Grok 4.1 Fast is supposed to be really

[full panther]

Do you have any more info on this by now @fiery finch ?

[fiery finch]

After a little more research I decided not to try grok models yet but I’m still interested - but I saw some people saying the grok model hallucinates a lot and is bad at code so I am testing kimi k2.5 im not sure how I feel yet idk if it really feels or does better than sonnet 4.5 - opus 4.5 definitely feels superior for openclaw but its just $$$

[full panther]

Exactly. The price is just insane. What are the differences you notice now though between opus and kimi? Are the responses dumb or do you have to itterate more?

[fiery finch]

I’m not 100% sure if it wasn’t something with my configuration I think my hooks were messed up… I went ahead and switched back to sonnet 4.5… Claude opus 4.5 recommended it when I had it review the code base said it would be better than kimi K2.5… a little suspicious, but from my experience at that point, I thought it was probably right - it said that k2.5 wasn’t as good for tool usage. I have had my open router pointed to sonnet overnight while it’s been running- a little bit scared to see how much it cost because I’ve never let it run off an API using anthropic models before.. but so far it seems with the hook configuration change and sonnet 4.5 it probably is running better overall and acting more agenticly and actually accomplishing stuff and less slop… I was having to tell it before over and over like remind it what it could do and it would be like “you’re right I should have been doing this….”

[full panther]

Exactly the experience that blows the magic. Weird it is, but sad is that people might settle with less just because of the token price.

Thanks for taking the time to share. I believe it would be nice if more people talked about this and what models, and if there is a way to fine tune things for cheaper models to accomplish somewhat the same in the talking part at least. I understand some models will always be better in coding, or tool calling etc than others. But simple stuff should be easy

[fiery finch]

I know people have set up for multiple models where it routes different stuff to certain models, but I am kind of dubious about it because you have to run it through a model to determine what model to run it through anyway and then if you end up sending stuff to a lower tier model that isn’t able to get you the right answers you’re just wasting double the tokens basically… I’m sure there’s some set up that works well, but I still I’m not sure. Need to keep looking into it.

[candid yoke]

you need to do the routing not a model

[fiery finch]

So like to set routes for specific models to specific functions or tasks? How do you figure out what goes to what route

[candid yoke]

with experience and systematic evaluation

[fiery finch]

Well I guess hopefully I can tell the bot to figure something out bc I’ve burned through about $120 of api credit since I switched to 4.6 like 4 hours ago lol 🦞🤪 yayyyy