My budget right now is ~$10/month, so my choices are based on cost, and the ability to get shit done. I’ve been using Opencode.
preference ranked descending
poolside/laguna-sx.2:free
BIG warning, it’s nothing short of a loose cannon. It needs a lot more instruction and it often duplicates work, makes mistakes, deletes things it shouldn’t, violates explicit policies outlined in AGENTS.md, circumvents read-only permissions to make file changes in planning mode, but it is capable of completing many small tasks and implementing simple features. If you use it, make sure you audit it’s output closely. I only use this when I run out of opencode free usage.
baidu/cobuddy:free
I’m experimenting with this one. Good initial first impressions, but sometimes it gets stuck and just sits there with a spinner, doing nothing. I can’t find any workaround and I have to switch to a different model to make any progress.
opencode-go/kimi-k2.5
This one seems good at everything. I’ve made a couple game prototypes (rust & typescript) with this and I was pleased. It usually fails the first few times, but I feed it the errors I see in my console and it eventually gets it right. The problem? Too much money. I have to spend about $5 a day to use it, which is over my budget. I use this only when free models fail, and it’s good at getting me unstuck.
openrouter/z-ai/glm-4.5-air:free
Models I no longer use
opencode/big-pickle
This was my favorite, but opencode’s free usage program is so broken that I could barely use it. I would have paid $10 a month for this, but opencode didn’t offer paid usage of this model as an option. This one was good at helping me find bugs. When I had it write code to fix a bug, new bugs often occur as side effects. It’s good at fixing the side effects too, but sometimes it takes a few attempts. I’ve seen improvements when I give the AI instructions in AGENTS.md such as, “validate all code before considering it done.” In AGENTS.md, I included examples of using tsc and svelte-check and that seemed to help.
openrouter/free
Let’s be real— this one is pretty bad. It returns large chunks of text in the thinking output for seemingly no reason. Sometimes it stops before the task is done. It’s really inconsistent because every request may be using a different model behind the scenes. It’s OK for autocomplete type tasks, but not good for implementing features.
inclusionai/ling-2.6-flash
Formerly called openrouter/elephant-alpha, this one is unusually fast, and it’s low cost, but it does a lot of bullshitting which limits it’s usefulness.
minimax/minimax-m2.5:free
This one seems like it could be good, but it regularly stops mid-thought. I am forced to re-send the query, but it stops mid-thought again. Is there some sort of limiter on this guy?