Recurrent Depth

What is recurrent depth?

If opaque recurrence is a model thinking in loops on the inside, then recurrent depth is the question: how many loops? It measures how many times those internal layers get run — like a treadmill, where recurrent depth is how many laps you've done.

How does it relate to opaque recurrence?

Opaque recurrence is the behavior; recurrent depth is the dial
The first describes "repeating a computation internally"; the second puts a number on it: it looped N times. One says what's happening, the other says how much.

What happens as depth grows?

More thorough thinking
A few extra laps give the model more chances to refine the result, and accuracy on hard problems tends to rise.
Higher cost
Every extra lap spends compute and time. Deeper isn't automatically better — you're balancing quality against cost.
Diminishing returns
Just like "thinking longer" doesn't always mean "thinking better", beyond a point the gains shrink and can even introduce new errors.

Why is the term suddenly everywhere?

As reasoning models shift from "writing their thinking out" to "quietly computing inside", we needed a word for how much internal computing is happening. Recurrent depth turns that hidden effort into an engineering parameter you can discuss and tune.

Bottom line: recurrent depth is how many laps the AI's internal thinking takes — more laps, deeper thought, bigger bill.