What is the draw rate of top Xiangqi engines today?
As long as a strong engine is used (for example Pikafish, the number one AI engine for Xiangqi), the draw rate is normally extremely high.
The optimal result of Xiangqi is very probably a draw for both sides, and there are a great many optimal solutions; the probability that one side has a forced win is negligible. (Note that Xiangqi is very, very far from being exhausted; see also Has Xiangqi been solved?. But empirically, a forced win for one side can be ignored.)
Also, normal balanced Xiangqi openings have a high margin for error: the higher the level of both sides, the more likely they are to find one of the optimal moves, and the smaller the first-move advantage becomes. In a normal balanced position, the stronger the weaker side is (strength includes the engine, computing power, thinking time and other factors), the higher the draw rate. The stronger the weaker side's engine, the longer its thinking time, and the higher its hardware computing power, the higher the draw rate in normal balanced positions.
Note that this is not specific to Pikafish; humans and any engine follow this rule.
Some test results
The tests used "random, out-of-book, unfamiliar balanced openings" with no fewer than 10 major pieces on the board, simulating the first-move advantage of the Xiangqi starting position.
Both sides on identical hardware, 8 threads, 3 minutes + 2 seconds, no pondering: software of the same level reaches a draw rate of 99%+.
If one side is the latest Pikafish and the other is the web version of Pikafish (whose strength is roughly that of the 20240917 version), the draw rate under these conditions is still 98%+.
128 cores against 8 cores on the same hardware, same software on both sides, 1 minute + 0.6 seconds, no pondering, 200 games: 195 draws, and the 128-core side won 5.
8 cores against 1 core on the same hardware, same software, 2 minutes + 1 second, no pondering, 2000 games: 1962 draws, and the remaining games were basically won by the 8-core side.
56 cores against 1 core on the same hardware, same software, 1 minute + 0.1 seconds, pondering on, 365 games: 345 draws, and the 56-core side won 20.
Red's first-move advantage gives Red more wins than Black, but even so, an engine cannot guarantee it will never lose as Red.
Note
Per-core computing power differs between hardware. For example, the 8 cores of an ordinary phone may be about as powerful as 1 core of a computer with a strong single big core. But even in the 4th, extreme test (56 times the computing power at very fast time controls), the weaker side still had a very high draw rate.
The rumor that "Xiangqi has been ruined"
Rumor
In 2024-2025, a widely spread rumor went: "Xiangqi is truly ruined, thoroughly and irrevocably; there's no point denying it anymore. After being trampled by AI, Chinese chess actually has an optimal solution. That is, when two top AIs play, Red opens with the central cannon and Black just resigns."
About this rumor: "an optimal solution has appeared" is correct and not remarkable at all, but "it appeared after being trampled by AI" is wrong. Empirically Xiangqi is judged to be a draw with optimal play, but there are very, very many drawing paths, so there are also very many "optimal solutions". Empirically, drawn game records must already contain a great many "optimal solutions", regardless of AI; before Xiangqi programs existed, humans had certainly already played plenty of them. But even today's programs cannot guarantee not losing a balanced position, although the draw rate between strong engines exceeds 99%. As for "Black just resigns", there is no basis for it at all. Empirically Xiangqi is a draw; a rigorous proof is impossible with current technology, but the chance that Xiangqi is a forced win is no greater than the chance that the Earth is destroyed next year.
Testing with high-advantage openings
To test whether engine a has improved over engine b, engine authors usually use high-advantage openings with a lower margin for error, which lets statistical tools reach a conclusion faster and save computing resources. For each opening, a and b each play Red and Black once, for two games.
If an opening has a 50% chance of being drawn, then under such testing about 40% of openings produce a result of 1 win and 1 draw. No matter how strong the machines are, even in multi-threaded same-engine self-play, it is still about 40%. And even with the same variables as last time, a considerable share of that 40% may turn into one win and one loss or two draws in the next test, while a considerable share of the previous 60% of ties may turn into one win and one loss (multithreading randomness).
Against humans
Of course, top software is far beyond humans. When an engine plays a human at slow time controls, as long as the opening leaves the book (which is very easy to do against a human), the human's draw rate will be very low. For the detailed analysis, see this Baidu Tieba post.