Claude 3.7 Sonnet: Extended Thinking

HRA42

Claude 3.7 does support extended thinking like R1 and o3-mini.

https://www.anthropic.com/research/visible-extended-thinking

I would suggest adding either a different model or a toggle, so the user can choose which mode is required for the task. By default, the normal mode should be active to safe cost.

azdanov

GitHub Copilot did it this way:

RoxyRoxyRoxy

azdanov This seems... inelegant to me ngl. I'd much prefer a separate toggle for extended thinking, that seems much more intuitive.

kaghat

The copilot way seems fine to me. It looks a little bit like what we already have with "deepseek chat"/"deepseek R1". The fact that anthropic decided to choose names that are similar is out of our control.

RoxyRoxyRoxy

kaghat Yes but V3 and R1 are different models. If I'm understanding this extended thinking right, it's a toggle that can be enabled for the same model, much like Kagi's Web Access. We don't have different models in the list for Web Access, neither should we for this extended thinking if it gets added imo.

HRA42

RoxyRoxyRoxy yes, it's just an API option that can either be enabled or disabled. The difference is that all models support the function call for web search, for thinking it could be the better option to have this a separate model as it is super expensive to run.

(it's a loop that feeds the context over and over up to 50 times to model)

HRA42

Openrouter did go the way of adding an extra model for thinking, so the general trend tends to be to add it as an extra model and not include an option in the parameters:
https://openrouter.ai/anthropic/claude-3.7-sonnet:thinking

kaghat

HRA42

That's pretty much how I thought about it, yes.
I don't know the details in the implementation, and I don't want to get pedantic about something like this, but "adding thinking", feels like it would be a different model, just because of the added capabilities.
Anyway, I'm on the "add it as a different model" team.

gladiator2339

I personally think that the ultimate goal here is to be as smooth and seamless as possible: no model selection needed, model decision happens based on the prompt automatically... That will eventually make the experience much smoother than it is now, and maybe Kagi can also cut cost by doing this (we don't need Claude 3.7 to summarize a few paragraphs).

RoxyRoxyRoxy

gladiator2339 I think the problem there is that (to my knowledge) Claude can't switch between extended thinking and not by itself. That would entail Kagi building their own classifier for whether extended thinking is needed or not. That's quite an investment I'd imagine. It's also unlikely to be correct 100% of the time, meaning you'd want a manual override, and at that point you might as well just build the override and not make your own classifier.

lordofgreen

Hey, please add Sonnet 3.7 as a thinking model. Would be great to have next to R1 and o3-mini.

gladiator2339

lordofgreen It's already available