Chess Engine Sacrifice Statistics (2023-2026): Which Engine Sacrifices the Most?

What counts as a sacrifice here

A computer cannot feel brave, so we need a strict rule. Our scan marks a move as a sacrifice when all of this is true:

• The engine lets a knight, bishop, rook or queen be taken by a cheaper piece (a pawn takes a knight, a bishop takes a rook, the king takes a bishop on h7...).
• After that the engine is down 2 points or more of material (pawn 1, knight and bishop 3, rook 5, queen 9) compared with the position before its offer, and it stays down for at least 4 half-moves. So a normal trade does not count.
• The piece is taken before move 30, the game has 80 moves or fewer, and it is decisive.
• No bullet games, no Fischer Random, no odds games, nothing before 1 January 2023.

One big limit: the scan only keeps sacrifices made by the side that went on to win. So this page can tell you how often engines win with a sacrifice, but not how often a sacrifice fails. We also leave out 144 sacrifices that were still inside the opening book (the tournament chose those moves, not the engine). That leaves 1,285 sacrifices that were the engine's own idea.

Filter = decisive, 80 moves or fewer, no bullet, no Fischer Random, no odds games, played on or after 1 Jan 2023. CCRL games in our archive start in November 2025. "Own idea" leaves out sacrifices that were still inside the opening book.

Every one of the 1,429 games was found again in the archive, its capture move was checked with python-chess, and it passes the filter. The full method and the scripts are at the end of the page.

Which chess engine sacrifices the most?

Raw counts are not fair: an engine that plays (and wins) more games will find more sacrifices. So we divide by the number of wins in games that pass the same filter. The result is own sacrifices per 100 wins.

Engines with at least 25 own sacrifices in the pool, all versions together (handicapped test builds such as Stockfish_15_1M, StockfishClassical and LCZeroCPU left out). Games played = every non-bullet, non-Fischer-Random game 2023-2026 in our archive. Wins (filter) = wins in games that pass the same filter as the sacrifice scan (decisive, 80 moves or fewer). The last column only for engines with 150+ such wins in CCC and TCEC, where the opponents are the strongest.

What stands out:

• Stockfish has the most sacrifices in total (272) and the highest rate in the table: 9.9 own sacrifices per 100 wins. In TCEC, where it meets the strongest engines at long time controls, the rate is even higher.
• Torch, the Chess.com engine, gives material in 7.9 of every 100 wins. See the full rivalry in Torch vs Stockfish.
• Leela (Lc0) is famous for its style, but by this measure it sacrifices less often: 4.6 per 100 wins, the lowest in the table. Our scan only sees material given before move 30 that is still missing 4 half-moves later, so a long, slow positional sacrifice later in the game is not counted.
• Smaller engines such as RubiChess, Caissa, Stormphrax are close behind Stockfish, but with fewer games.

Treat small differences with care: a few sacrifices more or less move a rate by a whole point, and the engines did not all play the same opponents.

Which piece do engines sacrifice most?

The most common engine sacrifice is the exchange: a rook given for a bishop or a knight. Next comes the knight, usually taken by a pawn. A whole rook for a pawn or nothing is much rarer (56), and real queen sacrifices are rare too: 110 in the whole pool. To check the point values, see chess piece values.

1,285 own sacrifices. "Exchange" = a rook given for a bishop or knight (still down 3 points or less 7 half-moves later). "Rook" = a rook given for a pawn or for nothing. A king that takes the piece counts too, for example after Bxh7+ Kxh7.

When do engines sacrifice?

Most sacrifices come in the middlegame, when the pieces are out and the kings are already castled. The busiest window is moves 21-25 (389 sacrifices). Very early sacrifices (before move 11) are rare once the opening book is left out: only 52.

White or Black? 1,107 of the 1,285 own sacrifices (86%) were played by White. One reason: tournaments like TCEC and CCC give the engines opening positions where White is pushed for a win, and the side that attacks is the side that sacrifices.

Which openings lead to the most sacrifices?

The Sicilian Defence is far ahead with 300 sacrifices: sharp pawn structures and kings on opposite sides give engines targets. The French and the Caro-Kann follow.

1,249 own sacrifices in games from the normal start position; opening named by position with the Lichess opening list. In CCC, TCEC and CCRL the organisers choose the first moves, so this shows their opening books as much as the engines.

Do engines sacrifice more than before? 2023 to 2026

Engines get stronger every year, and draws get more common. Did sacrifices go down? Counting only CCC and TCEC (the same kind of events every year), the rate stays about the same: between 4.8 and 5.7 sacrifices per 100 decisive games.

CCC and TCEC only, so every year is measured the same way. All decisive games that pass the filter are counted, whoever won.

Were the sacrifices sound? A Stockfish check

We let Stockfish 18 look at every position right after the capture (depth 16, a quick search) and asked: is the side that gave the material better, equal or worse?

• Better by 1 pawn or more: 882 of 1,285 (69%). The engine had real compensation at once.
• About equal (-0.5 to +1): 383 (30%). The sacrifice kept the balance and the engine outplayed its opponent later.
• Worse by half a pawn or more: 20 (1.6%). These look like gambles, or a quick search simply does not see the idea yet. In the first game below, depth 16 gives Stockfish -1.0 after 24...dxe5, but depth 18 already gives +1.0.

So engine sacrifices are mostly not wild: most of them are judged good the moment they are made. To learn how an engine decides this, read how Stockfish works.

Top 10 biggest engine sacrifices (2023-2026)

Which sacrifices gave away the most material and still won? To keep it fair we only use games between two named engines and measure what the winner was still missing 7 half-moves after the capture.

Own sacrifices between two named engines (no weak opponents, no test builds), normal start position. Points: pawn 1, knight and bishop 3, rook 5, queen 9. The table counts what the winner was still missing after 7 half-moves, so a quick trade that wins the material back at once does not count as big. Every one of these games was won by the side that gave the material.

A funny detail: in the TCEC Season 26 "viewer submitted openings" event (July 2024), Stockfish and Komodo Dragon both played the same queen sacrifice 15.Qxa5 in the same position, once each with White, and both won.

Replay the best engine sacrifices

Four real games from the pool, one per famous engine. Every move is from the official game file; the comments and the Stockfish 18 evaluations (depth 18) are ours. Use Copy PGN or Analyse to study them in the Analyser.

How to cite this data

Method and limits, in short:

• Games: CCC event files from Chess.com, TCEC official archive (Seasons 23-30 and Cups), CCRL broadcast games (from November 2025). 906 PGN files.
• The sacrifice rule is the one described at the top. It only sees material that stays lost for 4 half-moves, only before move 30, and only in games the sacrificing side won.
• Engine versions are grouped by name (all Stockfish builds = Stockfish). Handicapped test builds are left out of the engine table.
• Openings are named by position with the Lichess opening list; 36 CCRL games that start from a set position are not named.
• All counts come from scripts we can run again: sac_stats_scan.py, sac_stats_compute.py, build_sac_stats.py.

More about the events: TCEC explained and chess engine ratings explained.

Engine Sacrifice Statistics FAQ