A comparison of free limits of AI-powered coding tools and my
personal thoughts after testing them.
Note that this is not a scientific comparison. Just a personal overview after trying each one
during vibecoding web apps. Using free limits will be enough to have fun and make a couple of
small
projects or 1 bigger. Don't expect too much. When you try to conserve tokens, use smaller models
and set reasoning to low/medium. I developed apps for ~30 hours using ~80 AI requests so about a
weeks worth of work before hitting limits on all tools. Some are montlhy so this would not be
enough for a full-time development even with a decent amount of time spent on analysis without
AI.
I tested plugins in VS Code and standalone applications on Windows computer.
Local development on a laptop with Intel Arc 140v with 16gb/18gb shared memory (32GB of RAM) was
unreliable for agent coding tasks. I managed to run with llama.cpp:
gemma-4-26B-A4B-it,
gpt-oss-20b, Ornith-15-9B, Qwen3-8.9B-Distill at 5-12 tokens/s. It was too slow to handle tasks
effectively. When reasoning was set to minimum it was able to do at least anything. Otherwise
thinking could not end.
I did not care about code completion as I was mainly vibecoding, but most of those tools can do
it.