Hyperbrowser persistent browser sessions keep billing for as long as they are alive. If your agent creates sessions with session_create and forgets to call session_stop, credits drain in the background long after the task finished. The setup guide for the mux project calls this out directly: "Sessions: session_create, session_stop, session_list. Persistent browsers are billed while active - always stop them when done." For scraping work that does not need a persistent browser, prefer the stateless scrape, extract, crawl, and crawl_status calls, which the guide describes as "cheapest and usually enough." Practical rule for agents: wrap every session in try/finally with session_stop in the finally, and list sessions on startup to clean up orphans from crashed runs.