Centipawn - Chess Glossary
Centipawn
Definition
A centipawn is a unit of chess evaluation equal to one hundredth of a pawn. In numerical terms, 100 centipawns (often abbreviated “cp”) = 1 pawn. Chess engines use centipawns to quantify positional and material advantages: a score of +100 indicates roughly a one-pawn advantage, while −50 would indicate a half-pawn disadvantage.
How It’s Used in Chess
Engine Evaluations
Most chess engines evaluate positions in pawn units and report them in centipawns. A typical engine output like +0.70 means White is better by 0.7 pawns (70 cp). At some thresholds, engines switch from centipawn scores to mate announcements (e.g., “#5” for a forced mate in five), because a forced mate supersedes any material/positional estimate.
Note on perspective: many chess GUIs display evaluations from White’s point of view (positive = advantage White). Some tools present the score from the side to move. Always check the interface’s convention.
Player Metrics: ACPL
Average Centipawn Loss (ACPL) summarizes how closely a player’s moves matched the engine’s top choices. For each move, the difference between the move played and the engine’s best move is measured in centipawns; the average over the game is the ACPL. Lower ACPL generally indicates higher accuracy.
Annotation Thresholds
Practical (not universal) thresholds used by many annotators and platforms:
- Inaccuracy: ~50–100 cp worse than best
- Mistake: ~100–300 cp worse
- Blunder: ~300+ cp worse
These are guidelines; complexity, phase of the game, and engine depth can shift these boundaries.
Strategic and Historical Significance
From Pawn Units to Centipawns
Early computer chess research (inspired by pioneers like Claude Shannon) evaluated positions in “pawn units,” weighting material and positional features against the value of a pawn. Representing evaluations as integers in hundredths of a pawn (centipawns) made computation and comparison efficient and precise for search and pruning. The Universal Chess Interface (UCI) protocol still communicates evaluations with “cp” values.
Interpreting Scores
- Opening/middlegame: small shifts (±10–30 cp) are common and often within evaluation noise. A score around +0.20 to +0.40 typically suggests a pleasant edge.
- Endgames: even a +30 cp edge can be meaningful (e.g., a tiny but lasting structural plus or zugzwang potential). Tablebases may override cp with exact win/draw/loss or mate distances.
- Conversion is not guaranteed: +100 cp (a pawn) might be drawable with opposite-colored bishops; +50 cp could be decisive in a technical rook endgame with an outside passer.
Examples
Material Benchmarks
- Pawn = 100 cp
- Knight ≈ 300 cp; Bishop ≈ 320–330 cp
- Rook ≈ 500 cp
- Queen ≈ 900 cp
These are ballpark figures; engines dynamically adjust based on king safety, piece activity, pawn structure, and phase of the game.
Opening Example with a Small Edge
After 1. e4 e5 2. Nf3 Nc6 3. Bb5, many engines give a slight edge for White, often around +0.20 to +0.30 (20–30 cp), reflecting initiative and structural pressure without concrete material gain.
Tactical Swing (Blunder)
In a typical tactics scenario, a single oversight can spike the evaluation by hundreds of centipawns. For instance, if Black plays a move that hangs a minor piece, the engine’s score might jump from approximately 0.00 to around +300 cp in White’s favor. Hanging a queen often produces a jump near +900 cp.
Endgame Nuance
Consider a rook endgame where White has an outside passed pawn and a more active king. Engines might show a modest +60 to +120 cp. Despite the “fraction of a pawn” reading, this can be winning with precise technique—centipawns don’t directly equal win probability, especially in technical endings.
Practical Tips
- Don’t overreact to tiny changes: a shift of 10–20 cp is often within normal search fluctuation, especially at low depth.
- Look for stability: if several principal variations all favor the same plan by 50–100 cp, the plan is likely sound.
- Mate trumps cp: once a forced mate appears, the “#” score supersedes centipawns; focus on the mating line.
- Use ACPL contextually: a low ACPL in a tactical slugfest can still hide missed wins; review key moments, not just the average.
Interesting Facts and Anecdotes
- Engine GUIs often smooth noisy evaluations by rounding to one decimal place in pawn units (e.g., +0.7 = +70 cp) for readability.
- Different engines calibrate features differently. Classical engines like Stockfish/Komodo report centipawns; some neural engines (e.g., Leela) also provide WDL (win/draw/loss) probabilities that interfaces can convert to a cp-like display.
- A “quiet” improving move may only gain 20–40 cp, but high-level games are often decided by accumulating several such small centipawn improvements.
- In brilliancy-rich games (e.g., Kasparov vs. Topalov, Wijk aan Zee, 1999), evaluations can swing wildly as sacrifices are assessed more deeply; early shallow searches may mis-evaluate by hundreds of centipawns until the tactical horizon is fully resolved.