steel is a horrible material for a nervous system
how is THC meant to run on that
Overall my experience is similar. The top tier models are pretty damn good, add some facy skills and they get even better for single projects. But try to approach a huge legacy codebase, be it monorepo on 100+ repos and you get properly shafted by the cost of doing anything. A 20 ish $ sub lasts about an hour at best.
So you’re using a really crappy model for agent work and extrapolate all the problems you encounter to actually capable models? That’s not really how anything works.
If Cursor is this bad with quotas, get a ChatGPT or Claude subscription. The $20 tiers are fine even for large codebases. For extended autonomous work (goal or loop mode) you would want the higher tiers. Use at least Terra 6.1 or Sonnet 5.5 for your regular conversation and if needed, let them create cheaper subagents for large, repetitive tasks.
They did write why they are not using better models. Hint: prohibitive cost at scale!
IMHO. “Use a better model” ignores the real issue they are writing about.
I answered the cost point
Fair enough. Maybe I’m guilty of “using it wrong”. But even if you let the “main” model delegate to cheaper models it still tends to spend quite a large amount of input tokens on building context before it starts to delegate and later on output tokens on the changes itself.
Input tokens are cheap, those shouldn’t cause your quota to run out too quickly. What you can try is using some sort of YAGNI enforcement like the Ponytail plugin. Also giving some instructions in your personal AGENTS.md for the model to give more efficient answers can help.
What plan are you using? Currently I think that ChatGPT plus can be a bit tight if you’re not careful. On my Claude plan for work I can use Opus all day and it barely scratches the quota. I’m not sure which Team seat I have though, maybe it’s the premium one.
Also, install the superpowers plugin. I’m not saying it’s always the best under all circumstances, but if you are new to agentic development and don’t know how to write your own system, it’s pretty good.
