But how could they maintain an accurate count? Maybe they were just persisting the user-friendly format alongside the actual count...

The endpoint of the API is probably just rounding the accurate number and returning a friendly number... or it's all bullshit anyway.