Building Apps Without Burning Money
A simple workflow for building apps efficiently with coding models by combining expensive and cheap models.
I keep seeing the same discussion over and over again: what’s the best way to build an app with coding models without spending a fortune?
Most of the answers fall into two camps:
- Use the most powerful model for everything and accept the cost.
- Use a cheap model and accept the lower quality.
Neither option feels great.
I’m also not particularly interested in questionable workarounds to use SOTA models (fancy account bans?). At the same time, accepting mediocre output from a weaker model just to save money seems like a waste of time.
So instead I started leaning into something much simpler: use the different models for what they’re actually good at.
Tools
This basic workflow came about from testing with these open-source frameworks:
These frameworks already encourage using different models and have structured workflows. The superpowers framework in particular seems to scale nicely for complex tasks.
The real value here is that they let you orchestrate models instead of relying on a single one to do everything.
Current Workflow
The process I’ve settled on is pretty straight-forward:
- Document & Plan: Describe the application and create a detailed plan and task breakdown using a strong reasoning model (Claude Opus).
- Implementation: Execute the tasks using a cheaper but reliable coding model (Kimi K2.5).
- Review & Refine: Run the completed work back through Claude Opus to review, test, and refine.
For simpler apps, Sonnet works perfectly fine instead of Opus.
The key idea: spend the expensive tokens on thinking, not typing.
Planning and code review are where the smarter models shine and implementation is often straight-forward once the plan is in place. Though I can’t be too confident as issues may appear in future from poor code.
Result
This approach produces something surprisingly effective:
- Better structure from strong planning
- Lower cost from cheaper implementation
- Higher quality from strong review
The final result is an app built for a fraction of the cost of just letting a premium model handle everything.
It also produces cleaner architecture because the planning stage is forced to be explicit and identify issues before any code is typed.
Funny Consequences
An amusing thing happened during a review.
Claude Opus was looking over the project, finding a bunch of bugs and then frets to itself:
“Why did I do this?!”
You didn’t buddy.. but you will be fixing it.