Queen's Gambit Accepted: Can Black Keep the Pawn?

Every Black player who enters the Queen's Gambit Accepted with 2...dxc4 hits the same wall on move three. The pawn is there, it is extra, and ...b5 looks like a way to keep it. A chess.com poster named chasm1995 puts the question plainly: how exactly does one hold on to it? The honest answer is that it depends on which third move White picked, and most pages on the opening never separate the cases. We did. This article uses our maia3-5m human-move data, the same method as our chess openings by rating dataset, plus new Stockfish 18 checks at depth 22 on the pawn-keeping lines. For the full per-rating tables on Black's best replies after 3.Nf3, see our Queen's Gambit data page. The short version is that the engine's top move, 3...a6, stays rare below 2200 while 3...Nf6 and 4...e6 cost almost nothing. This page covers the other half of the story: what happens when Black tries to keep the pawn.

What the Queen's Gambit Accepted is

The Queen's Gambit Accepted starts 1.d4 d5 2.c4 dxc4 and carries the ECO codes D20 to D29. Wikipedia traces the modern ideas to the 1886 World Championship match between Steinitz and Zukertort, where Steinitz planned to give the pawn back but leave White with an isolated pawn on d4. The name is slightly misleading, because it is not a real sacrifice. Wikipedia says Black cannot retain the c4 pawn without disadvantage, and chess.com says Black typically does not retain the extra material.

That shapes how Black should think about it. The pawn is on loan. The real return is time: while White spends moves winning it back, Black develops, and chess.com lists piece mobility, an open long diagonal for the c8 bishop and counterplay against d4 as Black's gains. On the other side of the ledger it lists White's two to one pawn centre advantage, White's control of the c-file, and the fact that the bishop on f1 can recapture straight on c4.

The results are modest. In chess.com's database, as of September 2026, 2...dxc4 appears in 33,897 games after 2.c4, against 118,063 for the Slav and 115,120 for the Queen's Gambit Declined. White scores 38.1 percent wins, 35.7 draws and 26.3 losses, a lower White win rate than against the Slav (40.0) or the Declined (42.4). House of Staunton goes further and says the Accepted has the lowest percentage of White victories among Black's main second moves. Modern-chess.com, read in September 2026, quotes White at 58 percent, but it is measuring a different pool, so do not set the two numbers side by side.

White's third move decides everything

Wikipedia lists five third moves for White. They do not carry equal weight.

• 3.Nf3. The most popular main line. It prevents ...e5, and White follows with 4.e3 and Bxc4. Chess.com's sample of 33,844 games, as of September 2026, puts it at roughly 75 percent.
• 3.e3. Modest, prepares Bxc4 and invites 3...e5 according to Wikipedia. In the same sample it is 21.3 percent, the real second choice by a wide margin.
• 3.e4. The Central Variation, an old move that Wikipedia says became popular again in the 1990s. It is 3.0 percent of that sample.
• 3.Nc3. Wikipedia calls it misguided because it does not control d4 and e5 adequately. It shows up in 0.4 percent of games.
• 3.Qa4+. Rarely played, and Black gains time harassing the queen.

The chess.com figures are database shares from September 2026, not our Maia3 predictions. Our dataset begins one move later, after 3.Nf3, so we have no rating-banded split between 3.Nf3 and 3.e3. That matters for the next section, because the answer to "can Black keep the pawn" changes with the third move more than with the rating of either player.

How often do humans try ...b5?

We ran Maia3 on two positions: Black's third move after 3.Nf3, and Black's fourth move after 3...Nf6 4.e3. At each of seven rating bands from 1000 to 2200 the model gives the probability that a player of that rating picks each legal move, and Stockfish 18 at depth 20 scores every one. Percentages are model predictions, not counts of real games. For the method, see how Maia chess actually works and what Stockfish actually is.

The pawn-keeping move ...b5 is not a fringe idea in either position.

The evals in brackets are White's point of view. Both moves peak at 1400, with 13.7 and 13.9 percent, and at 1200 and 1600 they are the fourth or third most likely move. In the first position the engine's best move is 3...a6 at +0.29, so 3...b5 at +0.48 costs Black about 19 centipawns. In the second, 4...e6 and 4...a6 both sit at +0.26 and 4...b5 at +0.43, a cost of 17. Those are small gaps, which is why the move is so tempting: on a static engine score ...b5 does not look bad.

The decline afterwards is gradual. By 2000 3...b5 has fallen out of Black's top five in the first position, and in the second 4...b5 slides to 7.8 percent at 2000 and 5.5 at 2200. Stronger players do not stop considering the move overnight, but they play it less often.

A caution on the percentages: in the table the 3...b5 entry only answers one question, what Black plays after 3.Nf3. The moves a player might make against 3.e3 are not in our data at all.

When ...b5 fails and when it holds

This is where the pages we looked at disagree. Wikipedia calls 3...b5 fairly risky and rarely seen. A chess.com forum poster, TwoMove, says holding the pawn with ...b5 is risky for Black but playable, and that stronger players do it, though after e3 ...b5 really is bad. Another poster, Gambitknight, says the key is not to stubbornly hold onto that pawn. All three can be right, because the move is different in different setups. We checked these positions with Stockfish 18 at depth 22.

Against 3.e3 the problem is concrete. After 3.e3 b5 4.a4 c6 5.axb5 cxb5, White plays 6.Qf3. The long diagonal from f3 to a8 is open, because Black's b7 and c6 pawns have both left it, and Black has no good way to cover the rook on a8. The engine's line runs 6...Qc7 7.Qxa8, and the position scores +3.22 for White. Even before the trap, the engine scores 3.e3 b5 at +0.86 for White, with 4.a4 as its choice, which is already more than the +0.3 to +0.4 of the main lines. Wikipedia describes the same trap and says the defence is fairly risky and rarely seen. The Chess.com forum offers the same advice in plainer terms. BBostrom writes that e3 is the better choice because the queen can still get out and it sets a nice little trap if Black defends the pawn.

Against 3.Nf3 the trick does not work, because the knight already stands on f3 and the queen cannot use that square. After 3.Nf3 b5 4.a4 c6 the same engine gives White only +0.48 with 5.axb5 cxb5 6.Nc3 Qb6 as the main line. That is about a fifth of a pawn worse than the best Black reply, and only a fraction more than the +0.3 you get from sensible development. So on the engine's scale, holding the pawn after 3.Nf3 is playable. In practice it is committal: White gets the c-file and a lead in development, and Black has to know the a4 break.

The takeaway is that "can Black keep the pawn" has no single answer. After 3.e3 it is a mistake that needs only a few accurate moves from White to punish. After 3.Nf3 it is a modest concession. After 3.e4 the question is almost moot, because the best reply is a central counter, not a pawn defence.

The 3.e4 central line

3.e4 is rare, 3.0 percent of chess.com's sample as of September 2026, but it is the line with the most direct theory and the one shown in the Harmon and Borgov finale of the Netflix series. Ivanchuk against Wolff in 1993 reached the same structure with 3...Nc6 4.Be3 Nf6 5.Nc3 e5 6.d5. White's aim, according to House of Staunton, is to build pawns on d4 and e4 and then regain the pawn with Bxc4.

House of Staunton names five Black replies: 3...e5 (the McDonnell Defense), 3...Nf6 (the Alekhine System), 3...Nc6 (the Modern Defense), 3...c5 (the Rubinstein Defense) and 3...b5 (the Greco Variation). Our Stockfish 18 check at depth 22 prefers the counter-strike 3...e5, scoring the position at +0.38 for White, and the line it likes runs 4.Nf3 Bb4+ 5.Nc3 exd4. That is a long way from a clear edge. Wikipedia's summary of the theory says White's best reply to 3...e5 is 4.Nf3 and not 4.d5. It also says the Alekhine System with 3...Nf6 provokes 4.e5 Nd5 5.Bxc4 Nb6.

House of Staunton adds a warning for White: 4.dxc5 is not a good idea, because it allows Black to trade queens with 4...Qxd1+. The sensible idea is simpler, and it is the same one that runs through every White setup. Get the pieces out and take the pawn back with the bishop.

How White wins the pawn back

After 3.Nf3 Nf6 4.e3 e6 5.Bxc4 the pawn is back. Wikipedia's classical main line goes 5...c5 6.O-O a6, where Black challenges d4 while preparing ...b5 to harass the bishop on c4. Notice that ...b5 shows up here too, but as a plan with ...a6 first, not as a grab on move three or four. In our Maia3 data the earlier ...a6, one move before, enters the top five only at 2000 (4.5 percent) and 2200 (7.5).

For White, the practical plan after regaining the pawn is easy to state. Recapture with the bishop, castle, and use the centre. Irontiger's advice on the chess.com forum is to play e3 with the bishop on f1 attacking c4, keeping the a8 to h1 diagonal open so that a later Qf3 is available. That is the structure behind the trap above, and it is why 3.e3 is more dangerous than it looks.

For Black, the useful lesson is Steinitz's: give the pawn back on your own terms and aim to leave White with an isolated pawn on d4. We have no sourced guide for that middlegame beyond the idea, so we will not invent one. If you want to see how a 1.d4 player who avoids all of this sets up, see our London System guide.

What Black should play at each rating

Our data gives a clearer instruction than most forum threads. At 1000 the likeliest third move after 3.Nf3 is 3...Nc6 at 26.6 percent, with 3...Nf6 at 21.3. 3...Nc6 costs 43 centipawns against the engine's best, the most expensive of the three likeliest moves at that band, and our full tables on the Queen's Gambit page show it fading as rating rises. From 1200 up, 3...Nf6 is the likeliest move and costs only 5 centipawns against the engine's best, which is why it is the sensible default for a club player.

If you are around 1400 and tempted by ...b5, the data says you are in good company, since it is 13.7 percent of Maia3's prediction at that band. It also says to check White's setup first. If White has played e3 early, hold back: the Qf3 trick needs only one slip. If White has played Nf3, ...b5 is a sidelined but playable path. At 2000 and above, 3...b5 drops out of Maia3's top five, so stronger opponents rarely test it.

The wisest summary comes from NM ThunderClap on Lichess: you can accept the gambit and not hold onto the pawn, just develop your pieces afterwards. The main lines sit near +0.3 for White, and the pawn-keeping lines cost more.

What White should play at each rating

If you play White, the third move you pick is also a choice about which opponent you want. 3.Nf3 is the mainstream and the safest: it controls e5, leaves ...b5 as a mild concession, and matches three quarters of the games on chess.com as of September 2026. 3.e3 is the trap-friendly choice, popular in about a fifth of games, and BBostrom's forum comment says why. Its downside is that it allows 3...e5 according to Wikipedia, so White needs a plan for that. 3.e4 has the best story and the least traffic, and gives only +0.38 against best play, so use it if you enjoy open central positions and know the theory.

For players who dislike memorization, White's main task is simple: take the pawn back and develop.

Where other pages are better

Our page is deliberately narrow, and other sources do things we cannot. Wikipedia has the fuller variation tree and the Steinitz history. Chess.com's opening explorer has the largest game counts, and its paid lesson covers 3.e3, 3.e4 and 3.Nf3 in video form. As of September 2026, modern-chess.com names Sergey Rublevsky as the leading top-level practitioner, with 182 games, and lists engine evaluations for sub-lines. House of Staunton goes deeper into the names and ideas of the 3.e4 variations, though it is written for White only and urges memorising theory. The chess.com forum threads have the real questions from real players, including the honest disagreement about ...b5.

What none of them shows is how often humans at each rating try the pawn-keeping move, or what the engine says about it in each setup. What we do not measure: results, anything after Black's fourth move, and how often White chooses 3.e3 or 3.e4 at a given rating. Our data cannot tell you whether your opponent will play 3.e3 either. Use our chess analysis board to see what the engine thinks after a game of your own, and the openings by rating overview for the other lines we measured.

What to take from it

The Queen's Gambit Accepted is a sound opening that asks Black to give the pawn back with good grace. The temptation is ...b5, and our data says between 9 and 14 percent of Maia3's prediction at ratings from 1000 to 1800 goes to it, in both the early and late position. The engine says the danger depends on White: +3.22 after the 3.e3 trap line, +0.86 after 3.e3 b5 alone, and +0.48 after 3.Nf3 b5 4.a4 c6. If you are Black, develop first and keep ...b5 for the moment White has shown which setup it is. If you are White, 3.e3 is a good way to set a trap, and 3.Nf3 is the way to play the position you already know.

Maia3 is a model of human choice built from the maia3-5m checkpoint, not a count of real games, and every percentage above should be read that way.