What Is a Centipawn? How to Read a Chess Engine Evaluation

What +0.4, −1.8 or M3 on an engine bar really mean, how centipawns map to winning chances, and when an evaluation misleads you.

Every engine shows you a number: +0.4, −1.8, M3. It sits on the evaluation bar, next to the best line, at the top of the analysis panel. Most players glance at it, see whether it is green or red, and move on. That is a waste, because the number tells you a lot more than “who is better” once you know how it is built and where it lies.

A centipawn is one hundredth of a pawn

The unit is a centipawn. One hundred centipawns equal one pawn, so +1.00 means the engine thinks White stands about as well as if White were a clean pawn up. Positive numbers favour White, negative numbers favour Black, and 0.00 is a dead-equal position.

The scale is anchored to material: a pawn is worth 1, a knight or bishop about 3, a rook about 5 and a queen about 9. But the engine also prices in everything that is not material — king safety, active pieces, passed pawns, weak squares, space. That is why you see +0.7 in positions where nobody has won or lost a single pawn.

A rough scale for reading the number

These bands are conventions rather than laws, but they are a useful first reading:

• 0.00 to about ±0.30: equal. Either side can play for a win without taking a risk.
• About ±0.30 to ±0.80: a slight edge. The better side has the easier game, but nothing is decided.
• About ±0.80 to ±2.00: a clear advantage. With accurate play it is usually converted.
• Beyond ±2.00: a large advantage, comparable to being a piece or a couple of pawns up. Beyond ±3.00 it is normally decisive.

Why a centipawn score is not a probability

The engine assumes both sides play perfectly from here on. Humans do not. A +1.0 position is a long way from a guaranteed win when the player with the edge has to find a series of only-moves, or when the clock is nearly out.

That is why modern analysis tools convert the evaluation into an expected score. Lichess publishes the curve it fitted to real games: +1.0 corresponds to roughly a 59% expected score for the better side, +3.0 to about 75%, and +5.0 to about 86%. Even a three-pawn lead leaves the opponent a real chance in practice. The curve also explains why going from +9 to +6 is not a blunder — the winning chances barely moved — while going from +0.5 to −2.0 is.

Reading mate scores

When you see M3the engine has found a forced mate in three moves. A minus sign (M−3) means it is Black who delivers mate. Unlike centipawns, mate scores are exact: there is no “slightly winning” version of a forced mate. If the number of moves keeps growing as you step through the game, the attacker is missing the fastest route but still winning.

Why the same position can show different numbers

Depth matters. At low depth an engine can be wildly optimistic about a sacrifice that a deeper search refutes. Different engine versions and neural-network files also disagree by a few tenths of a pawn. As a rule, do not chase differences smaller than about 0.2 — they are noise. What deserves your attention is a large swing between two consecutive moves.

When the evaluation misleads

• Fortresses. An engine may show +5 in a position that is a dead draw because the defender can lock up the position and never be broken down.
• Closed positions. Many closed structures look roughly equal to an engine, while a human has no idea how to make progress or how to hold.
• Human practicality. A move rated −0.3 that needs five precise follow-ups is harder to survive than a +0.2 move that plays itself.

How to use the number in your own games

Start with the swing, not the absolute value: find the move where the evaluation changed the most and ask what your opponent’s reply threatened. Then check the best line and work out why it works before reading the next suggestion. If you just copy the engine move you learn nothing; if you understand the idea behind it, you gain a pattern you can reuse.

Our guide to move classifications explains how those swings become labels like Mistake and Blunder, and what a good accuracy score is shows how they roll up into one number for the whole game. To see both on your own game, import it into Chesslume for a free Stockfish report.