Playing10 min read

Chess960: Rules, Castling, and What 338,809 Games Show

Chess960 shuffles the back rank into one of 960 setups. We read 338,809 rated Lichess games from October 2026 and found that the setup you draw does not measurably change White's result at club level, and that blunder rates look like standard blitz.

The starting position

Which of the 960 starting positions you draw makes no measurable difference to White's result at club level. In 338,809 rated Chess960 games from Lichess's open database for October 2026, a position's White score in one half of its games had a correlation of about -0.06 with its score in the other half, which is what you get when there is nothing to predict. Below: the setup rules, castling, the data, the history, and whether it helps your standard chess.

The short answer

Chess960 is chess with the back rank shuffled into one of 960 legal setups. The setup does not decide the game: White scored 51.0% across all our games. Club players blunder at close to standard blitz rates once the first five moves are done. The one rule that needs explaining is castling, covered below.

What Chess960 is, and the setup rules

White's pieces are shuffled along the first rank under two constraints, and Black copies the arrangement mirrored. The pawns stay where they are.

  • The bishops stand on opposite coloured squares. One on a light square, one on a dark square.
  • The king stands between the two rooks. This is what keeps castling possible for both sides from the start.
  • Black mirrors White. Each of Black's pieces starts on the same file as its White counterpart.

Count the ways: 4 choices for the dark squared bishop, 4 for the light squared one, 6 for the queen, 10 for the knights, and the rooks and king take the last three squares in one fixed order. 4 x 4 x 6 x 10 = 960. The standard start, RNBQKBNR, is one of them, number 518 in the Scharnagl numbering that the format uses (SP 518). Setups are normally picked by software or by dice.

Fischer's stated aim, as Wikipedia summarises it, was chess "not degenerated down to memorization and prearrangement." He stopped short of full randomness because with pieces placed anywhere, one side would almost certainly start with an advantage. For the standard layout, see our guide to how to set up a chess board.

How to castle in Chess960

The rule is about where the pieces end up, not where they start.

  • Castling toward the a file (queenside, written 0-0-0). The king ends on the c file and the rook on the d file.
  • Castling toward the h file (kingside, written 0-0). The king ends on the g file and the rook on the f file.
  • The path must be clear. Every square between the king's start and end squares, and between the rook's start and end squares, including the end squares, must be empty except for the king and the castling rook.
  • The king's path must be safe. The king may not castle out of check, through an attacked square or into check. The rook may pass through or land on attacked squares.
  • Neither piece may have moved. Same as in standard chess.

Two examples. With the king on b1 and the rook on a1, castling toward the a file puts the king on c1 and the rook on d1, and both squares must be empty. With the king on e1 and the rook already on f1, castling toward the h file means the king alone steps to g1 while the rook stays on f1. The rule only fixes where the pieces finish, so a piece that is already there simply stays.

How to enter the move on Lichess and Chess.com

On Lichess, the variant page says the only way to castle is to move the king onto the rook. On Chess.com, the advice we saw in search results is to drag your king onto the rook on the side where you want to castle, and community posts report trouble when the two stand one square apart. Over the board, FIDE's recommended procedure, per Wikipedia's summary, is to announce castling, move the king next to its final square, move the rook to its final square, then move the king to its final square. FIDE's own guideline page returned an error when we checked.

What 338,809 games say about castling

We streamed Lichess's open database file of rated Chess960 games for October 2026 (CC0 license): 367,964 games, of which 338,809 remained after dropping BOT accounts and games where the players' ratings differed by more than 400. Bands use the average of the two ratings. All 960 positions appear, with 283 to 436 games each and a median of 352.

White's castling by band, with the average ply at which a castle happened.

Average ratingGamesWhite castledNeither side castledQueenside share of castlesAverage castle ply
Under 100052525.3%63.6%54.0%19.9
1000 to 139915,67935.6%49.2%46.7%20.9
1400 to 1799163,82063.3%21.1%41.8%20.5
1800 to 2199148,66473.0%12.7%39.7%20.1
2200 and above10,12177.7%10.0%42.3%19.2

Source: Lichess open database, October 2026 Chess960 file, CC0. The under 1000 row rests on 525 games, so treat it as a rough guide.

What the table says

  • Castling is a rating habit. White castles in about a quarter of games under 1000 and in 77.7% at 2200 and above. Short games explain part of it: under 1000 they end before many players castle.
  • Players castle around move 10. The average castle ply sits between 19.2 and 20.9 at every band, which is move 10 or so.
  • Queenside is common. Between 39.7% and 54.0% of castles go toward the a file, depending on band. That is why the castling rule matters: you often read where your king and rooks start before choosing a side.

Does your starting position matter? Not in our data

The question every Chess960 page eventually asks is which setups favor White. The published answer comes from engines. A Stockfish 9 evaluation of all 960 setups by Sesse, summarised on Wikipedia, puts White's advantage between 0.00 and 0.57, averaging 0.18, with the standard SP 518 at 0.22 and 27 setups equal. Engines see differences. Do human games? We split each position's games into two halves and asked whether White's score (a draw counts half) in one half predicts the other.

MeasureValue
Positions, all with at least 283 games960
Split half correlation of White score-0.0559
Observed spread (SD) of White score across positions0.0254
Spread expected from sampling noise alone0.0267
White score at the 10th and 90th percentile of positions47.8% and 54.3%

The observed spread across positions, 0.0254, is no larger than the 0.0267 that coin flips would produce at these sample sizes. The correlation between halves is slightly below zero. Put plainly, a position that scored well in one half did no better than any other in the other, so the 47.8% to 54.3% range is chance. The standard position illustrates this, since White scored 48.8% from it in 333 games, below even.

That is why we do not list the best and worst setups for White. A list built from a few hundred human games per position, ours included, would rank noise. The noise estimate is slightly conservative, since draws reduce real variance a little, but with a correlation near zero the conclusion holds.

White still scores about 51%

Across all 338,809 games White scored 51.0%, with draws counting as half, and 3.9% of games were drawn. By speed, White scored 51.0% in both the 192,703 blitz games and the 112,376 bullet games. Among the bands with enough games it ran from 50.6% at 1000 to 1399 to 51.4% at 2200 and above. Draws are rare: lowest at 1400 to 1799 (3.5%) and highest at 2200 and above (5.5%). Most of the data is fast chess (32,149 rapid and 1,581 classical games), so read it as a description of online blitz and bullet 960.

Do shuffled openings make you blunder more?

We used the 49,801 games with Lichess computer evaluations and counted blunders, meaning a move that drops the mover's win probability by 30 points or more. The standard comparison is our Lichess blitz data from the blitz chess guide. Both count moves from ply 10 onward by the mover's own rating, so the first five moves are left out. The caveat: the standard figures come from a different dump, part of one day of standard games against a full month of Chess960.

Mover's ratingChess960 blitz, blunders per 100 movesStandard blitz, blunders per 100 moves
1000 to 13994.834.42
1400 to 17993.583.42
1800 to 21992.622.63
2200 and above1.741.87

We skipped the under 1000 band because its Chess960 sample is small. At 1800 to 2199 the rates match, and at 2200 and above the 960 rate is lower. The biggest gap is 0.4 blunders per 100 moves at 1000 to 1399. So shuffled openings do not make club players blunder much more past move 5. The first five moves are excluded from both columns, and that is where unfamiliar positions would bite hardest.

Keep the reading narrow: the players differ, and evaluated games are a self selected subset because someone asked for analysis. For the same measure across speeds, see our bullet chess data.

What people open with in Chess960

The first move is where the shuffle shows up. e4 and d4 still lead, but not by much, and the fianchetto moves g3 and b3 are far more common than in standard chess. For comparison, in our standard Lichess blitz data at 1400 to 1799, e4 was 61.5% of first moves, d4 25.5% and g3 1.3%. Each figure counts that move in whichever setup the game started from.

Average ratingGamese4d4g3b3
1000 to 139915,39826.0%13.7%7.7%6.6%
1400 to 1799162,47525.0%15.0%11.5%9.7%
1800 to 2199148,08726.0%15.8%10.5%8.7%
2200 and above10,09320.2%13.2%12.2%9.8%

At 2200 and above g3 is nearly as common as d4. We did not read the games to find out why. Our guess is that with bishops often starting in corners, g3 and b3 open a long diagonal in many setups. For standard chess, see the best first move in chess and which openings players choose at each rating.

A short history, current as of November 2026

Shuffled setups are older than Fischer: an anonymous French book of 1792, randomised games in Mannheim in the 1840s, Georg Capellen's Freischach in 1915. Sources disagree on some dates.

  • Fischer, 1996. Bobby Fischer announced his version, Fischerandom Chess, in Buenos Aires in June 1996. ChessBase reports that a planned exhibition there was cancelled when sponsors withdrew, so we do not claim a game was played. His 1999 radio line: "I didn't invent Fischerandom chess to destroy chess."
  • Names. Fischerandom in 1996, Fischer Random Chess in 1998, Schach 960 in 2001 and Chess960 in 2002.
  • FIDE. FIDE added Chess960 to its rules around 2008 to 2009. It now sits in guidelines, not binding Laws.
  • World championships. Wesley So beat Magnus Carlsen 13.5 to 2.5 in the 2019 final. Hikaru Nakamura won the 2022 FIDE title over Ian Nepomniachtchi, 2 to 2 on games and then the armageddon tiebreak. No further FIDE title was contested until 2026.
  • Freestyle Chess. The pro circuit uses the name Freestyle Chess. Carlsen won the 2024 Freestyle Chess G.O.A.T. Challenge, beating Caruana in the final, and was overall champion of the 2025 Grand Slam Tour. FIDE and Freestyle Chess signed a cooperation agreement in January 2026, and at Weissenhaus on 13 to 15 February 2026 Carlsen beat Caruana to become the first official FIDE Freestyle Chess World Champion.

Carlsen's reason for backing the format, from the 2024 event, is this quote: "We're basically thrown into an unfamiliar middlegame from move one, where the price of a mistake is incredibly high." We did not verify that every Freestyle event used the same 960 setups.

Where to play Chess960

Lichess lists Chess960 as a variant, and Chess.com has one with its own rating, per a forum thread we found. We could not confirm either site's rating structure on an official page, so check the rating list. Chessdrive is not a Chess960 server. Our guide to playing chess online for free compares the main sites. Before your first move, take a minute to see where the bishops and queen stand, which pawns are undefended, and where your king and rooks start.

Does Chess960 help your standard chess?

We found no study showing it either way, so this is reasoning, not evidence. The honest case for it: with no memorised lines to lean on, you have to judge development, king safety and pawn structure yourself from move one. That is the skill our article on learning plans instead of memorizing openings argues matters more than lines at club level. The case against: ChessBase has published a critique titled "The problem with Chess960", and 960 will not teach you the Ruy Lopez.

Our data shows blunder rates near standard blitz and no setup effect on results. If you enjoy it, play it, and keep most of your improvement time on standard games and reviewing them.

Method and sources

Data: Lichess open database, rated Chess960 games, October 2026 file (CC0), the whole month, streamed with no local copy. We read 367,964 games and counted 338,809 after dropping BOT accounts and games with a rating gap over 400. Bands use the average of both ratings, blunder bands the mover's own. Speed is Lichess's estimated duration, base plus 40 times the increment. A blunder is a move that drops the mover's win probability by 30 points or more, on games with Lichess evaluations. Standard comparison: our Lichess standard games from part of 1 October 2026, same rule. Rules, history and the setup evaluation: Wikipedia's Chess960 article, ChessBase, the Lichess variant page, Chess.com support and news. Championship results are current as of 16 November 2026.

Frequently asked

What is Chess960?
Chess960 (Fischer Random Chess) is chess with a shuffled back rank. The bishops must stand on opposite coloured squares and the king must stand between the two rooks, Black mirrors White, and the pawns stay on the second rank. That gives 960 possible starting positions, and the usual setup is one of them.
How do you castle in Chess960?
The king and rook finish on the same squares as in standard chess: king on the c file and rook on the d file when castling on the a side, king on the g file and rook on the f file on the h side. Every square the two pieces cross or land on must be empty apart from them, and the king may not castle out of, through or into check. On Lichess and Chess.com you move or drag the king onto the rook.
Is Chess960 the same as Fischer Random?
Yes. Bobby Fischer announced it as Fischerandom Chess in 1996, and the name went through Fischer Random Chess and Schach 960 before settling on Chess960 in 2002. Pro events now say Freestyle Chess.
Is one Chess960 starting position better for White?
Engines say the differences are small: a published Stockfish 9 evaluation of all 960 setups puts White's edge between 0.00 and 0.57, averaging 0.18. In 338,809 human games on Lichess, a position's White score in one half of its games did not predict the other half, so at club level we could not measure any difference.
Does Chess960 make you better at standard chess?
We found no study either way. It removes opening memorisation, so it trains judging a position from first principles. Treat it as a complement to your normal games.
Who is the Chess960 world champion?
As of November 2026, Magnus Carlsen is the first official FIDE Freestyle Chess World Champion, after beating Fabiano Caruana at Weissenhaus on 13 to 15 February 2026. Wesley So won the 2019 event and Hikaru Nakamura the 2022 one.

Play 960 for fun, review your normal games for progress

Chessdrive is not a Chess960 server, so play it on Lichess or Chess.com. For the standard games you play every week, a Stockfish review in Chessdrive checks every move and explains the ones that mattered at your level. On iOS and Android, $5.99/month or $39.99/year on the US storefront with a free trial. Prices vary by region.

Keep reading

All guides

Updated