The book "Artificial Intelligence: A Modern Approach" (4th edition, global version) says
"With an admissible heuristic, A* is cost-optimal...".
An admissible heuristic is one that never overestimates the distance to the goal, while a consistent heuristic is one that satisfies the triangle inequality, meaning that the cost of reaching the goal through a particular path is no more than the cost of reaching the goal through any other path. A* is cost-optimal in a tree search when using an admissible heuristic function, but it needs the heuristic function to be both admissible and consistent for optimal results in graph search.
I'm a bit confused by the AIMA statement. Are they saying that A* can be non-optimal if it uses an admissible but inconsistent heuristic with the graph search version?