I just upgraded to Cubase 9.5 after a looong time without upgrading.
I work with lots of MIDI tracks and sample libraries, all in the box, no hardware instruments or effects. I remember it used to be good practice to export audio mixdown in realtime to be on the safe side when printing a mixdown for delivery. I was wondering if this is still the case. No point in sticking to that old habit if there are no benefits 
.
Myself I don’t hear any difference in online/offline Mixdown.
I don’t hear any difference in general, but there are situations in which realtime might be suitable.
For example, I use BFD which has an “Offline Mode” meaning that in offline rendering it doesn’t return control to Cubase until it has completed the tails of cymbals. Generally I freeze BFD when a production is complete and then I can render offline, but if I am making a work-in-progress mp3, then rendering in real time stops those cymbals potentially being cut off short. I am guessing that there are other vstis that might behave in the same way.
That’s a bunch of nonsense.
It has never been recommended, unless you use buggy plugins that forces you to do so.
I haven’t run in to any the last decade.
Realtime is only needed if there is something realtime used in the project aka real synths, or real outboard.
When the whole project is ITB you should just render.
Such certainty and ignorance in just a few words, breath-taking.
Here’s a direct quote from BFD3 manual.
"Offline mode
This switch is intended for use with hosts that provide an offline (non-realtime) bounce or mixdown facility. When the Offline switch is activated, BFD3 waits for all data to be properly delivered from the hard disk before allowing the host to continue, ensuring that sounds are not cut off before their full decay is complete.
Please note that Offline mode is non-realtime, and should only be used during offline mixdown or rendering in your sequencer, if it even offers this feature. Do not leave the Offline switch enabled during regular realtime playback.
If your sequencer only provides real-time bouncing facilities, you should leave Offline mode turned off at all times."
And what exactly makes you think djw was responding to you rather than commenting on the OP question? 
Lol Grim, good point, low self esteem maybe? …nah but apologies to all concerned.
It just seemed like a brutal put down (aimed at whoever?) with no material content or constructive criticism, in fact of no value at all. That’s a bit of a red rag to me.
In direct response to OP. There are situations in which both on and offline make sense. Don’t just go with herd mentality, figure out what works for you. You can always run a null test. Keeping this simple: Make pair of renders on and offline respectively. Put them into a new project on two stereo tracks, in exact alignment and same volume, then reverse the phase of one of them. If silence is the end result, then wavs are identical, of not then that is the difference that you would need to be aware of. Also be aware that freerunning choruses phase and pan effects will ruin the test, so use material that avoids that sort of thing. (Edited)
Null testing is often a good way for people to convince themselves that something actually is wrong when it really isn’t.
If you must null test, be sure to run two passes without changing anything.
If this also fails to null to a similar degree then your test is not telling you anything.
Does automation enter into the OP question?
For example, if the automation is dense with abrupt value changes, would it be better to render non-real time, “so Cubase can take all the time it needs”?
Except if you have modulation effects or synths which emulate analog gear which might not produce the same exact results on each pass. That would mess up your null test, wouldn’t it?
You’re quoting a single line of my sentence out of context which is missing the point I’m making.
Say you run 2 realtime and they null completely but then you run 1 offline and it doesn’t null you maybe proved some difference between realtime and offline. To be safe you should run another realtime and that should null to the other realtime.
If however the 2 realtime don’t null, you can look at how far they do null…if the realtime and offline null a similar level you can assume no significant difference, but if one nulls 5db different then you proved a difference again.
If you didn’t run two passes without any change and only ran one realtime and one offline then you have not validated the test.
Thanks everyone, for your input. I’m going to summarize this as follows:
In general realtime audio mixdown is not necessary, unless you’re using hardware instruments/effects or if you run into audible problems with certain plugins (in which case a realtime mixdown is worth a try).