We're rolling out a feature that lets you vibecode Excel models, the same category an a16z-funded startup has been getting attention for. Building it out, I realized we don't just match what's currently considered best. We beat it.
Here's our output for the prompt "build a detailed LBO model for ExxonMobil": https://docs.google.com/spreadsh...
Run that same prompt on any other platform and the gap is hard to miss.
How we did it: we generate Excel server side instead of client side, using openpyxl as the engine instead of SpreadJS. We built on top of openpyxl to support pivot tables and the rest of what a real model needs. The reasoning behind the bet is that openpyxl is heavily represented in training data, so the model actually knows how to drive it. That's the difference between a spreadsheet that looks right and one that is right.
