On Claude Pro, Kagi Assistant and Charm Hyper

I’ve cancelled my Claude Pro subscription. It wasn’t out of any dissatisfaction with Claude; on the contrary, I still think it’s the leading LLM for general use and especially for coding.

No, my choice was more because I realised I was wasting money.

I already pay for Kagi, and even upgraded to their Duo plan so I could give my wife a search engine that works for her (versus the other way around). As part of that subscription, I can use Kagi Assistant.

For general-purpose ‘consolidated search’ — answers — I ran tests of asking Claude and Kagi the same question, and more often than not they came out very similar, if not the same.

I don’t do a lot of programming, but I probably end up running Claude Code once a week to do something far faster than my rusty tech chops can muster. All of my server machines run FreeBSD, which Claude Code recently ended support for when they moved to shipping their own package versus distribution via npm.

I’ve been a fan of Charm’s mods for some time, which led me to their CLI coding tool, Crush. Now Charm have extended Crush with a subscription inference service called Hyper.

Crush itself, being a Go-based binary, is considerably snappier to use than CC anyway, plus it’s entirely LLM-agnostic, simply needing API tokens. With Hyper, one can now also access some very-keenly priced LLMs that are often almost as good as Anthropic’s models. Although Hyper doesn’t yet support Kimi K3, released this week and looking like a serious competitor to the leading frontier models, you can bet it will soon (Charm are moving quickly with Crush and Hyper).

So I’m returning to buffering up some Anthropic tokens for when I really want to use Claude, either in Crush or for chat in TypingMind. Otherwise, Kagi Assistant works really well.

Not in the slightest bit important, but as an aide, Kagi’s logo doesn’t also look like a butthole.