Hamel Husain suggested offering an adaptive mode to balance both.
when you use codex do you want to choose the model and reasoning effort yourself?
should the product just pick based on the task?
what would make you trust that choice?
Are you someone who always uses extra high, or do you think about using low when you don't need that much thinking
Positive users support adding an auto router to Codex for easier model selection and token savings, while negative users reject it to retain control and avoid past disliked adaptive features.
No Digg Deeper questions have been answered for this story yet.
I wouldn't hate an auto option for reasoning effort. I would not want it for model choice personally, although you can add that to as long as they are customizable.
But I have an exactly seen a router that works as amazing as I would like. The fact of the matter is that I don't need high or extra high reasoning for every single query that I make so there's definitely value in routing to lower reasoning levels for low complexity tasks
That said, even the higher reasoning levels seem to do this to some degree anyway because for simple questions they still respond, more or less, right away
@jxnlco I would 100% rather the product routed it for me. Several coding agents like @ampcode @droid already do this for you and people seem to like it
when you use codex do you want to choose the model and reasoning effort yourself?
should the product just pick based on the task?
what would make you trust that choice?
Are you someone who always uses extra high, or do you think about using low when you don't need that much thinking
@jxnlco The time wasted and hassle from any misclassification into a dumber model almost always outweighs the savings, for me.
The misclassification cost is much worse in one direction. You'd need to convince me that calculation has changed.
when you use codex do you want to choose the model and reasoning effort yourself?
should the product just pick based on the task?
what would make you trust that choice?
Are you someone who always uses extra high, or do you think about using low when you don't need that much thinking
@jxnlco If i could trust the routing logic worked really well it be much nicer not to chose but still have the option if i wanted.
@jxnlco I want to choose, I use low most of the time because my goals are focused, narrow, and usually straight forward.
XHigh for complex tasks or brainstorming things.
I would enjoy suggestions on when to go up/down in reasoning effort by codex though.
@jxnlco The name low means I will never use it. The option for both would be good tbh auto or choose the reasoning. I would NEVER want it choosing the model though.
Make an auto mode available and make it consume less usage as a perk for using it. Leave the ability to select a model and effort at the standard usage rate.
Codex has really fallen behind on usage allowance when compared to Claude Code. This might be a simple solution to offer more for less.
I’d only want it for reasoning effort, not between models. Does that break cache moving between different reasoning levels? I know it used to but haven’t seen anything about it in a month or two. Would be nice to know if I’m silently screwing models when I change it back and forth trying to preserve my usage.
@jxnlco For me to use low, it has to not hallucinate, it has to actually do what I ask it to do, and I don't need it to be creative, and I don't need it to have to think scientifically. I just need it to execute tasks without fail, and it doesn't do that consistently rn.
@jxnlco If you can make an auto option + the same we currently have I wouldn't mind.
In my use case I tend to play around with effort/fast mode for each task I tend to throw at codex.
Also, a slow-mode would be cool, cheaper but slower, for the time I'm sleeping :)
@jxnlco Never let the product ratchet it down from the user’s configured default or ad-hoc selection. But if you want to create an “automatically crank it up without slowing it down” option, that would be fine!
As an option.
:)
@jxnlco I want the choice, but I wouldn't mind for some tasks to delegate that choice to a competent model selector if it makes significant savings on usage
@jxnlco Depends on my subscription tier. I’m stuck on the $20 rn, so I shoot for medium, I use xhigh when I need to do make a plan, and I use high for UI related stuff
@jxnlco I like picking because I like to get a feel for it but if you could somehow learn my preferences for when I pick high vs medium vs low then I’d be okay with it auto picking for me
So manual, with auto based on my preferences
@Tally__DE @jxnlco So I neverrrrr use low either. But.. last night I was on a new thread I didnt care about and stuck it on low to see what it could do and dig up about it self (Codex) from whats saved locally on my Mac and I was blown away... it is very capable. I had no idea
@jxnlco i want to choose 5.6 but it's not there
@jxnlco can u just let me flex
@jxnlco I would absolutely never trust that choice and would not want it. That control helps me build confidence in the outputs.
I’ve been noodling on this quite a bit, I think there’s a lot of opportunities in breaking down a coding task into different levels of difficulty, and then using different effort levels based on true thinking depth needed.
I would trust it to make the right choice as long as there was transparency around the choice it made, and the ability to tune this over time.
I have to use xhigh or it makes too many bad choices on its own, so an auto select would probably assume it's okay to lower it, make bad decisions, be technically correct and leave me a pile of stuff to cleanup
there are very occasionally light tasks that can be low but even high mode succeeds at the tasks in a pretty good way, it's just.. small details, small mess ups, missing some finer things, always have to come back to it