The biggest factor: tool call count

How long a turn takes is dominated, more than anything else, by how many tool calls it actually ends up making, not by how long or detailed the original prompt was.

A single, well-scoped fix, read one file, make one edit, is usually the fastest kind of turn, since each round trip in that chain, read the file, get the result, make the edit, is quick on its own and there are simply few of them chained together.

A broader investigation, tracing a bug through several files with repeated grep calls before ever proposing a change, naturally takes longer for the straightforward reason that more steps are genuinely involved, each one waiting on the real result of the last before deciding what comes next.

This is also why the length of your own prompt is a poor predictor of turn duration on its own. A long, carefully detailed request that removes ambiguity can finish faster than a short, vague one that ends up needing several exploratory steps just to figure out what was actually meant.

Phrased differently: you are paying, in time, for however much the model has to discover on its own, not for however many words it took you to describe the goal. A precise request is often the faster one for exactly this reason.