It seems to have been a combination of a traffic surge and a third-party API being temporarily slow. The slow API meant that I ended up paying for more idle CPU time during the request round-trip.