Can cost_drift catch a cost that creeps up slowly, not just a sudden spike?

Yes, because cost_drift checks two references at once, and each one is blind to what the other catches. One is the call site’s rolling recent normal, which catches a sudden jump, like a deploy that doubles token usage overnight. The other is pinned at the last deploy, a fixed point rather than a moving one, so a change that inches up over weeks still reads as far from where things stood at launch.

A rolling baseline alone would absorb slow creep a little each window until expensive became the new normal with no finding ever firing. A pinned baseline alone would keep flagging a change forever, since it never resets. Running both means creep gets caught by the one that doesn’t move, and a real jump gets caught fast by the one that does. Only a person resets the pin, once they’ve confirmed the new normal is actually fine.

keep reading

More on this.

Does cost_drift work for a scheduled agent that has no user sessions?Yes. cost_drift tracks dollars per turn at the call site, not per user session, so a scheduled job's runs build a baseline the same way.answer →Is cost_drift turned on by default?Yes. cost_drift is one of three classifiers Tessary calls deterministic and turns on by default, alongside duration_drift and tool_error.answer →How do I tell traffic growth apart from a broken agent in my LLM bill?Look at dollars per turn, not total spend: growth adds turns at the same per-turn price and leaves the ratio flat, while a defect moves the ratio.answer →Everything on tessary cost driftThe definition, and every question answered under tessary cost drift.concept →

Send us the traces you already emit.

book a demo