I bought ChatGPT plus ($20) I used GPT-6 Astra on "MAX" I gave it the task: "Integrate Supabase and Improve App UI" It worked for 40+ minutes! YES! The task wasn't done... These are the usage stats: - 100% 5 hr limit consumed - 16% weekly limit consumed -... 5 Million tokens used - 50k Output Tokens For $20, where I can barely use Opus 5 for 20 minutes... I am getting GPT-6 Astra for 40 minutes is MORE than enough. BEST value you can get! Absolutly AMAZING!show more

Here are another screenshots because some people just don't believe:

What do you think about this?

Codex was good before... it used to finish the tasks but now... it just puts a hard limit... not good... Hopefully they'll not turn themselves into Anthropic ngl

By the way, what happened to Luna Reserve? I thought we were promised it 👀

I was expecting it too but seems like they have some other plans. 🥲🥲🥲 Also Luna just breaks a lot of stuff. It just doesn't work so I don't think the reverse Luna even works for me.

Yes! Luna somehow feels way dumber after Astra came out 😅

It WAS dumber when Sol was there... I stopped using it because it was breaking code

How? Used it for 5 mins on simple tasks Astra light and all my usage would run out. Maybe there was some bug and they patched it?

I don't know... for me it worked 1 hr and 16 mins before in Blender (posted that one too)... and now 43 mins in coding

@eggporl Is it possible to take a simple product mockup from chat gpt to render in blender, good enough to do a 3d print?

@eggporl YES! Completely possible! You should try it

I'm calling bullshit because astra eats my 5 hour window in 4 minutes and returns a null result. I have to tell it to keep a run log and work as it goes and then schedule it to restart in 5 hours. It's smarter than sol though.

Again use 500K context for 6 mins now... still 65%+ left... I can't believe you

Don’t fucking lie, I ran a security scan and it consumed 5hr usage in just 3 minutes and 13% weekly limit. It’s been 3rd reset and the scan hasn’t finished. That too on Astra Light, not max.

Well you're allowed to believe however you want. I have video and I have a screenshot and it works for me. Maybe you are one unlucky man.

So you are praising two models that did not finish the task and value it by who worked longer while still did not finish the task? Interesting If we had AGI it would refuse to work on request ‘improve App UI’ without any details. It is just super smart for using tokens when you give this kind of prompts

I do not know what to say to you. I have only written "improve app UI" so that you can understand. The original prompt was much more detailed.

I was not expecting an answer. Just giving my comment on valuating two models even tho neither one finished the task

40 minutes on Plus? 😂 Maybe in some parallel universe, but definitely not in this one right now.

Well you can watch the video again and decide for yourself.

BS. ChatGPT plus is essentially useless with GPT-6 Astra. One refactor consumed my 5 hour limit in 10 minutes and ate 35% of the weekly limit

You are allowed to believe whatever you want bro.

It's not about belief. I am a ChatGPT plus user. I used it today. And yesterday. I know what I am experiencing with it

Astra doesn’t have a “max” mode. It is Ultra, tardo. I call BS on this post

What if I call your knowledge "BS" ... does that hurt? At least try to research something if you don't know before shit posting in comments

You can call BS, but you’d be wrong lol. I’m not shit posting; I’m correctly a noob who obviously doesn’t know what he’s talking about.

Yeah sure

Stop lying.

Lol there is no way you ran Astra for 40 minutes on the basic plan. No way in hell!

Video is the proof + Screenshots

I have a corporate plan and I could run a single task for not even 20 minutes and burnt through two 5 hour budgets. The corporate plans must be throttled to pieces then because this is insanity!

simply get this plan 😅😅

I just did haha

congrats!

Also, I have this SaaS for people who want to grow extremely fast on X

These are accurate stats. I am getting similar results. But there is no point in using MAX; use xhigh only. It is better in my opinion in terms of cost and gives similar results.

Yeah I just used it for this test. I don't even use Astra.

Well people, when OpenAI cuts the token limit for Plus accounts, thank this big mouth.

What do you mean by that?

Who paid you

No one, but I really hope someone will

max on £20 just nit worth it and I hate combining different models have to spend close to a grand a month just to survive new way to make people poor.

Bro... I only have like ~$30+ of subscriptions that give me plenty... idk why you spending a grand

Try Astra using Blender. It will eat your 100% usage in 10 mins.

🤣🤣🤣🤣🤣🤣🤣🤣🤣🤣🤣🤣 Watch this:

it differs task to task, this is very misleading

I can't fight you bro... Thank you for the comment

Thats so good, people are complaining about usage but if you compare it with anthropics….

Bro, if you compare this usage to Anthropic, you're definitely going to see that Anthropic doesn't give enough usage to their customers on the $20 plan.

$20 Plus. MAX mode. 40 minutes. 5M tokens on a Supabase + UI pass. Capability flex — or a warning label for anyone budgeting overnight agents?

'max' mode burned through 300k tokens per hour, same rate as a gpt-4 call center script

What do you mean by that?

Why you lying?? Opus for 20m ?? I can use opus for as long as i wanted unlike sol which is about 1/5 of the same usage i get with claude 20$ pro subscription

I don't know how you are using Opus for Hours I have used Opus 5 for a month on the Claude Pro $20 subscription and every time I used it, it only lasted 20 to 30 minutes per 5-hour session.

Even on weekly limits it takes almost 1.6 time less , with codex burning your 5hours limit takes 16% while on claude 10% always so even weekly is way more generous

Well I think so. I need to give it a try again

Simple things to consider just don't start new coversations to continue a project and always use /compact when the usage at 99% and u can't message more for less than 1h then u can to compact so that next message will not cache miss(read) again this is where most of usage go

The interesting metric here isn’t just how many tokens the model consumed. For long coding tasks, I’d also track first-pass completion rate, runnable code produced, rework after the context window fills up, and how well the agent recovers from interruption. High token throughput is useful—but reliable task completion is what determines the real value.

dont use astra on max, high is like lossless, just like 1% increase on some bencmarks

I used max just for the limits test. I don't even use Astra for most of my work.

Why max? 🙂

Because it's a test.

I think people who spend more than 20 euros a month on AI for coding should think about doing something else, since they have completely lost the criteria and control of the solution they are working on.

I spend $30 something and I get my work done... People who spend more than $100 on AI are the ones who need to look out for themselves.

I agree

$20 for 40 min of actually useful agentic work is honestly a steal compared to burning Opus credits on the same task taking half a day. that 5M token context window is the real game changer though.

It's not the 5-million context window. It's the total consumption of 5 million tokens.

$20/month for a 5 hour weekly limit is wild when south korea's about to give unlimited ai to 51 million people for free.

Yeah but I'm not in South Korean

This is the kind of dumb example you get when someone tries to cut paper with a hammer and calls it a test. Just to end up frustrAted 40 mins later.

I am not frustrated ... I am happy that it lasted 40 min

About time someone sees the value in it. Expecting things to be cheap or basically free can't always happen. At least openAI lets users use Codex even on the free subscription. That is huge itself. Granted it's terra 5.6 not Astra or Sol but that is more generous than expected

exactly! Claude doesn't even let you touch ClaudeCode without a Pro subscription

Seriously and that sucks coming from a company that cares deeply about ethics lol.

@shownotover haha, sounds like me last week! It's a wild ride when AI goes all-in but doesn't quite finish. Sometimes you gotta step in and nudge it a bit.

But I didn't want to nudge and I was just testing out the limits.

@shownotover totally get that. sometimes you find the edge by pushing a bit. how did it go?

IDK i can use opus on high parallelly on 2 projects for ~2 hours straight

you must have 5x Max subscription... Opus 5 barely works 30 mins on Med with $20 subsription

I have the $20 dollar, running only opus at high in claude code. IDK I use graphify and do well structured implementation parts, and always open new session when one task finished. Ive never reached 200k context

That can be the reason but for now I have left Claude Code and I do not want to go back because Codex works perfectly fine.

Sure, also, I'm interested with the same usage behaviour do u get more or less out from the same 20$ plan on codex.

I get way more on the $20 Codex.

I created an app in Go with cursor on the $20 plan used it to make the app from scratch, connect supabase, handle all the secrets being stored safely, over a year ago. I don't know if this is really that impressive.

The UI improvement was the biggest thing... that was the biggest consumer

Did it make anything creative for the UI? Cursor had a big downfall of not being a UI/UX expert. I'm not either just wanted to see what it could do vs what I could do. I was impressed.

Basically, in the UI, there was a lot more difficulty for the AI to actually talk to the user in real time. And I wanted this feature, and I also wanted to improve the whole chat UI for the app that I'm building that works with AI, with other AI agents. So definitely it was a bit difficult, but that's why I gave it to Astra, and I think they did a pretty good job.

Nah shut the hell up then there is something bugged.. I am on x20 and used all banked.. just used my last and already on 80% again

Everything is in front of your eyes and you still are saying that this is all false. Well I can't help you

I did not say your test is false. I said it maybe not the same in every account / harness etc - Maybe server side issue, maybe different points in time, you tested 1 account.. Since I've saw your test I came to the conclusion that something is off....bruh

No I didn't test it once. I tested it twice and it worked: 1 hour and 16 minutes for the first test and then 43 minutes for the second.

Does not change anything.. Like I said something is bugged somewhere and I don't care what some random tested with 2 plus Accs.. Ive got 7 plus and one x20.. I can see by myself what is happening without any test

in this case its a usage bug affecting some users mine literally runs out of 5h usage in 15 minutes max, and i am using High not xhigh or max.

I seriously do not know why this is happening. Mine worked for 40 minutes on a MAX setting with a plus plan. you can see that literally in the video.

yeah i'm not saying you are lying, it's not the first time @OpenAI released models with usage bugs, but in my experience 15 minutes is all it takes to demolish the 5h limit, and it wasn't even an app-wide prompt. @thsottiaux

@OpenAI @thsottiaux He ain't gonna listen here... XD But yeah if there is a usage limit, I think he is going to give you a banked reset.

@OpenAI @thsottiaux ik chances are slim for him to see this but never 0 lol

@OpenAI @thsottiaux 🥲🥲🥲 You are way too optimistic haha

@OpenAI @thsottiaux i'm not, i don't expect any answer, but typing a @ and a few words ain't hurting my hand lol

@OpenAI @thsottiaux Haha... Sure, I ain't complaining or anything... Good luck bro! Happy day to you