What this is, in plain words
AI coding assistants (like the ones built into modern code editors) can now write working software. To do it, the AI has to read and understand where your data lives before it can write a single line.
We gave two different AI assistants the exact same job, twice: once where the data lived in MongoDB, and once where the same data lived in Postgres. These are the two most popular ways to store data. They organize it differently: MongoDB keeps everything about one thing together in a single record; Postgres splits it across many connected tables that the AI must stitch back together.
We then measured how much work each AI had to do to finish the job — how much it had to read, how long it took, how much it cost. Same result, same test, every time. The only thing that changed was the database. So any difference is the database's doing.
Same job, two databases — here's the difference
Shorter bar is better — it means less work for the AI. We only ever compare each AI assistant to itself.
Text read — how much the AI had to read to finish. Cost — the bill for that work. Time — how long it took. False starts — tries that errored and had to be redone.
Does the gap grow as the database gets more complex?
We built the same feature on three Postgres designs, from simple to detailed. MongoDB stays one record every time; Postgres spreads the data across more and more connected tables.
Why the difference happens — from the AI's own session logs
These counts come straight from each AI's recorded work session. On Postgres, a JOIN is the step that stitches separate tables back together; the AI writes one for every connection it needs.
How we kept it honest
Inspect the evidence
Every number on this page comes from a real file you can open. Each file is fingerprinted, so if anyone changed a number after the fact, this page would refuse to load it.