Minions

Uber's AI cost per session fell 18 points further than its cost per request

The gap implies each session now uses roughly a quarter fewer requests, so the compaction threshold and the prompt cache in Uber's own harness are carrying part of a saving usually credited to cheaper open models.