Date: September 8th, 2026 4:51 PM Author: Spruce galvanic jewess
It designed a website for me in one prompt and on the next prompt as soon as I hit enter it ran out of usage. Have to wait three hours to reset. Used 16% of my weekly usage.
Date: September 9th, 2026 12:49 AM Author: dashing temple
In the chat Astra is a single agent doing every single task in order when Astra is really only needed at the beginning and end, and NOT doing a bunch of front end design grunt work that terra/luna or sonnet 4.6 is perfectly capable of
In codex the main agent will be smart enough to use dynamic workflows, orchestrate lower models and review their shit and make them fix it until Astra approves. Plus all the built in tools that are literally for coding shit like this firing automatically vs the chat having to invent novel coding solutions for every new step
And chat is just a glorified harness/wrapper for codex at the end of the day. If you ever look at its thinking for deliverable tasks 90% of it is coding shit on the fly. It's eating up usage for an unneeded layer on every prompt and MCP and tool firing
Date: September 9th, 2026 4:32 AM Author: maroon step-uncle's house
don't forget, a good amount of chatgpt reasoning is wasted trying to crawl web sites that block automated access, which limits its RAG capabilities. by contrast, codex cli can write a python script that scrapes the web site locally, from your computer and network.
chatgpt has to create the environment where the solution can be designed and validated, whereas codex already has access to it. instead of wasting usage on the testing environment, codex can spend more time actually implementing the solution using information otherwise inaccessible to chatgpt.
Date: September 8th, 2026 4:55 PM Author: khaki disrespectful forum electric furnace
Just pay for the premium, brah--it resets every week instead of every 5 hours. Although you can still blow through that limit fast. I am fuarked bc I have to use it to prep for a senate hearing tomorrow and it does not reset until tomorrow COB
###
This has been a message from United States Secretary of Health and Human Services Robert F. Kennedy, Jr.
Date: September 8th, 2026 4:57 PM Author: Irate frozen french chef
astra medium could have made that and likely better. ultra is turbo autism mode when you need it recursively rethinking the same thing 1000 different ways like optimizing an algorithm. it writes far shittier for example
Date: September 8th, 2026 5:16 PM Author: Irate frozen french chef
lot's of argument on X that using astra for any typescript/python shit is a mistake. it is too good at manipulating high level interpreted languages for its in-app use cases with hardcoded shortcuts etc I'm no CS masterman but TSINAH or someone could explain perhaps.
your astra max website might be an incomprehensible TypeScript "Any" gordian knot using a single class and 1 function that handles everything. so if you want to change the color of a single asset you will need more astra max credits or risk breaking the entire thing
Date: September 8th, 2026 5:21 PM Author: provocative circlehead
gpt is probably the best with python out of any other language. but I think it depends on your use case for whether to use astra or not. if you are doing experiment design I think you should use astra.
Date: September 8th, 2026 5:34 PM Author: provocative circlehead
yeah, probably still not even necessary though imo, you can set up tensor flow and pytorch training runs with pretty basic scripts. I was using o3 for that back in the day.
Date: September 8th, 2026 5:23 PM Author: provocative circlehead
it ran out of tokens lol. cheapest way is to put it on a lower reasoning level. it is fine on high. seems better than sol ultra on high. I was looking at the stats actually for sol, and the higher reasoning levels only give marginal improvements, it makes much less of a difference than on earlier models. and it gets much more expensive for little gain.
Date: September 8th, 2026 5:27 PM Author: provocative circlehead
High doesn’t mean it keeps running the prompt against itself. It’s basically a larger reasoning budget for the same agent. I think you’re thinking of Ultra, which can actually delegate to additional agents and workstreams.
Date: September 8th, 2026 5:25 PM Author: Spruce galvanic jewess
It actually completed the website build. I uploaded the zip to cloudfare, checked it out, looked good but the opening page was missing. So I told Astra, "You removed the opening page with the quotes, put them back," and it was all, "Oh, I'm sorry let me fix that for you," or whatever it said and then about fifteen seconds later I got the message that I was out of work credits.
I could have used the chat function to build it, that's what I used before but I wanted to see how high-end it could go so I told it to build me an incredibly high-end website, spare nothing. I like seeing what they are capable of. Everytime a new model comes out its like Christmas.
Date: September 8th, 2026 5:27 PM Author: provocative circlehead
Don't upload it as a zip to cloudflare use github with continuous deployment. It makes it much easier to make changes. Every time you commit it to github it is connected to cloudflare and pushes the changes.
Date: September 8th, 2026 5:35 PM Author: provocative circlehead
no. it charges you for compute when you commission a vm, but only if you leave it running. you can just press the stop button. just having the aws account doesn't charge you anything. I use some of the ai models on aws too through bedrock mantel, and just give codex my aws credentials. you can put api keys in a .env hidden folder on your computer and give aws full access. useful if you are building something or running an experiment that requires some form of gen ai, language or vision.
Date: September 8th, 2026 6:22 PM Author: provocative circlehead
you don't need aws, that's for a static site on cloudflare pages. that would be for if you were using your own vm because you had backend logic and a database. but you should hook it up to git for sure instead of pasting the zip. github is basically a remote repository for storing code and tracking all changes to it. codex is actually already designed around it and you don't even have to paste any token credentials open ai exposes a connector for it where you just authorize it through clicking a setting in the app and entering your password. then you can just push all your changes to your code every time you make them to github, there is actually a button for it on codex or you can auth codex to do it automatically. you can link this up with cloudflare so it automatically pushes it to cloudflare pages when codex makes a change.
Date: September 8th, 2026 8:17 PM Author: Spruce galvanic jewess
All I do is open Cloudfare, click on update website or whatever the button is, drag the zip over, click deploy and I'm done. Its hard for me to imagine that process being easier. What am I missing? Sorry for being retarded.
Date: September 8th, 2026 8:22 PM Author: provocative circlehead
Difference is that that requires downloading a zip and uploading it and you also are not using git, so you don't have a history of changes to the code, and you can't just push or pull to github whenever, you have to make sure you have to have a copy of the code. It isn't a good practice to not use git and it isn't that hard. Plus once you set it up, every time you make a change all you have to do is push a button and push it to git hub and it will go to both github and cloudflare if you have them integrated. You can also do rollbacks that way. There is nothing inherently wrong with doing it the way you are doing it, but it is considered amateurish, plus its cumbersome if you are making a lot of changes.
Date: September 8th, 2026 5:37 PM Author: Effete ivory national
anthropic and openai both make ~$50 billion in annualized revenue rn which is about what delta and lockheed martin do. i know it's fun to be a bitchmade contrarian on xo but maybe ask why that might be the case & whether ur missing something.
Date: September 9th, 2026 1:08 PM Author: Spruce galvanic jewess
I had Astra run a big task this morning so I got up and ran it at 6:50 am to start my five hour window. I had 55% left after the task finished, got to work and my 5-hour and weekly windows reset.
Date: September 9th, 2026 12:39 AM Author: dashing temple
For the past week I had fable 5.1 in code/chat AND codex 5.6 sol ultra independently generate and then cross-review and perform synthesis and reconciliation until they separately approved a reversioning of my PI case valuation algorithm.
Then I had Astra pro review it all with one chat prompt with copy and paste of every prior chat and the final files attached to the same prompt. It spent 3 hours reviewing and found 13 significant corrections and enhancements. Both SOL and fable insta-approved all of them. And that was a fucking CHAT session motherfuckers
Date: September 9th, 2026 12:53 AM Author: provocative circlehead
lol its not anywhere near $30k worth. most you could use in a month as a single person would be like $6k worth. Also you get 30 percent off if you hook your credit card up to buy $1000 in credits every time you run out. I doubt having extra 200 dollar accounts is cheaper than one $200 account hooked up like that.
Date: September 15th, 2026 3:48 AM Author: Spruce galvanic jewess
You really don't need the work tab unless you are using chatgpt as your personal assistant. In that case, Astra is 180 if you have unlimited usage. I asked my bosses to upgrade me to premium and offered to pay the $80/month upgrade out of my bonuses. Still waiting . . .
But in reality, the chat tab is fine. You can run Sol on high all day long, you won't run out.