I get my monthly salary for being a senior dev in a startup.
Claude Code User
I have been using mainly Claude Code till now. As I use Neovim as my code editor, using a CLI based coding agent was easier.
I have a team account at work and a personal account that my wife and I share. She is using the web interface and I use Claude Code for hobby projects.
My Claude Code experience has been mostly flat from Opus 4.6 onwards, the 1M context window was probably also the deciding factor. Ever since then, I feel that the model performance has stayed flat for my day-to-day work. Of course the drama around Fable was interesting, and is also a reason why I am interested in open weight models.
I listen to some dev podcasts and heard many people praise pi, I want to give it a try, both at work and personally.
Considering Pi and Open Weight Models
Hackernews is my news source/social media provider. Every now and then in the past few weeks I have seen discussions around open weights models. With the drama around Fable release and usage in the EU, I want to evaluate switching over to pi and open weight models. I want a quick way to test out open weight models and came across openrouter. I have heard about in one of the podcasts but never really had a chance to actually understand what it is. I took some time, opened an account, added some credit and started using it in my local pi setup.
Openrouter is positioning itself as a model marketplace, they don’t self host any of the open weight models. They are routing the traffic to one of the many providers who do. From the developer perspective, this is pretty neat as I need one API key to access any of the models, including the paid ones. Openrouter has the first mover benefits, but the space would get competitive and margins drop eventually as there is not much USP for a single company.
Initial trial with Pi and GLM 5.2
I tried pi in my blog project running Hugo. I wanted to include pre-commit checks for spelling and grammar using prek. First slight surprise was that pi did not include a web fetch tool, but it built one for itself using node. It took awhile and some /tree reverting to get it right, as I made the mistake of polluting the context with the task of pre-commit configuration and the web fetch tool building for pi itself.
Eventually it settled with using typos for spell check and harper. Pi was printing the full conversation history and it was nice to see the back and forth of GLM 5.2 reasoning. Several times it said I'm overcomplicating it. As Claude Code doesn’t print the full conversation by default, and I did not bother to enable it, this statement was surprising. But knowing that it is a next token predictor that is harnessed into solving a narrow task, I am not surprised.
The full pre-commit setup and the pi’s web fetch tool building took around 0.8 &euro. I am impressed with the first trial, I will keep using it in my personal projects from now on.