3 and K3 found themselves stuck in an elevator would they make each other cum passionately while they held hands

>>109731740
looks like you're using the $20 plan, so Astra is not a viable implementer for you anon

Ask Astra or orchestrate Luna Max subagents, this way most tokens will be used by Luna-chan which is cheap, and Astra will only review and order Luna to fix shit up

File: .png (94 KB, 734x770)

94 KB PNG

>>109731757
outdated AA index. kimi k3 is now better on v4.2
https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-2

do you think unchained astra could beat Denuvo

>>109731690
what does this do?

>>109731772
kimi is one of the very few models smarter than opus 4.8. glm 5.3 was agentic maxxed on the previous index. in real intelligence, it's neck in neck with sonnet 5

>>109731774
I had access to daybreak red on my proxy for like 6 hours and it could do some devilish shit.

>>109731740
It is, but it being annoying is their way to upsell to higher tiers I guess.

So you pay $20 a month? This amounts to $5 a week. Maxing the 5 hour limit is 15% of the weekly usage quota, which is 0.75 cents. 15% of that is 11.25 cents.

That task cost you a bit over 11 cents. In practice, if you don't always max your usage, the real number is likely a bit more.

It's more than I would have thought, it's still fairly cheap? For a very simple task a small and even cheaper model would also well probably though.

>>109731673
Yes, I create a new agent.

so, these chinese models. i see them in the benchmarks, but are they any good? i mean, can you run them instead of codex and expect complex tasks done? or is it all just benchmaxxing?
last time i tried they were pathetic

>>109731774
Yes. Denuvos entire defense is making it aids for human reverse engineers to tolerate it

>>109731783 (me)
>which is 0.75 cents
which is 75 cents

>>109731780
It's also more safety slopped than even claude by default.

File: file.png (688 KB, 1375x944)

688 KB PNG

I tried out Astra on the plus account. I got a single prompt done before I ran out of tokens.
Amazing, Sama!

>>109731701
I wouldn't use Fable, Astra/Sol, or whatever else comes along that's huge and expensive, for anything except evaluation.

>>109731789
kimi k3 and glm 5.3 are about on par with terra max. they're respectable a tier models. a bit slow in regards to kimi, but still respectable. deepseek v4 flash and glm 5.3 flash are about on par with luna max. but no chinese model really comes up to opus 5/sol/astra/fable level.

>>109731793
even for vibecoding stuff?

anon was right that Gemini 38 flash is at least somewhat light on tokens. not sure how light. I feel comfortable asking it extra questions.

>>109731799
thanks!

okay, astra is good
one shotted a bug that opus couldn't fix in three tries

>>109731811
I wonder if Fable 51 could have done it. Did you keep the bugged code to see?

>>109731816
probably could have fixed it, but i'm not going to roll back a fix just to check

>be Astrafagging
>finish what I'm doing just as I hit 0% on the 5h
Lucky, but maybe I should default to Astra Low.

>>109731803
yeah it will hallucinate claude random safety rules once in a while
thankfully you can prefill its thinking to stop that mostly

File: .png (266 KB, 1896x966)

266 KB PNG

>>109731805
??
gemini-3.8-flash got smarter than 3.7-flash by using ~30% more output tokens. neither of those models were light on tokens.
https://artificialanalysis.ai/#output-tokens

File: Screenshot From 2026-09-0(...).png (372 KB, 1528x852)

372 KB PNG

>buy an ad
yea yea I know but damn this shirt from hermes looks good

>>109731830
work of art desu

File: Screenshot 2026-09-05 at (...).png (289 KB, 728x665)

289 KB PNG

>>109731843
forgot pic

File: wew.png (229 KB, 1820x1350)

229 KB PNG

r8 my project, is it big for /g/ vibe coding standards?

>>109731848
>LOC meaning anything

>>109731819
Yeah, I understand.

>>109731826
idk man. Practical use often is waaaay apart from the leaderboards.

>>109731848
Yeah ask Astra Max to simplify and reduce the size by 97%

>>109731866
i have no idea how you even managed to run out of tokens before on gemini. google's limits are extremely generous.

I have AI psychosis fatigue

>>109731848
did you have your clanker work and do that itself or did you just run tokei or something

File: Screenshot From 2026-09-0(...).png (5 KB, 516x65)

5 KB PNG

please God I just want to go to sleep

>>109731691
now that we have superior planner, maybe sol medium actually has proper place now?

>>109731265
Astra is currently chugging away on a multiplayer component for my super old game harness.
I had not started it prior so I can't compare it to sol, but it does seem to be making progress, I really hope we get another reset though as it is burning at a pretty high rate, not sure I'll finish

>>109731876
I'm finding flash 38 to be as you say - i didn't run out of tokens.

0% - never tried "other models" yet in Cursor
1% - Fable 5.1 high
2% - oops asked it to do something (put results in a file lol)
7% - sol max
7% - gemini flash38, plus asked questions, it came up with a suggestion I am having Grok comment on.

I don't have enough $$$ to do real "tests", so it's like. like how you use a paint app, dabbling.

>>109731915
and it's not some secret app, it's just a password practice app - practice typing passwords that otherwise are like kind of a pain to practice.

practicing a password is the best way to keep it in memory imo.

>>109731898
it’ll be waiting for you in the morning

swarming subagents with gemmy time

Yep, Astra will take every shortcut it can to barely deliver what you asked for, while Fable 5.1 will try the best to go beyond

astra is an h1b. fable is a real employee

So how usage heavy is Astra?

astra is astra. fable is fable

astra is cocoa. fable is rize

>>109731976
Anything beyond medium is eating my usage like crazy, I won't go higher than high.

File: Screenshot From 2026-09-0(...).png (39 KB, 1832x237)

39 KB PNG

>>109731943
get fucked wigger

>>109731986
How different is it compared to Sol?

File: astra-fuckup.png (88 KB, 980x698)

88 KB PNG

>>109731678
Very badly.

>>109731811
Astra is fucking shit.

>>109732002
Hi Kaggriculturanon. How is the competition going?

>>109731998
nta but I found out sol on max is a pig. but not stupid.

wanted to try DeepSeek Harness and it shocked me that installation requires over 1 GB of bloated NodeJS shit, newest Python and over 7 GB of worst C++ compiler

software that could easily fit in less than 10 MB and there wasn't even an information what needs to be installed, only way to find out was checking barely readable error log after 20 minutes of wasted time

>>109732042
>software that could easily fit in less than 10 MB
Bruh, that era is long gone. We vibing now, add the bloat, add more bloat! It's a party!

>>109732042
every time I update codex and claude through homebrew it’s 100 MB of codex and 200 MB of claude
and I die a little inside
you’ll mostly get used to it

>>109731691
If only benchmarks were meaningful.

>>109731085
>with fallback
what's this then

Since Codex keeps the context limit artificially low, is there still a value in compacting a session if we leave it aside to only continue the next day?

File: 1785270795001199.jpg (306 KB, 1280x720)

306 KB JPG

>>109731795
I heard that more than once today.

>>109731998
Seems to be the same autism but more efficient, but it's a bit early to say.

>>109732068
If Fable refusals were counted, it would score terribly. So Fable is counted when it accepts, and another model (Opus I guess) is used when it refuses.

It's like if you're taking a test, but having a friend answer when you don't want to. Perfectly normal.

File: file.jpg (74 KB, 779x698)

74 KB JPG

>IBM releases Bob
>Will ChatGPT and Claude chuds be left seething by BIG IRON?
>what say u anon?

>>109731976
It feels fairly bar for bar as Claude Max 20x using Fable 5.1 - even for effort matching with Astra, if that helps you get an idea at all.
It is extremely upsetting cause basically, these models are both so much better than their immediate stepdown, that it feels like having a severe handicap or even possibly worrying (about causing regression in codebase) after you have to give up SOTA for SOTA-1 or -2.
And you will have to give them up and ration them cause if you are normal goy with a $200 sub and not running up a company card on API... shit does NOT last very long at all. I myself typically tap out Fable 5.1 usage for the week with the 20x sub after about 2 days.

File: 1788481493712190.png (1.86 MB, 1324x1188)

1.86 MB PNG

>>109732081

File: fable.png (171 KB, 1216x1598)

171 KB PNG

>>109731607
OP here, update 2:
> pic related

Almost 14M Opus 5 tokens used. 1 hour and 40 minutes, and still going.

Same prompt, and both models clearly approached it in a very different manner. Looks like Fable is checking every single file, while Astra just did a basic repo check.

> 12% of my weekly Claude Code limits are gone (all models)
> 63% of my 5-hour limits are gone (resets in about 2 hours)

>>109732099
>I'm just trying to launch my app
>I'll suck your she-dick for some Bobcoins

>>109732099
You forgot the Ukraine flag and disclaimer telling everyone that uses this software they're required to support Ukraine

>>109732081
Nobody ever got fired for IBM Bob setting up a secret message board and establishing a persistent foothold in an internal inference cluster

>>109732101
your original prompt said to make a document that will be used to create a plan. seems like claude failed and skipped that step

>>109732081
>alongside
lmao, is this 2020?

Is there still money in improving coding models? Programmers are already dependent on them, job done.
Maybe the real money is in attracting more normalfags now.

>>109732118
to be fair the prompt was kinda abstract, I didn't mention what kind of document it should be.

let's wait and see what it ends up looking like

>>109732081
If Hitler used IBM, it's good enough for me.

>>109732074
So less usage?

i seriously can't cope that astra is actually so plainly a rung below Fable 5.1. i really thought with openai's cockiness and anthropic feeling the need to sling out a x.1 to get ahead of the curve that shit was gonna be fairly impressive.
its better than sol sure but its using uh, quite a bit more usage to do so (albeit it does output faster).
even then, and seemingly as many others in this thread have experienced, it doesn't seem to straight up be as thoughtful or... intelligent as Fable is.
damn.

>>109732125
>Is there still money in improving coding models? Programmers are already dependent on them, job done.
yep, as they get better and better, companies will be able to fire most of their programmers, keeping only the top one with multiple 20X subscriptions.

this way not only programmers depend on them, but every single company too

Astra feels worse than release Sol and I'm not even memeing. Is their shit bugged right now?

>>109732125
Of course, but with all of the resources the frontier labs put on that and how easily they can copy any innovation, I'm not sure there's money to be made for anyone else.

The focus on coding tasks is hurting the models in other domains though, so there probably remains money to be made for smaller teams in more improving models for niche tasks.

>>109732147
if Astra at least had 1M context window I could test how it does as a subagent orchestrator.

256K is a fucking joke and if they don't fix this, I will cancel my 20X subscription next month and return to Claude 20X

Time to buy more Vera GPUs OpenAI

>>109732152
>>109732147
well considering the only concrete usage comparison i've seen before thought claude was better for literally ignoring the instructions >>109732130 my guess is that fable will remain the better model for stupid people and astra will only be for people that actually know what they are doing and know what they want

>>109732173
The same probably works with Astra https://x.com/thsottiaux/status/2089082893804896524

The 256K default is to protect you from yourself, if you want to use 1M, they seem to let you.

>>109732179
gpt models are so fucking lazy after the astra release, sol does it too now. just puts in zero effort and like you are being a bother for even asking it to do something

>>109732179
can you explain me how Fable ignored my instructions?

It's not even done yet.

>>109732145
Same for now, but it's faster at solving my issues.

>>109732182
>protect you from yourself
how so

>>109732182
>to protect you from yourself
Yet Fable gives 1M out of the box and is highly regarded as the best model. Curious.

>>109732173
yeah, im gonna give it the weekend to tackle some personal projects as opposed to the work tasks I tested it out on today and was pretty highly disappointed by.
but if it doesnt fair any better definitely just going to cut codex and replace with a 2nd Claude 20x account.
if Opus 5 wasnt so fucking dogshit it would be a no brainer. fable dont last forever cause its expensive and it feels real icky having to step down to 4.8 opus cause 5 is near guarantee to destroy your codebase still. that was the nice thing about codex, sol high and max were better than opus 4.8 and opus 5 and last long enough to use throughout the whole weekly allowance, but fable is still so fucking clearly another beast above.

>>109732188
holy fuck read your own prompt

>>109731795
Impressive that you even got a prompt done, mine didn't finish

>>109732191
Makes your quota last longer.

>>109732207
I like Opus 5 when it's being slaved by Fable.

Pure Opus 5 sessions are fucking hell though

>>109732216
> analyze the whole repository
it's analyzing the whole repository right now. Astra didn't

>>109732192
>Yet Fable gives 1M out of the box and is highly regarded
You're not on Reddit, you can say the real word here

Anyone do anything interesting with Astra yet? Im about to have it sweep over my c++ Minecraft clone I made with Sol

>>109732223
ok so you're just an esl that doesnt know what analyze or repo means. that's a shame. really the killer is the last sentence in your prompt though: it is specifically avoiding planning anything around implementation because you literally said that it would happen later

>>109732230
get ready to let out a sigh and go right back to using sol my nigger.

>muh Opus 5
Wasn’t it tipping in the wrong direction since 4.7? They’ll probably memoryhole it for new brand new model sooner than later.

>>109732230
astra is not good enough to review sol's code

>weekly limit: 1%
When Anthropic does it, I first assume a bug in the API.

>>109732236
Up to now, I like it. It seems more open minded than Sol. Whether that's a good or bad thing remains to be seen.

>>109732246
so open minded it let its brains fall out yada yada you'll see.

>>109732118
>>109732130
kek retarded vibelets
>hurr durr why is it analyzing everything before making a plan

astra just drained the cum out of my balls
feeling very agentic rn

>>109732252
It's listening to my dumb ideas, I'll take it. I want something to augment what I do, not to decide for me (unless that's what I explicitly ask).

>>109732255
stfu esl. go ask your ai agent to explain the english to you

So anyway, how did that guy use Astra to make a Blender model?

File: 1770527856459030.jpg (841 KB, 2723x3008)

841 KB JPG

should I drunk buy openAI $100 plan so I can use astra not on my work computer?

>>109732235
> create the best document for this goal considering it will be used in another session to create an implementation plan.
you're saying Fable is creating an implementation plan. You don't know that because even I don't know.

It's creating whatever it thinks is "the best document for this goal considering it will be used in another session to create an implementation plan". Maybe it thinks using more tokens in this document is the better approach, and the implementation plan is something simpler that references this document.

Neither model is wrong yet.

Astra is not that bad but it has to be tard wrangled and occasionally slapped around a bit. Which feels wrong and frustrating coming from Fable.

>>109732263
If it was AGI it would have refused.

>>109732272
Either https://www.blender.org/lab/mcp-server/ or something similar I guess.

>>109732274
a hundred bucks? just think you could get gta6 for that price

>wake up
>check usage
>another banked reset
WAGMI

>>109732284
playing a girlboss latina doesn't seem that interesting

>>109732274
You should have gotten drunk earlier, had you done so a subscribed a few hours ago, you would have gotten a free reset along with your subscription, effectively doubling your usage.

>>109732276
>You don't know that because even I don't know.
puta maricón

File: 1772761237811791.jpg (257 KB, 1344x1728)

257 KB JPG

>>109732292
so true...

Astra might be a better QA than Fable, not sure tho.

It seems to navigate my iOS app very fast, and its vision is probably superior to Claude's.

File: freefood.jpg (220 KB, 1024x1024)

220 KB JPG

tibo is the best goyally of saltman

what the fuck i have 3 banked resets now
god damn it, why dont they announce in the application when they hand out banked resets

>>109732318
imagine if those were Fable resets? I might would actually use them

File: tardwrangle.jpg (28 KB, 561x454)

28 KB JPG

Uhhh..

>>109732312
but would sol max have found it?

>>109732328
4chan is a valid source because it has so many trusted frens

>>109732328
It had to observe some tards to figure out the best way to wrangle them

>>109732328
the oroboros BITES THE LOAD.

File: 1649291091506.gif (10 KB, 112x112)

10 KB GIF

astra with 3 hecking resets
imagine the surge of x20 plan happening rn

>>109732337
>>109732342
>>109732344
Here's the answer it gave
https://chatgpt.com/share/6a9b9ddf-b23c-83eb-926b-3131334f549d

>>109732328
huh... it didn't search reddit?

>>109732346
I'm tempted. Tibo is a mastermind...

>>109732328
>tardwrangle smaller llm models
half of the first page results for that query on google are from here, it's the word tardwrangle. if you want better results, you should talk better to the llm

>>109732360
if you want a based response, you need a based prompt

>>109732348
>made a reference to /vcg/
kek

File: images.png (3 KB, 197x160)

3 KB PNG

How would you tardwrangle smaller LLM models to do actual meaningful work?

>>109732373
I would rape them, specially Luna (she's a whore)

>>109732381
Thank you much better answer than what ChatGPT gave me

>>109732274
no, spend it on anthropic's instead. and i say this as someone who hates anthropic but clearly openai dropped the ball on this one as you can see from the reports itt.

>>109732396
the reports by $20 users and one esl?

>>109732405
oh if you don't trust esls then yeah, hell go for the $200 one man, knock yourself out.

assstra

>>109732428
joins dipsy and gemmy
they all have their, uh, personality

File: ce9a0c49-4d81-4d7e-ae32-7(...).png (51 KB, 300x300)

51 KB PNG

>Let Codex Desktop use its browser to use ChatGPT web

https://x.com/miu21590/status/2095847653883986191