One engineer building a full finance SaaS, cycling through four AI models before finishing it.

Claude for debugging and refactoring. Gemini for UI design. Free models for early prototyping. Paid OCR once the accuracy actually had to hold.

What I found interesting: not one model for everything, just because it’s convenient. Each one where it actually works.

I’ve ended up in the same place without planning to. When I’m building an extraction workflow, Claude is better at reasoning through edge cases and fixing logic errors. For processing a high volume of documents cheaply, a lighter model handles the first pass and only uncertain results go to something heavier. For visual tasks — reading a poorly scanned document, parsing handwritten fields — a vision-optimized model is worth the extra cost.

The thing he mentioned about free OCR: 80% accuracy sounds acceptable until you remember these are financial documents. One in five invoices with at least one wrong field. At low volume you catch it in review. At scale it becomes a real operational problem.

Accuracy requirements in finance are stricter than most contexts. 95% isn’t enough either. You want 99%+ with a clear review queue for the rest.

The part of his story that sticks: one person, the right tools, building something that would have required a team five years ago. That’s not a curiosity anymore. It’s changing what’s realistic for small firms and solo operators.

The models keep getting better. The question is whether you’re using the right one for the right job.