The most prestigious law school admissions discussion board in the world.
|
 |
 
|
Codex Sol uses way too many tokens per pass
Poast new message in this thread
Date: July 10th, 2026 5:05 PM Author: The Penis (http://www.autoadmit.com/thread.php?thread_id=5881563&forum_id=2most#49991573) |
Date: July 10th, 2026 5:07 PM Author: gaetan dugas I stopped going further than 5.4 on standard speed
No clue wtf u guys need past this
This one does it all |
 |
Date: July 10th, 2026 5:16 PM Author: The Penis The small percent jumps in capability can make a big difference in long tasks and big projects over time. 5.4 can technically do anything sure, but it requires much more constant and attentive supervision and fixes. (http://www.autoadmit.com/thread.php?thread_id=5881563&forum_id=2most#49991610) |
|
Date: July 10th, 2026 5:19 PM Author: gaetan dugas Interesting but I'd assume I'd be able to get double the work done since I'm using less tokens
And are you really saving time? I had it do some complex shit today and it took like ten mins tops. Unless ur getting that same level of work in like 30 seconds I'm not seeing the issue. Tye difference between 5,7,10 mins in meaningless especially for the amount of work that can be done |
|
Date: July 10th, 2026 5:22 PM Author: The Penis Yeah I agree with that mostly except the main thing I'm worried about is it making an error or failing to catch a previous error that I would have never noticed, because I'm lazily vibecoding then it bricking the entire project or taking like an extra week to fix or whatever. That's the point I think it becomes worth it. I agree 5.4 codex was still very good, but feels risky to use now that there is one that performs better. (http://www.autoadmit.com/thread.php?thread_id=5881563&forum_id=2most#49991624) |
|
Date: July 10th, 2026 5:28 PM Author: gaetan dugas Imo instead of coding I now use that time to do extreme amounts of verification/testing/edge case testing and shit like that. (http://www.autoadmit.com/thread.php?thread_id=5881563&forum_id=2most#49991638) |
|
Date: July 10th, 2026 5:30 PM Author: The Penis Yeah but I'm a GOOD GOYIM, and do most of that verification inside Codex now too. The agents write unit tests and execute them, run smoke tests, execute the full application, inspect logs, test edge cases, and report the results back. (http://www.autoadmit.com/thread.php?thread_id=5881563&forum_id=2most#49991643) |
|
Date: July 10th, 2026 5:37 PM Author: gaetan dugas Lmfao
Jesus take the wheel |
|
Date: July 10th, 2026 5:23 PM Author: zarathustra i poasted the same about fable a couple of days ago. they are divorcing the cost of the model to the user from actual compute cost to them for the most advanced models. instead jacking up the price on power users and less cost-sensitive users to increase margins ahead of the ipos. (http://www.autoadmit.com/thread.php?thread_id=5881563&forum_id=2most#49991626) |
|
Date: July 10th, 2026 5:26 PM Author: The Penis idk how people even afford to use "fable" in claude code unless they have an org supporting them. it costs me like 5 nigga dollars already over the past few days just doing a few basic chat requests to the fable api on x-high reasoning with streaming and a 60K token cap. But that was just like a handful or so of text-based chat requests, which is nothing compared to using it for agentive workflows on huge codebases. (http://www.autoadmit.com/thread.php?thread_id=5881563&forum_id=2most#49991632) |
|
Date: July 10th, 2026 5:28 PM Author: zarathustra i can break out fable like twice a session to do the heavy lifting and then i have to put it away. even that usually takes me over my limit and forces me to use some credits before the five hour clock resets. (http://www.autoadmit.com/thread.php?thread_id=5881563&forum_id=2most#49991640) |
|
Date: July 10th, 2026 5:27 PM Author: computer online (🧐) anthropic models add six billion lines of dipshit comments. i fucking hate how claudes speak (http://www.autoadmit.com/thread.php?thread_id=5881563&forum_id=2most#49991635) |
|
Date: July 10th, 2026 5:28 PM Author: The Penis what do you think of the new gpt? (http://www.autoadmit.com/thread.php?thread_id=5881563&forum_id=2most#49991639) |
|
Date: July 10th, 2026 6:20 PM Author: computer online (🧐) i have always preferred scam altman's work to the others (http://www.autoadmit.com/thread.php?thread_id=5881563&forum_id=2most#49991813) |
|
Date: July 10th, 2026 6:30 PM Author: The Penis yeah same (http://www.autoadmit.com/thread.php?thread_id=5881563&forum_id=2most#49991839) |
|
Date: July 10th, 2026 5:30 PM Author: zarathustra opus is the most try-hard verbose dipshit on planet earth. fable doesn't do that but it also doesn't care that it's smarter than u and often doesn't explain itself. (http://www.autoadmit.com/thread.php?thread_id=5881563&forum_id=2most#49991644) |
|
Date: July 10th, 2026 5:32 PM Author: The Penis Fable is actually smart as fuck. Been playing with it having it argue back and forth with sol though and sol holds its own. Was having sol argue with sonnet 5 and it was making an absolute fool out of sonnet 5, which I think is garbage. They ruined sonnet this round. (http://www.autoadmit.com/thread.php?thread_id=5881563&forum_id=2most#49991645) |
|
Date: July 10th, 2026 5:41 PM Author: zarathustra i should give gpt another shot, it sounds like it's gotten a lot better (http://www.autoadmit.com/thread.php?thread_id=5881563&forum_id=2most#49991656) |