To add to that, Image Server (or Image Server extension pre 10.5), is an additional license, so that may have to be taken into consideration.
However, even if you end up caching, in my tests, caching of an image service is much faster (~1/8 the time) than caching a map service. Dropping a 8 week cache job to ~10 days, while adding 5 more scales makes it pay for itself, in my opinion. Plus, the Image Server gives you the option to do either.
re: viewing as a basemap. If using a cached service, you will be limited to those scales when creating a web map (ArcMap will interpolate between cache levels/scales). This can work to your advantage if you need to control how users are viewing the maps, but can work against you if you do not have enough scale levels.
These are just my opinions. Results may vary.