Vibe Coding Without the $200 Price Tag: Multi-Model Workflow.

How much do you spend on monthly subscriptions?

If you’re already on a premium plan like Claude Max or ChatGPT Pro, stick with it, unless you really want to optimize your "vibe coding" workflow.

If you want to optimize your output while lowering your monthly spend, this is the stack that works best for me:

1. Designing Functional Specs: Gemini 3.5 Flash Use "high thinking" mode (30% AI / 70% Human) via AI Studio, then export to Markdown. While some people have started using HTML for specs, Markdown still works best for me.

2. Architecture: Claude 4.7 Opus (50% AI / 50% Human feedback). This stage is critical. Once the AI understands the spec, you must review the plan. It is vital to plan the folder structure and instruct your coding agent to document it. Define each folder and its purpose; store this in a reference or docs folder and link to it from CLAUDE.md or AGENT.md. Spending more tokens and time here will save you hours later.

3. Testing: Yes, write your tests first. I usually use Opus 4.7or GPT 5.5, though "max effort" usually isn't necessary. Don’t just say "make the tests." Ask the AI to draft a test plan, then review it. Brainstorm edge cases and iterate a few times. Once the cases are finalized, implement them using GLM (if you have), Flash, or Sonnet with "medium effort."

4. Implementation: Whether you divide tasks by user stories, build the full stack simultaneously, or focus on the UI first, here is my model breakdown. Always ask to run test after implementation:

   1. Backend: Opus 4.7 or GPT 5.5 for complex logic / Sonnet 4.6 for simpler logic

   2. Frontend: Gemini 3.5 Flash or GPT 5.5 (you can use stitch MCP if you like)

   3. Refactoring: GLM-5.1/Sonnet/Haiku (if it is simple one)

5. Code Review: GLM-5.1 (or GPT-5.5 if you’re rich) (80% AI / 20% Human). As others have noted, GPT models are exceptionally good at code review.

6. Once finished, update the spec document and create a handoff document. This stage doesn't require a powerful model or high reasoning effort.

This has been my personal experience. If you have a different workflow, I’d love to hear it.