I'm still surprised by how much usage you can get from a Claude max $200 subscription, using Opus 5.5 medium. My weekly reset on Friday morning, i've been running multiple agents on large codebases since then and i just killed it this morning, 0% left. That's 3 full days of heavy work where a codex $200 lasts barely 1.5 days with this kind of usage. I still have 2 codex $200 subs that i plan to move to Claude but in the end i might just need 1 Claude max to replace 2xCodex $200

1/ they tell you the reset gonna be end of day. 2/ you use one of your banked reset and plan to max out everything till end of day 3/ the reset actually comes by mid day. used a banked reset for nothing, thanks for playing.

i pulled the trigger a few weeks ago on a Mac Studio M5 Ultra. My intent was to secure my spot in the line and lock the price in while waiting for the 512Gb version to see if it makes sense. Part of the reason is also to secure some local compute in case Anthropic & OpenAi subscriptions get nerfed post IPO. It makes little economic sense to spend $10k on local inference hardware while we can use generous subscriptions, i'm still pondering if i should cancel the order. Anybody having similar thoughts ?

GPT Astra is better than Sol, but still so dumb.

Astra, 1 hour later. "Have you actually done something ? " "I’ve spent far too long in preparation and review. Sorry."

Opus now reviews an implementation plan drafted by gpt Astra ... Let me guess, over engineering ?

as expected, sol just messed up everything. Opus comes in and reviews what happened, basically gpt over engineered once again

gpt 6.1 sol strikes again. Why is it slow although we have 3 agents working concurrently on the codebase ? well, it appears gpt messed up once again ...

never change gpt 6.1. Sol doing over solengineering once again. I drained my 2xClaude max subs so i have to wait for the reset and i'm using gpt for some stuff and it keeps messing things up, i'll have to clean up the mess as soon as Opus is back online

today's numbers: sol 6.1 'normal': 50 tk/s sol 6.1 'fast': 50 tk/s ... results vary from one day to the other, sometimes sol 'normal' mode is as fast as 'fast' mode although 'fast' is priced 2x higher and sometimes sol 'normal' mode runs at 25 tk/s. In the end it's a joke, either you have slow mode instead of normal mode, or you're paying for 'fast' and just get regular 'normal' speed ... Thanks OpenAi, i wonder if they plan a refund for people paying for 'fast' ?

Opus 5.5 is putting it nicely, Astra just dropped the ball. OpenAi keeps giving resets but i'm having a hard time giving gpt work to do now that i've worked with Opus 5.5 for quite some time, the model just is smarter, and it even is superior to Astra.

audit the whole app. Astra high, fast. go.

the speedup didn't last long, already back below 30 tok/s for gpt 6.1 sol ...

it's been going on for a while now, Codex just suddenly thinks it doesn't have access to docker or a mcp server. It usually happens after a few hours, you then need to restart codex and the problem goes away ...