Comment by olliepro

I bet they lack good long context training data and need to start a flywheel of collecting it via their api (from willing customers)