# What is centipawn loss?

> What is centipawn loss? How much engine evaluation a move gives up, in hundredths of a pawn. How Lichess computes ACPL, and where it misleads.

- URL: https://thechesscoach.netlify.app/answers/what-is-centipawn-loss/
- Updated: 2026-11-02
- Author: Ignacio Soler

**Centipawn loss** is how much of the engine's evaluation a move gives up compared with the best move, measured in hundredths of a pawn. If the best move keeps you at +1.50 and your move leaves +0.90, that move lost 60 centipawns. **Average centipawn loss (ACPL)** is that number averaged over all your moves in a game; lower means closer to the engine.

One pawn equals 100 centipawns, as the [Stockfish FAQ](https://official-stockfish.github.io/docs/stockfish-wiki/Stockfish-FAQ.html) explains (checked November 2026). Below is exactly how Lichess computes ACPL, what my own numbers look like, and two real moves that show why a centipawn isn't always worth the same.

## How Lichess computes average centipawn loss

Lichess shows ACPL in the summary of every computer analysis, next to accuracy ([Lichess analysis board explained](/guides/lichess-analysis-board-explained/)). Its open-source code is short and specific (checked November 2026):

| Step | What Lichess does | Source |
|---|---|---|
| 1. Evaluate | Takes the engine evaluation before and after each of your moves | [AccuracyCP.scala](https://github.com/lichess-org/lila/blob/master/modules/analyse/src/main/AccuracyCP.scala) |
| 2. Cap | Limits every evaluation to ±1000 centipawns (10 pawns); a forced mate counts as the cap | [eval.scala in scalachess](https://github.com/lichess-org/scalachess/blob/master/core/src/main/scala/eval.scala), [eval.scala in lila](https://github.com/lichess-org/lila/blob/master/modules/tree/src/main/eval.scala) |
| 3. Measure | Loss = how much the evaluation dropped from your point of view | AccuracyCP.scala |
| 4. Floor | A move that improves the evaluation counts as 0, never as negative | AccuracyCP.scala |
| 5. Average | ACPL = the mean of those losses, rounded | AccuracyCP.scala |

The cap matters. Without it, one move that throws away a forced mate would add thousands of centipawns and swamp the whole game.

## What my centipawn loss looks like

I computed ACPL the Lichess way on my own Stockfish analyses: 1,470 games of at least 20 moves each side, played as nacsol3 between January 2024 and October 2026, rated 707 to 1385.

| My games | Median ACPL |
|---|---|
| Wins (733 games) | 47 |
| Losses (678 games) | 80 |

Across those games, ACPL and accuracy track each other closely (correlation −0.82: higher loss, lower accuracy). They aren't the same thing, though, and the next two moves show why.

## When a big loss barely matters

December 21, 2025. I'm White (1320) against julermitmuler (1318) ([game on Chess.com](https://www.chess.com/game/live/147009194938)). I was completely winning, with a pawn one step from promoting.

> **Move 38: a rook check was faster, but I was winning anyway**
> Position (FEN): `4R3/pppR1P1p/7b/6kB/8/r5P1/1r3PKP/8 w - - 5 38`
> Played: g4 — Pawn to g4, still completely winning
> Better: Rg8+ — Rook check on g8, then the pawn promotes with check

The engine's evaluation went from **+11.06 to +6.87**: more than four pawns. Lichess's cap counts it as a loss of 313 centipawns (from the capped +10.00 down to +6.87), a big number for one move. But my winning chances only went from 97.5% to 92.6% by Lichess's Win% formula, under 5 points, so the move gets no inaccuracy, mistake or blunder label. I won the game.

## When a small loss decides the game

April 10, 2024. I'm White (989) against machoem (1108) ([game on Chess.com](https://www.chess.com/game/live/106523877781)). My bishop on f4 was attacked by a pawn, and my knight was attacked by the queen.

> **Move 15: taking the knight let the pawn take my other bishop**
> Position (FEN): `2rq1rk1/p4pbp/1pn2np1/4p3/Q4B2/4PB1P/PP1N1PP1/R4RK1 w - - 0 15`
> Played: Bxc6 — Bishop takes the knight; the pawn then takes the bishop on f4
> Better: Bg3 — Bishop to g3, slightly better for White

I took the knight on c6. Black didn't take back: the pawn took my other bishop instead, and the position became slightly better for Black. The evaluation went from **+0.71 to −0.95**, only 166 centipawns, about half the loss of the move above. But this one cost me 15.2 points of winning chance (56.5% to 41.3%), which is a **blunder** on the Lichess scale ([the thresholds](/answers/inaccuracy-vs-mistake-vs-blunder/)). I lost the game.

That's the weakness of centipawn loss: a pawn near equality is worth far more than a pawn when you're already winning. Lichess's [accuracy page](https://lichess.org/page/accuracy) says it directly: centipawns are great for developing engines, but not so much for human comprehension. That's why it built accuracy on winning chances, which I cover in [what a good accuracy is](/answers/what-is-a-good-accuracy-in-chess/).

## How to use centipawn loss

- **Compare games only with your own, analysed the same way.** Engine, depth and caps all change the number.
- **Don't chase a low ACPL in won positions.** Big losses there may not matter.
- **Sort your moves by lost winning chance, not by centipawns,** to find the moves that actually changed the game. The [guide to reading Stockfish analysis](/guides/how-to-read-stockfish-analysis/) shows how.

## How Chess Coach uses the evaluation

Chess Coach, a connector for Claude and ChatGPT, reviews finished Chess.com and Lichess games with Stockfish on our own server. It doesn't show ACPL. It turns each evaluation into winning chances with the Lichess formula (also capped at 10 pawns), finds the moves that lost the most, and your assistant explains them in plain words on an interactive board, with the evaluation graph and red and green arrows. The free plan includes one in-depth review a day. [Connect Chess Coach](/connect/).

## FAQ

### What does average centipawn loss mean?

It's the average, over all your moves in a game, of how many hundredths of a pawn each move lost compared with the engine's best move. Lichess shows it in the summary of a computer analysis.

### Is lower centipawn loss better?

Yes. A lower average means your moves stayed closer to the engine's choices. In my own games, my median average centipawn loss was 47 in wins and 80 in losses.

### What is a good average centipawn loss?

Lichess doesn't publish a target. Compare your number only with your own games analysed the same way, because the engine, depth and caps change it.

### Why does Lichess use accuracy instead of centipawns?

Lichess's accuracy page says centipawns are great for developing engines but not so much for human comprehension. Its accuracy is based on lost winning chances, which depend on how close the game was.

### Is centipawn loss the same as accuracy?

No. Centipawn loss counts evaluation lost in pawns; Lichess accuracy counts winning chances lost. A four-pawn drop in a won position can cost less accuracy than a one-and-a-half-pawn drop in an equal one.

## Related

- [How to read Stockfish analysis (and why it says blunder)](https://thechesscoach.netlify.app/guides/how-to-read-stockfish-analysis.md)
- [Lichess analysis board explained: every panel and button](https://thechesscoach.netlify.app/guides/lichess-analysis-board-explained.md)
- [What is a good accuracy in chess? Real numbers (2026)](https://thechesscoach.netlify.app/answers/what-is-a-good-accuracy-in-chess.md)
- [Inaccuracy vs mistake vs blunder in chess: the real cutoffs](https://thechesscoach.netlify.app/answers/inaccuracy-vs-mistake-vs-blunder.md)
- [Chess MCP server for Claude and ChatGPT: Chess Coach](https://thechesscoach.netlify.app/connect.md)