Playing11 min read

Smith-Morra Gambit: The Better the Player, the More They Decline

The Smith-Morra pages we opened name the setups and the Siberian Trap. They do not say what Black actually plays at each rating. Chessdrive's Maia3 and Stockfish data covers two positions at seven rating bands. Accepting with 3...dxc3 is the engine's best move (-0.37) and the likeliest human move at every band, but its share peaks at 73.1 percent at 1600 and falls to 52.3 percent at 2200 as declines rise. After 4.Nxc3, 4...Nc6 is best everywhere, and the low-band mistake is 4...e5. The engine also says the gambit is slightly dubious, while the traps keep it practical at club level.

The starting position

The Smith-Morra Gambit is 1.e4 c5 2.d4 cxd4 3.c3, an anti-Sicilian gambit. White offers a pawn for fast development and attacking chances. After 3...dxc3 4.Nxc3, White has a developed knight and a central pawn, and Black keeps the extra pawn and a central pawn majority. ECO code: B21 (B20 in the ECO volumes), according to Wikipedia, which says the Encyclopaedia of Chess Openings uses B20 and Chess Informant uses B21. If you are new to the Sicilian itself, start with the Sicilian Defense page.

The pages we opened name the setups and the famous trap. None shows what Black plays at move 3 or move 4 by rating, or what each way of declining costs. We ran the maia3-5m checkpoint (see how Maia chess actually works) on two Smith-Morra positions at seven rating bands from 1000 to 2200, and graded every legal move with Stockfish 18 at depth 20. Percentages in the tables are the model's predicted move probabilities for a player at that rating. They are not game database frequencies and they are not win rates. We have no game results for this opening.

Every evaluation is from White's side, so +0.37 means White is better and -0.37 means Black is better. The two dataset positions are the one after 3.c3 and the one after 3...dxc3 4.Nxc3, both with Black to move. Our line checks are separate local runs of Stockfish, ten seconds per line, called local checks below. Table evaluations come from the dataset; line evaluations come from the local checks.

Where the Smith-Morra comes from

The name honours Pierre Morra (1900 to 1969) of France and Ken Smith (1930 to 1999) of the Dallas Chess Club. In Europe it is usually called the Morra Gambit, and older names such as the Tartakower Gambit and the Matulovic Gambit have fallen out of use. Smith did not have an easy time with it himself: Wikipedia records that he tried it three times at San Antonio 1972 and lost all three, against Byrne, Evans and Mecking.

Wikipedia says the gambit is popular at club level and played occasionally by masters. Marc Esserman wrote the book Mayhem in the Morra! (2012) and, per Wikipedia, beat Loek van Wely in 2011 and Gukesh D in 2017 with the Nd5 sacrifice. Hikaru Nakamura played it against Caruana at the 2023 Chessable Masters, winning one rapid game and losing the other. The Gukesh year rests on Wikipedia alone, and we found no move lists for these games.

Move 3: should Black take the pawn?

After 1.e4 c5 2.d4 cxd4 3.c3, Stockfish ranks 3...dxc3 first at -0.37. Then come 3...Nf6 at +0.05, 3...e5 at +0.08 and 3...d5 at +0.19. Taking the pawn is the engine's choice, and it is also the model's favourite at every band.

Rating1000120014001600180020002200
3...dxc3 (-0.37)66.8%70.6%72.4%73.1%71.7%64.3%52.3%
3...Nc6 (+0.37)19.7%17.3%14.4%11.2%7.8%5.2%3.4%
3...Nf6 (+0.05)2.4%2.5%3.0%3.7%4.8%7.9%14.9%
3...d3 (+0.65)1.2%1.4%2.1%3.3%5.8%10.0%13.2%
3...d5 (+0.19)not top fivenot top fivenot top fivenot top five2.7%5.6%9.4%
3...e5 (+0.08)6.0%4.0%2.9%2.4%not top fivenot top fivenot top five

Here is the pattern that makes this opening unusual. The share for 3...dxc3 rises from 66.8 percent at 1000 to a peak of 73.1 percent at 1600. Then it falls: 71.7 percent at 1800, 64.3 percent at 2000 and 52.3 percent at 2200. In the same bands the declines grow. 3...Nf6 goes from 2.4 to 14.9 percent, 3...d3 from 1.2 to 13.2 percent, and 3...d5 enters the top five at 1800 and reaches 9.4 percent at 2200. So from 1600 up, the stronger the player, the more often the model turns down the objectively best move.

The sources only say that many Sicilian players decline. Our data shows that the named declines are a high-rating habit. Below 1600, 3...Nf6, 3...d3 and 3...d5 together stay in single digits. Low-rated Black players still skip the pawn about a third of the time at 1000, but they do it with the quiet 3...Nc6, which no source lists as a named decline. It falls from 19.7 percent at 1000 to 3.4 percent at 2200, and at +0.37 it concedes about 0.74 against taking the pawn. So the low bands avoid the gambit with a passive move, and the high bands avoid it with the book declines.

Wikipedia gives a reason: if Black keeps the pawn, the endgame chances are good but hard to realise, so many Sicilian players decline. Our data cannot test that. What the engine does say is that every decline we checked is worse for Black than accepting, by roughly 0.4 to 1.1 of a pawn.

The declines, checked

Wikipedia and chess.com give the declines slightly different names, so we use both where they differ. All numbers below are local checks, White's side.

  • 3...Nf6 4.e5 Nd5 (Alapin transposition). +0.11. This is the most popular decline at 2200 and the one closest to equality for Black. If you play the Alapin as White, the Sicilian Defense page covers the family.
  • 3...d5 4.exd5 Qxd5 (Scandinavian structure). +0.21, with 5.cxd4 the engine move. For the pawn structure after the queen recapture, see the Scandinavian Defense page.
  • 3...d3 (Advance, Dubois Variation, called the Push Variation on chess.com). 4.c4 gives +0.43, with 4...e5 the best reply. The dataset grades the pawn push itself at +0.65. It grows from 1.2 to 13.2 percent.
  • 3...e5 (Center decline). +0.10 after 4.Nf3 Nc6 5.cxd4 exd4 6.Nxd4. The dataset gives +0.08.
  • 3...Qa5 (Wing) and 3...g6. Both are named in the sources. The dataset grades 3...Qa5 at +0.31 and 3...g6 at +0.73, and neither reaches the top five at any band.

Move 4: 4...Nc6 is best, 4...e5 is the low-band error

After 3...dxc3 4.Nxc3, Stockfish ranks 4...Nc6 first at -0.30, then 4...e6 at -0.27, 4...d6 at -0.23 and 4...a6 at -0.19. Black has four good moves, all a few hundredths apart. The human distribution is different.

Rating1000120014001600180020002200
4...Nc6 (-0.30)46.9%50.7%51.5%50.1%47.4%45.8%48.4%
4...e6 (-0.27)8.7%10.6%12.9%16.1%20.8%26.9%30.8%
4...d6 (-0.23)7.8%8.3%9.1%10.2%11.5%12.0%11.1%
4...e5 (+0.63)12.9%10.7%9.0%7.3%5.3%not top fivenot top five
4...g6 (+0.00)not top five6.7%7.3%7.6%7.6%6.6%4.5%
4...Nf6 (+0.77)10.2%not top fivenot top fivenot top fivenot top fivenot top fivenot top five

4...Nc6 is the likeliest move at every band, between 45.8 and 51.5 percent. It also begins most of the named setups in the sources: the Scheveningen, the Siberian, the Morphy and the Pin. So the mainstream move is also the engine move, and it takes nothing special to find.

The surprise is below it. At 1000 the second most likely move is 4...e5, at 12.9 percent, and no source we opened mentions it. The dataset rates it +0.63, which flips the evaluation from a third of a pawn for Black to a clear edge for White, a swing of about 0.9. A deeper local check of the same position gives +0.53 with 5.Bc4 Nf6 as the engine line, so the exact figure moves but the verdict does not: 4...e5 is a mistake. It fades as ratings rise, to 5.3 percent at 1800, and 4...Nf6 at +0.77 is a similar low-band slip at 10.2 percent at 1000.

The other trend runs the opposite way. 4...e6 climbs from 8.7 percent at 1000 to 30.8 percent at 2200, and it costs almost nothing against the best move (-0.27 against -0.30). The Chicago Defense 4...d6 holds a steady 7.8 to 12.0 percent. At no band does the model's favourite differ from the engine's, so for Black the real advice is simple: play 4...Nc6 and avoid 4...e5.

The named setups, checked

Wikipedia calls 4...Nc6 5.Nf3 d6 6.Bc4 e6 the Scheveningen Formation, and chess.com agrees. Wikipedia labels 5...e6 6.Bc4 Nf6 the Siberian Variation, while chess.com calls 5...e6 6.Bc4 a6 the Paulsen Formation. Do not treat one naming scheme as universal. We checked the lines that matter.

  • Scheveningen main line. 7.O-O Nf6 8.Qe2 Be7 9.Rd1 e5 is +0.12 for White in a local check, with 10.Bb5 the engine move. Wikipedia gives 10.h3 or 10.Be3 as the classical continuations. Chess.com lists "Black can gain a slight edge with theoretical knowledge" as a con. The engine says White is a hair better here, so in this line the Black edge shown at move 4 is gone by move 9.
  • Pin Defense 6...Bb4. After 4...Nc6 5.Nf3 e6 6.Bc4 Bb4 we get -0.26, with 7.O-O Nge7 the engine line, so Black is comfortable.
  • Chicago Defense. 4...d6 5.Nf3 e6 6.Bc4 Nf6 7.O-O a6 is -0.05, with 8.e5 dxe5 9.Qxd8+ Kxd8 the engine idea. That is dead level.
  • US Chess novelty 6...Nge7 and 7...h6. A US Chess article presents it as putting the gambit's soundness to the test. We could not read it, so the move order is unverified. In a local check, 6...Nge7 7.O-O h6 is +0.13 for White with 8.h4 the engine move, so the engine sees nothing close to a refutation.

The Siberian Trap

The moves, per Wikipedia and its trap page: 1.e4 c5 2.d4 cxd4 3.c3 dxc3 4.Nxc3 Nc6 5.Nf3 e6 6.Bc4 Nf6 7.O-O Qc7 8.Qe2 Ng4. Black threatens to jump to d4. If White plays the natural 9.h3??, then 9...Nd4! attacks the queen on e2 and the knight on f3. Per Wikipedia, 10.Nxd4 runs into 10...Qh2#, so White has to give up the queen.

What our engine confirms is the cost. After 9.h3 Nd4, the local check gives -2.96, with 10.hxg4 Nxe2+ 11.Bxe2 the best line. White takes the knight, loses the queen for a knight, and plays on a long way behind. The trap sits at move 9, outside our dataset, so it has no rating band.

The alternatives at move 9 are where the sources and the engine part ways. Wikipedia's trap page calls 9.Rd1 Bc5 another try that does not work for White. The local check disagrees: +0.19 for White, with 10.Nb5 Qb6 11.Bf4 the engine line. The recommended 9.Nb5 Qb8 from MCO-14 is 0.00 in our check, which matches the "roughly equal" the sources claim. So both 9.Rd1 and 9.Nb5 are fine for White, and the only error is 9.h3.

Esserman against Van Wely

The famous game is Esserman against Van Wely at the U.S. Open in Orlando, 2011. Per Jeremy Silman's review of Mayhem in the Morra!, the moves went 1.e4 c5 2.d4 cxd4 3.c3 dxc3 4.Nxc3 Nc6 5.Nf3 e6 6.Bc4 a6 7.O-O Nge7 8.Bg5 f6 9.Be3 Ng6 10.Bb3 b5 11.Nd5. The review printed 4.Nc3, which we read as 4.Nxc3, and we replayed the line to confirm it is legal. Silman marks the sacrifice 11.Nd5 as brilliant, and says it is especially strong with a pawn on e6, because ...exd5 opens the e-file.

The engine is more measured. A local check puts the position after 11.Nd5 at +0.38 for White, with 11...exd5 12.Qxd5 Qe7 the best reply. That is a real edge, not a winning position.

Is it sound?

Opinions split. A chess.com blog says it may not be entirely sound but is dangerous against the unprepared, and a US Chess article presents a Black setup as a test of its soundness. Wikipedia does not call it refuted, and says that to survive a prepared White, Black must first avoid the traps and then withstand long-term pressure.

The dataset gives a consistent reading. Black is about a third of a pawn better after the best replies, -0.37 at move 3 and -0.30 at move 4, and the engine's side of the debate is that the gambit is objectively slightly dubious. That is not a refutation, and it is smaller than the swings the engine finds when either side errs: 4...e5 is +0.63 for White, 9.h3 loses White's queen, and every decline we checked gives back between 0.4 and 1.1 of a pawn. A gambit that is a third of a pawn short but wins games through traps and unfamiliar structures is a sensible choice at club level, which is exactly where Wikipedia says it lives. Chess.com lists it as suiting tactical players who like sharp positions and sidestepping Sicilian theory, and notes Black can steer into quieter lines. How its acceptance rate compares with five other gambits at each rating is in our comparison of chess gambits.

If you like the same idea in a different opening, the Danish Gambit page covers a gambit with the same trade of a pawn for development, and the Evans Gambit page is the classical 1.e4 e5 version.

Who should play it, by rating

Our data does not include game results by rating for the Smith-Morra, so this is what the model and the sources suggest, not a result table.

  • Around 1000 to 1400, as White. Black usually takes the pawn (3...dxc3 at 66.8 to 72.4 percent) and usually finds 4...Nc6 (46.9 to 51.5 percent), but this is also where Black is most likely to err, with 4...e5 at 12.9 percent at 1000 and 3...Nc6 at 19.7 percent. Learn the Siberian Trap and 4...e5 first.
  • Around 1600 to 1800, as White. Black accepts most reliably here, at 73.1 and 71.7 percent. Expect 4...Nc6 and 4...e6.
  • 2000 and above. The declines arrive: 3...Nf6, 3...d3 and 3...d5 together are far more common than at 1400. This is where a repertoire matters, and the advice in why you should stop memorizing openings and learn plans applies to the structures that follow.
  • As Black, at any rating. Take the pawn, play 4...Nc6, and know the Siberian Trap and the 9.Rd1 and 9.Nb5 escapes.

For how all of this fits with other openings by strength, see chess openings by rating.

Where other resources are better

This page is narrow on purpose. The dataset holds two positions, and Maia3 5m is a small model of human choice, so its percentages are predictions. Our line checks are quick local runs, not a substitute for a repertoire book. Wikipedia is the best reference for the named setups and the history. We have no game results by rating, and we did not open the Lichess explorer, which is the place for real game counts. Several pages in the search results, including the TheChessWorld guide and the US Chess article, were unreadable. None of the pages we did read compares the engine's best move with the likeliest human move at each rating, which is the gap this one fills.

What to take from it

After 3.c3, taking the pawn with 3...dxc3 is the engine's best move at -0.37 and the model's favourite at every band, but its share peaks at 73.1 percent at 1600 and falls to 52.3 percent at 2200 as 3...Nf6, 3...d3 and 3...d5 rise. After 4.Nxc3, 4...Nc6 is best and likeliest everywhere, the low-band mistake is 4...e5, and 4...e6 grows from 8.7 to 30.8 percent. The engine says the gambit is slightly dubious, with Black about a third of a pawn better, and the traps are why it is still worth playing. Keep every percentage for what it is: the prediction of a model, not a count of real games.

Frequently asked

Is the Smith-Morra Gambit good?
It is playable but slightly dubious by engine standards. In our Stockfish data the best reply to 3.c3 is 3...dxc3 at -0.37 from White's side, and after 4.Nxc3 the best move is 4...Nc6 at -0.30, so Black is about a third of a pawn better. Wikipedia says it is popular at club level and played occasionally by masters, and does not call it refuted. White gets quick development and traps, and pays a pawn for them.
Should Black accept the Smith-Morra Gambit?
By the engine, yes. 3...dxc3 is the top move at -0.37, and every decline we checked is worse for Black: 3...Nf6 is +0.05, 3...e5 is +0.08, 3...d5 is +0.19 and 3...d3 is +0.65. The model also predicts accepting as the most likely human move at every band from 1000 to 2200. Wikipedia notes that many Sicilian players decline anyway, and our data shows that habit growing at 2000 and above.
What is the Siberian Trap in the Smith-Morra?
It is a Black trap on move 9. After 4...Nc6 5.Nf3 e6 6.Bc4 Nf6 7.O-O Qc7 8.Qe2 Ng4, the move 9.h3?? loses to 9...Nd4. Our local Stockfish check gives -2.96 for White, with 10.hxg4 Nxe2+ 11.Bxe2 leaving White a queen down for a knight. Per Wikipedia, 10.Nxd4 runs into 10...Qh2#.
Is the Smith-Morra Gambit the same as the Morra Gambit?
Yes. Wikipedia says it is named after Pierre Morra of France and Ken Smith of the Dallas Chess Club, and in Europe it is usually just called the Morra Gambit. Older names such as the Tartakower Gambit and the Matulovic Gambit have fallen out of use.
What is the ECO code for the Smith-Morra Gambit?
B21 (B20 in the ECO volumes). Wikipedia notes that the Encyclopaedia of Chess Openings uses B20 for this move order while Chess Informant uses B21, so both codes appear in the wild.
What is the most common mistake for Black?
After 3...dxc3 4.Nxc3, the model's low-band favourite error is 4...e5, predicted at 12.9 percent at 1000 and 10.7 percent at 1200. The dataset rates it +0.63 for White against -0.30 for 4...Nc6. A deeper local Stockfish check gives +0.53, so it hands White a clear edge either way.

Want to face the Smith-Morra against a model that answers like your rating?

Chessdrive runs the real maia3-5m checkpoint on-device as your opponent at any rating you set from 600 to 2600, so you can play the Smith-Morra from either side against a model that picks its moves the way your band does, then see where the game turned with a separate Stockfish review. Available now on iOS.

Keep reading

All guides

Updated