What is stopping you from switching to GLM 5.2 now? Have you tried it out yet?
I'm playing with Deepseek a lot more via OpenRouter recently and the only major downside I can see is the usage billing over the monthly plan
What is stopping you from switching to GLM 5.2 now? Have you tried it out yet?
I'm playing with Deepseek a lot more via OpenRouter recently and the only major downside I can see is the usage billing over the monthly plan
Existing setup. I'm already used to Claude Code I guess. I actually spent time and tokens customizing and fixing it. Have a patcher utility that modifies the binary in order to disable telemetry and remove performance reducing language in the system prompts. Every update I spend some Claude tokens dissecting the newest executable and integrating it with my patcher.
The monthly subscription is also a major hurdle for me. The "high end frontier models for low prices" aspect is a major reason. I think I'm getting a lot of value from my subscription, given that the API prices are like a hundred times higher.
However, there's also a psychological factor here. These subscriptions are like the gachas of the software world. I got "addicted" to them. I developed workflows around achieving 100% weekly usage. Sometimes Anthropic randomly resets weekly usage and I scramble to get the most out of it. I'll point Claude at things and then just have it run hundreds of code review agents. I ran out of projects to do this on and started doing it to my favorite open source projects instead, looking for things to contribute.
I think with usage-based pricing I wouldn't use LLMs quite so freely. It'd probably cure my "addiction" too, but the problem is I'm not sure whether that's a good thing, since this "addiction" has been a somewhat positive force in my life. It's driven me to start new projects and also make major progress on existing ones. It brought me out of a slump. I'm a little afraid of moving away from Claude and not being as driven as I was before.
Are you me? Anyway I wonder if you've played with the following, they've been what I've been using to patch CC:
https://github.com/skrabe/tweakcc-fixed https://github.com/skrabe/lobotomized-claude-code
Yeah I know about tweakcc, even had some custom prompts committed to my dotfiles repository for a while. For some reason it wasn't working for me, don't remember why.
At first I used this shell script here:
https://gist.github.com/roman01la/483d1db15043018096ac3babf5...
Someone posted it on HN on the Opus degradation thread:
https://news.ycombinator.com/item?id=47660925
I just kept using it to fix the prompts for a long time. Eventually I asked Claude to port it to Python to make it easier to maintain. Now every month or so Claude dissects the Claude Code binary and integrates the script with the newest version.
https://github.com/matheusmoreira/.files/blob/master/%7E/.lo...
Meanwhile, as an irregular user of heavy AI, I like the usage-based billing of deepseek. Made a bunch of optimizations to my vibe-coded codebase, created some extra modules, using the chat... has costed me 33 cents so far in a couple weeks.
Yeah for a side project I just switched over to DS4 and between Pro and Flash a few days ago I've spent $1.96 both for code work and some large LLM text processing tasks.
It feels to me as good as the last version of Opus I was using (I think 4.6 or 7?). I'm going direct through the DeepSeek API, not Openrouter.
33 cents? That's insane... I set up an API cost meter on my Claude Code status line and it frequently runs into the hundreds, sometimes thousands of dollars.
My build/work is probably a lot simpler than yours, but still, a huge value for me for just 33 cents.
Minimum spend/fill is US$2 plus Chinese vat (dunno why but not gonna complain about 6%)
For the 100$ I'm paying for Claude, I'm pretty sure I can use Deepseek way more than I can use Opus via the plan.
The downside are opportunity costs - not using other models to make better decision, I guess, right?
usage billing over the monthly plan when deepseek is over x25 cheaper?