A Position That Baffles Computers

EVERYONE knows analysis engines are very strong these days, but there are still positions they find hard to evaluate correctly.
Most such positions occur when there are lots of men still on the board, but in my most recent tournament, the Devon County Congress at the Livermead House Hotel earlier this month, the following example arose after White's 60th move.

Spanton (169) - Daniel Gibbs (153), Rowena Bruce Challengers 2017

The game continued routinely: 60...Nd7 61.Kd4 Kxh6 62.Kxc4 Kg6 63.Kd5 Kf6 64.a6 Nb6+ 65.Kd4 Nc8 66.Nb5 Ke7 67.a7 Nb6 68.Nc7 Kd7 69.a8Q Nxa8 70.Nxa8 Kc6 1-0.
But the next day DG told me he actually had a draw, and this does seem to be the case - but you have to give analysis engines plenty of time to find it.
The key line runs: 60...c3 61.a6 Nxe4! 62.a7 c2 63.Kxe4 (63.a8Q c1Q+ doesn't change the outcome) c1Q 64.a8Q Qh1+ with a trivial draw.
So it seems 60...Nd7 should have a question mark (under the John Nunn endgame convention).
Analysis engines eventually realise White is not winning in the diagram position (although many continue to believe White is much better), but it certainly baffles them at first.