A centipawn is one hundredth of a pawn — the standard unit chess engines use to express their evaluation of a position, so an advantage of +1.50 means the engine judges White to be better by the equivalent of one and a half pawns. The unit underlies almost every number you see in modern chess: the evaluation bar, the "+0.3" next to an opening line, the average centipawn loss in your game report, and the accuracy percentage on your post-game summary. Learning to read centipawns fluently — and to know when the number is lying to you — is one of the highest-leverage skills of the engine era.
What Is a Centipawn?
Engines need to compare positions numerically, and chess already had a folk currency: the pawn. Every club player knows the traditional piece values — knight and bishop about 3, rook 5, queen 9, all denominated in pawns. Engine programmers simply took that currency and subdivided it by one hundred for precision, the way dollars are subdivided into cents. One pawn of advantage equals 100 centipawns; a +0.25 edge is 25 centipawns; the difference between a move that keeps +0.80 and a move that keeps +0.75 is five centipawns — real to the engine, invisible to a human.
By convention the number is always from White's perspective: positive values mean White stands better, negative values mean Black does. An evaluation of -2.10 says Black is better by roughly two pawns' worth of position. When the engine sees a forced checkmate it abandons centipawns entirely and reports a mate score — "M5" or "#-3" — because no amount of material compares to mate.
The crucial conceptual step, the one I hammer on with students, is that centipawns measure everything, not just material. In the starting position both sides have identical armies, yet the engine says roughly +0.2 — that is the value of the first move. A position can be +1.5 with equal material because one side has a crushing initiative, or 0.00 with a rook deficit because the defender has a fortress. The centipawn is the unit in which the engine prices activity, king safety, structure, and time — the entire content of chess — in terms of the humble pawn.
Where the Number Comes From
Classical engines computed evaluations the way an accountant would: a hand-written function summed material, then added and subtracted bonuses — mobility, king shelter, passed pawns, doubled pawns, rooks on open files — each worth some tuned number of centipawns. The output was literally a weighted sum, and you could ask why the engine liked a position and trace the answer through the terms.
Modern engines are different animals. Stockfish evaluates with a neural network trained on hundreds of millions of positions; the network outputs something closer to an expected score — how likely is this position to be won, drawn, or lost — and the engine converts that back into a centipawn figure because centipawns are what humans expect to read. Recent Stockfish versions even recalibrated the scale so that +1.00 corresponds to a specific healthy winning probability rather than to "one pawn of material." I find this detail liberating for students: the modern evaluation was never a count of pawns you could point to on the board. It is a forecast dressed in the old currency, kept because three generations of players learned to think in it.
Depth matters too. An evaluation is only as good as the search behind it: at depth 12 the engine may love a piece sacrifice that at depth 30 it recognizes as unsound, and evaluations of quiet positions can drift by half a pawn as the search deepens. When you analyze, let the engine settle before you trust the number — a flickering +0.9/+0.4 readout at low depth is an opinion still being formed.
What +1.5 Actually Means
Here is the practical decoding table I give students, applying to typical middlegame positions between reasonably matched players:
0.00 to ±0.3 — balance. Nobody has achieved anything meaningful; the game is level even if it is sharp. Most main-line openings live here, which is exactly why they are main lines.
±0.3 to ±0.8 — a pleasant edge. One side has something: the bishop pair, a better structure, more space. Between grandmasters this is the difference between pressing for two hours and defending for two hours, but with accurate play it should not be decisive.
±0.8 to ±2.0 — a serious, likely winning advantage. Around +1.5 the engine is telling you that White has the equivalent of a clean extra pawn and a half — perhaps an actual extra pawn plus the better position, perhaps no extra material but a dominating bind. At grandmaster level, +1.5 positions are converted far more often than not; at club level, they change hands three times before move 40.
Beyond ±2.0 — winning. The engine has effectively decided the game and is now measuring how wide the margin is. The practical difference between +3 and +5 is real for the engine and irrelevant for you: both mean "win the game with normal care."
Two caveats make this table honest. First, the same number means different things in different position types: +1.5 in a sterile symmetric endgame is a nearly mathematical statement, while +1.5 in a mutual kingside attack means "White's attack objectively lands first," which is worthless information if you cannot find the follow-up. Second, the number assumes best play by both sides — it is the value of the position in the engine's hands, not in yours. Part of practical strength is knowing when a +0.5 engine line is harder for the opponent to meet than a +0.9 one; this is the entire theory of practical compensation and of playing for counterplay.
Centipawn Loss and Accuracy
The centipawn's most useful derivative measures not positions but decisions. Every time you move, compare the evaluation of the engine's best move with the evaluation after yours: the difference is your centipawn loss on that move. Play the best move (or an equally good one) and you lose zero; play a move that drops the eval from +0.8 to +0.1 and you have lost 70 centipawns. Average this over the whole game and you get ACPL — average centipawn loss — the closest thing chess has to a single objective quality score for a performance.
Rough calibration from thousands of student game reports: an ACPL under 20 in a slow game is excellent, strong-master territory; 20-40 is solid club level; 40-70 says tactics are leaking; above 80 means pieces were regularly hanging. The number shrinks in fast games' shorter, simpler battles and swells in wild ones — a sharp game honestly played can carry a high ACPL for both players because every move walked a cliff edge.
The familiar labels are centipawn thresholds too. An inaccuracy, a mistake, and a blunder are conventionally marked at roughly 50, 100, and 200+ centipawns lost, with modern sites adjusting the cutoffs by how winnable the position remains — losing 300 centipawns of a +8 advantage matters less than losing 150 of a +1.2. "Accuracy percentage" is a friendlier repackaging of the same data, mapping your centipawn losses onto a 0-100 scale calibrated so that 100 means engine-identical play. And in fair-play work, abnormally low centipawn loss across many games — especially in sharp positions where humans err — is one of the statistical signals investigators examine, though never in isolation: a forced sequence of only-moves produces perfect accuracy from any honest player.
One warning from the coach's chair: do not chase accuracy for its own sake. I have seen students steer every game into dull symmetry because trading pieces keeps ACPL pretty. The metric is a thermometer, not a target — a sharp, fighting game with an ACPL of 45 can teach you more than a sterile 12.
When the Number Lies
The evaluation is the best single number in chess, and it still fails in reproducible ways. Knowing the failure modes is what separates using an engine from being used by one.
The fortress problem. Engines evaluate by searching, and a fortress — a position the defender holds forever despite a material deficit — can sit beyond any finite search horizon. The engine shows +2.5 for many moves while the position is a dead draw; only very deep search (or the newest evaluation networks) grudgingly slides toward 0.00. When material is up but no progress is possible, trust your fortress instincts over the bar.
The 0.00 that isn't equal. A 0.00 can mean "sterile symmetric draw" or "both sides must find eight consecutive only-moves in a perpetual-check maze." The number is identical; the practical experience is not. Before an adjournment-style decision — which ending to enter, which defense to choose — always ask how the engine keeps its 0.00.
Horizon and hidden difficulty. The evaluation prices a position under best play, silently assuming you will find the tactical justification on move 6 of the main line. A +1.8 that requires a single precise quiet move in every variation is worth far less over the board than a +1.0 you understand completely. My rule for students choosing between engine lines in preparation: take the evaluation, subtract your confusion.
Style-blindness about imbalance. The engine happily reports a near-zero eval for a massive material imbalance — queen versus three minor pieces, rook and pawns versus two bishops — but "objectively balanced" positions like these are lopsided in practice, favoring whichever player understands the structure better. This is where a human still adds value on top of the number: not in out-calculating the engine, but in pricing human difficulty, which centipawns do not measure.
Using Centipawns to Improve
The engine's number becomes a training tool the moment you stop asking "what was best?" and start asking "where did the evaluation swing, and why did I not feel it?" My recommended post-game routine — expanded in our guide on how to analyze your chess games — is centipawn-driven but understanding-focused. First find the three biggest evaluation swings in the game; those are your real mistakes, whatever the annotation symbols say. For each, write down what you believed at the board, then identify what the position was actually saying — the loose piece, the wrong pawn break, the underestimated counterplay. Only then look at the engine's line. A 150-centipawn error whose cause you can name is worth more than ten best moves you copied without comprehension.
Equally important is what to ignore. Differences under about 30 centipawns between candidate moves are noise at human level — both moves are fine, and agonizing over them post-game is procrastination dressed as work. Chasing the engine's first line through a 0.15 difference teaches you nothing; understanding why your "natural" move lost a full pawn teaches you chess.
Centipawns at ChessMind AI
Our tools are built around exactly this philosophy: the number finds the moment, the explanation builds the player. Run any game through the analysis board and you get the evaluation graph, the swing moments, and — the part I care about — plain-language explanations of why the eval moved, not just a wall of variations. Played over the board? Photograph your scoresheet with Scan and the same review runs on your real-life games. If you want to know what your centipawn losses say about your overall game, the assessment locates your weakest phase, and the fix is usually a diet of tactics puzzles for the sudden 200-point swings and positional puzzles for the quiet 40-point leaks that decide equal positions.
Frequently Asked Questions
Is 100 centipawns the same as being a pawn up?
Not necessarily. The scale is denominated in pawns, but the evaluation prices everything — activity, king safety, structure, initiative — so a +1.00 can describe an extra pawn, or equal material where one side's pieces dominate. Modern engines calibrate the number closer to winning probability than to countable material.
What is a good average centipawn loss?
In slow games, under 20 is excellent, 20-40 is strong club level, and above 70-80 indicates regular tactical accidents. Context matters enormously: sharp games inflate ACPL for both players, fast time controls change the baseline, and a simplified drawish game deflates it. Track your trend across many games rather than judging any single number.
What does an evaluation like +1.5 mean in practice?
The engine judges White better by the equivalent of a pawn and a half under best play from both sides — a serious, usually winning advantage at master level. In practice its meaning depends on the position type: in a quiet ending it is close to a verdict, while in a double-edged attack it only says White's resources objectively suffice, with no promise the path is findable.
Why does the engine sometimes show 0.00 when one side is up material?
Because evaluation is a forecast of the result, not a count of the pieces. If the side down material has a fortress, a perpetual check, or enough positional compensation to hold, best play leads to a draw and the honest evaluation is 0.00. Conversely, engines can be slow to recognize fortresses, showing a healthy plus for positions that are actually unwinnable.
What does M5 or #-3 mean?
The engine has found a forced checkmate and switches from centipawns to mate distance: M5 means the side to move mates in at most five moves, and a negative sign means the opponent is doing the mating. Mate scores outrank any centipawn figure — +9.00 is a smaller advantage than M12.