ChessGrader / Guides / Centipawn loss explained

Centipawn loss explained

A centipawn is 1/100 of a pawn of engine evaluation. Centipawn loss is how much evaluation your move gave up compared to the engine's best move, and ACPL — average centipawn loss — averages that over a whole game.

By the ChessGrader team · Updated July 10, 2026

That's the whole definition. Everything else — accuracy scores, move grades, "estimated game rating" — is built on top of this one number. Here's how it works, what a good ACPL looks like at each rating, and where the metric breaks down.

A worked example

Say the engine evaluates your position at +1.50 — you're better by the equivalent of one and a half pawns, which is 150 centipawns. The engine's best move keeps it at +1.50. You play something else, and the engine now says your position is worth +0.90.

You lost 60 centipawns — a bit over half a pawn of advantage. Do that on every move and the engine keeps a running tab. Sum the losses, divide by the number of moves, and you have your average centipawn loss (ACPL) for the game. Playing the engine's best move costs 0. There is no negative centipawn loss — you can't beat the engine's top line by its own accounting.

Good ACPL by rating

These are rough ranges from public community data, for rapid-to-classical games. Treat them as honest approximations, not lab results. Fast time controls push everyone's ACPL up.

This relationship is stable enough that ChessGrader uses it for its estimated game rating — a log fit of ACPL against community rating data, rounded to the nearest 25. It's explicitly a vibe, not a measurement: one game is a tiny sample, and time control changes everything.

Why centipawns get converted to win probability

Raw centipawns have a problem: the same loss matters wildly different amounts depending on the position. Dropping 100 centipawns in an equal game (0.00 → −1.00) turns a draw into a losing position — a real disaster. Dropping 100 centipawns when you're up +9.00 changes nothing; you're still completely winning.

So modern grading converts evaluations to win probability first, using a logistic curve — ChessGrader uses 50 + 50·(2/(1+e^(−0.00368208·cp)) − 1), the same family Lichess uses. The curve is steep near 0.00 and flat at the extremes, which encodes exactly the intuition above: centipawns are precious in balanced positions and nearly free in decided ones. Move grades like Inaccuracy, Mistake, and Blunder are then defined as win-probability drops (10%, 20%, 30%), not raw centipawn drops. The full formula and thresholds are on the methodology page.

ACPL vs accuracy: two views of the same data

ACPL and accuracy are computed from the same per-move losses — they're two summaries of one dataset.

• ACPL is a plain average, in centipawns. Lower is better. It's simple and comparable across games, but a single huge blunder in an otherwise clean game gets diluted.
• Accuracy runs each loss through an exponential curve and combines them with a volatility-weighted average and a harmonic mean, producing a 0–100% score where one catastrophe still leaves a mark.

If you want one number to track over time, ACPL is arguably the better one — it's less flattering and less noisy. Accuracy is the friendlier headline.

Limitations

Centipawn loss is the best simple metric we have, and it still has blind spots:

• It depends on the engine and depth. A deeper search can change the evaluation of both the best move and yours, so the same game produces different ACPL on different sites. ChessGrader uses a fixed 100,000-node search per move so its numbers are at least consistent with themselves.
• It doesn't know why you lost the centipawns. A principled positional decision the engine dislikes and a hung knight can cost the same 300 centipawns.
• It rewards easy games. Forced sequences and one-sided positions produce low ACPL regardless of skill — the same inflation problem accuracy has.
• Mates break the scale. "Mate in 3" isn't a centipawn number, so forced mates need special handling — ChessGrader's rules for that are part of its move classification system.