Limiting the amount of Fable usage makes us reserve Fable for our hardest problems, which is probably a very important signal when it comes to how valuable the training data is.
In other words, I think they've found a way to manipulate their customers into labeling sessions as "high quality data" (Fable sessions) and "low quality data" (non-Fable sessions).
In other words, I think they've found a way to manipulate their customers into labeling sessions as "high quality data" (Fable sessions) and "low quality data" (non-Fable sessions).