I'm using OpenRouter for GLM5.2, but if there's a cheaper option out there I'd love to know about it!
In the few days I've been using it, my expenses have been higher than prorating my Claude subscription to 20 working days per month.
My experience with GLM5.2 is that it doesn't overthink nearly as much as Claude Code, has better and far more concise responses (I'm so siiiiick of 10 paragraph Claude babble trying to fill out some sort of answer length target by going on tangents I'm uninterested in... I'm sure that performs better on whatever eval they're doing, but apparently their evals don't include SNR?)
I think that there are some subscriptions to go by. Z.ai subscription might still be interesting. I once haggled with kimi to get it for 1$ per month. I can only help in providing pointers:
If you wish to go Non-API but rather subscription route: Z.Ai subscription/ Kimi subscription / MiniMax subscriptions are good. You could also take a look at ollama subscription and opencode subscriptions.
If you wish to go API route: Deepseek v4 pro /mimo v2.5 pro models are comparably good if your work can do that. Codex for all its failure and for as much respect that I had within Anthropic when they had fought against the govt. which Anthropic is slowly losing again by doing some pretty dystopian actions again so Codex subscription might make sense as well.
It depends on multiple things but hopefully i am able to provide some interesting things
If you wish to run models locally, unless you are specifically buying gigs for running them locally which is almost always about privacy rather than costs, then you are always better off with qwen models so if you got a 64-128GB laptop for example. You could run Qwen models and see where things go.
Extremely helpful, thanks! I think I'l go the OpenRouter route for a while to explore various models, then weigh the option.
I do kind of like basing decisions somewhat on the API costs, because they reveal what the true costs will be after the eventual rug-pull on subscription pricing.
Even seeing the API costs of Claude Code today to a year ago are pretty eye-watering. I think there's a ton of room, at least for my workflows, to go back to far less capable models.
I've run local models in the past a bit, and explored LLM ops somewhat, and have zero desire to do it anymore, haha. It's fun as a hobby, but there's tons of other homelab stuff for me to play with.
Yeah I think that the API route model is good and it is at cost as it gets and there are some efficiencies which can be gotten from say how deepseek does its inference but at the moment as it stands, API prices are the most stable thing to go through and I wish you luck!
> I've run local models in the past a bit, and explored LLM ops somewhat, and have zero desire to do it anymore, haha. It's fun as a hobby, but there's tons of other homelab stuff for me to play with.
True. I personally haven't played enough because of my hardware being quite modest than even personal hardware recommendations but I have had sometime playing with 350 (M with million!) models like the recent LFM model and very small qwen models. They are just experiments though but I would one day like to see even more standardized models that we could use on our laptops or desktops themselves.
> Even seeing the API costs of Claude Code today to a year ago are pretty eye-watering. I think there's a ton of room, at least for my workflows, to go back to far less capable models.
Yeah exactly. I would constitute that even by using GLM 5.2 as you are originally doing even with API costs is probably much more sustainable over long run as you are currently doing. And it keeps you away from the problems of proprietary models and issues surrounding that.
In the few days I've been using it, my expenses have been higher than prorating my Claude subscription to 20 working days per month.
My experience with GLM5.2 is that it doesn't overthink nearly as much as Claude Code, has better and far more concise responses (I'm so siiiiick of 10 paragraph Claude babble trying to fill out some sort of answer length target by going on tangents I'm uninterested in... I'm sure that performs better on whatever eval they're doing, but apparently their evals don't include SNR?)