Queen's Gambit: Certain When Declined, Split When Accepted

On a chess.com thread titled "Queen Gambit Accepted or Declined," Optimissed makes the case for taking the pawn: "I think you learn more, quicker, by accepting. When you stop winning in the QGA as black but start drawing...then learn two lines of the QGD thoroughly." A few replies down, TheGreatOogieBoogie argues the opposite: "I'd recommend declining as black with 2...e6 until they learn the ideas behind the QGA." Neither one has access to the thing that would actually settle their own argument for their own games: what a real opponent, at their own rating, is likely to play back at them on each side of that choice.

That is the gap Chessdrive's own Maia3 data fills, and the answer is lopsided. We measured four positions the same way as every opening in our chess openings by rating dataset: the maia3-5m checkpoint, see how Maia chess actually works, predicting what a human at each of seven rating bands, 1000 to 2200, is likely to play, graded against what Stockfish actually is, here Stockfish 18 at depth 20 scoring every legal move. Two positions sit inside the Accepted, two inside the Declined. On the Declined side, the human favorite already matches the engine's own favorite at every rating we tested. On the Accepted side, it doesn't, and that gap is the most interesting number in this dataset.

Where the name comes from, and why it isn't really a gambit

The opening is old enough to have settled its own history. Per Wikipedia, it is mentioned in the Göttingen manuscript of 1490 and analyzed by Gioachino Greco in the 17th century; Philipp Stamma of Aleppo recommended it strongly enough in the 18th century that it was sometimes called the Aleppo Gambit, though it did not spread widely until the 1873 Vienna tournament. Despite the name, it was never a true sacrifice: per Wikipedia, Black cannot retain the c4 pawn without ending up worse, so White typically regains it or gets ample compensation for leaving it alone. Wikipedia's own illustration is concrete: after 3...b5? 4.a4 c6 5.axb5 cxb5??, the a8-h1 diagonal collapses and 6.Qf3 wins material outright, which is why the article calls that greedy defense "fairly risky and rarely seen." Jahtreezy puts the same point more simply on a chess.com forum thread: "It's one of the few gambits so sound that you aren't really gambiting away a pawn."

The family tree: what this data covers, and what it doesn't

After 1.d4 d5 2.c4, chess.com's own opening database, directly fetched and read in September 2026 across 281,248 sampled games, shows Black splitting three ways at real volume: 2...c6, the Slav Defense (ECO D10-D19), in 118,063 games (White scores 40.0/37.2/22.8 win/draw/loss); 2...e6, the Declined (D30-D69), in 115,120 games (42.4/31.5/26.1); and 2...dxc4, the Accepted (D20-D29), in 33,897 games (38.1/35.7/26.3). The Slav and the Declined are now close to equally popular at this level, worth knowing before assuming the Declined is still the automatic second choice it was a century ago.

Three more named replies trail far behind those three, and our own dataset measures none of them: 2...Nc6, the Chigorin Defense (D07), in 5,957 games, Alexander Morozevich's own signature line per Wikipedia; 2...e5, the Albin Countergambit (D08-D09), in 3,951; and 2...Bf5, the Baltic Defense (D06), rarest at 1,925. Within the two branches this dataset does cover, the move orders are the real ones players reach: White's 3rd move after 2...dxc4 is 3.Nf3 roughly 75 percent of the time, and after 2...e6 it is 3.Nc3 in roughly 96.7 percent of games, near-automatic rather than one option among several.

Accepting: the engine's favorite almost nobody plays

After 1.d4 d5 2.c4 dxc4 3.Nf3, Black has 30 legal moves. Stockfish's own ranking is close at the top: 3...a6 is best, then Nf6, e6, c6, b5 and Bf5, all within a third of a pawn of each other. Further down sit Nc6 and Bg4, each measurably worse, Nc6 costing 43 centipawns against a6 and Bg4 costing 71.

Read the top row first: it rises at every step we measured, 0.3 percent at 1000 to 9.6 percent by 2200, and even that top figure is a move fewer than one in ten players finds. At 1000, Nc6 leads at 26.6 percent, costing 43 centipawns against a6. From 1200 on, the field converges on a different move: 3...Nf6, climbing from 23.3 percent at 1200 to 48.6 percent at 2200, the most common reply at the top of our range. Nf6 is not the engine's top choice, but it is close, only 5 centipawns worse than a6 at every one of those bands. By 2200, the top five are Nf6 48.6, c5 12.4, a6 9.6, e6 8.2 and c6 6.5, and the engine's actual favorite still sits third.

Chess.com's own master database, fetched directly at the identical position and read in September 2026, adds a useful bridge past our top band: across 18,084 sampled games, 3...a6 appears in just 5.7 percent of them, roughly where our own club-rating curve would land if extended past 2200. That is a different population from Maia's online games, not a seamless extension of the same curve, but it points the same direction: even strong players mostly reach for Nf6 over the engine's own preference.

One move later, the Accepted gap nearly closes

The more useful next position to measure is the one chess.com's database shows is standard: after 3...Nf6 4.e3, Black has 33 legal moves. Stockfish rates e6 and a6 tied for best, c5 just 3 centipawns behind; b5, Bg4 and Bf5 trail further, and Nc6 costs a full 60 centipawns.

This is the one place in the Accepted where the gap actually closes. At 1000, Nc6 is still likeliest, at 26.7 percent, a 60 centipawn cost against e6. From 1200 on, e6 itself becomes the most popular choice and stays there, rising to 50.5 percent by 2200, matching Stockfish's own top choice exactly at every one of those bands. Bg4 is the steady runner-up from 1400 onward, holding between 17 and 21 percent. Chess.com's own database at this exact position, 16,829 sampled games, shows 4...e6 in roughly 83 percent of games, well above even our own 2200 figure, one more sign the gap keeps closing higher up the ladder.

Declining: the move that already matches the engine

After 1.d4 d5 2.c4 e6 3.Nc3, Black has 34 legal moves. Stockfish's ranking: Nf6 is best, then Bb4, Be7 and c5 within 10 centipawns of it; c6, Nc6, dxc4 and f5 trail well behind, dxc4 costing 41 centipawns and f5 costing 69.

Here is the headline fact of this article: the likeliest human move is the engine's own top choice at every single rating band, 1000 through 2200, no exception. Nf6 climbs from 33.0 percent at 1000 to a peak of 64.3 at 1800, a real majority of the field, then eases to 64.2 at 2000 and 58.8 at 2200, a genuine non-monotone wrinkle worth stating plainly. dxc4, an immediate transposition toward the Accepted a move later than usual, is the clear second choice at 1000, 19.4 percent, then fades steadily, 12.8, 8.4, 6.0, 3.9, and drops out of the top five by 2000. What fills the gap as Nf6 eases back is c6: 8.2 percent at 1000, climbing to 12.5 by 1800, 16.9 by 2000, then easing to 16.5 by 2200.

One move later, a small dip that isn't a warning sign

After 3...Nf6 4.Bg5, Black has 31 legal moves. Stockfish rates Be7 clearly best, with dxc4, c6, Nbd7 and c5 within 20 centipawns of it, Bb4 a little further back, and h6 costing a full 77 centipawns.

Be7 matches Stockfish's own top choice at every band here too, including the bottom of our range, already leading at 43.5 percent at 1000. It climbs cleanly to 70.9 percent at 2000, then dips slightly to 70.0 at 2200, a small wrinkle worth naming rather than rounding away. h6, costing 77 centipawns against Be7, is a real 1000 rated habit at 15.6 percent, fading steadily, 10.9, 7.7, 5.3, 3.1, gone from the top five by 2000. Bb4 opens around 14 percent in the low bands, 13.9 at 1000 and 14.8 at 1200, before fading to 3.8 percent by 2200, a second beginner habit rating quietly cures.

Why rare doesn't mean wrong here

Put the two branches side by side and the honest reading is not "the Accepted is worse prepared" or "the Declined is safer." It is that findability, our own measure of whether a real opponent finds the engine's own move, punishes the Accepted unfairly. At 1000, the real cost sits almost entirely in 3...Nc6 and, one move later, 4...Nc6 again, the moves that cost 43 and 60 centipawns respectively, and exactly the moves that fade fastest as rating rises. The move everyone notices as rare, 3...a6, costs the field remarkably little once you account for what players actually play instead: Nf6 and e6 are both within 5 centipawns of the engine's own top choice from 1200 up. A low findability score here is not proof of a weak position. It is proof that two good moves sit almost on top of each other, and, on our reading below, humans keep choosing the one that looks like a plan on its own rather than the one whose point is invisible until you already know it.

Research for this piece turned up no source, competitor page or forum thread explaining, in plain language, why club players walk past 3...a6 specifically. Our own reading, not a sourced fact: per Wikipedia, 3...a6 is the Alekhine Variation, and it "prepares ...b5" so Black can actually hold the extra pawn a move or two later rather than simply returning it. Nf6 satisfies a principle every beginner has heard a hundred times, develop a piece toward the center; a6's point, setting up ...b5 and a following ...c5 or ...Bg4 plan, is invisible unless you already know it. Theory does not actually skip the idea, it just delays it: the Accepted's own classical main line, 3.Nf3 Nf6 4.e3 e6 5.Bxc4 c5 6.O-O a6, reaches ...a6 anyway, three moves later than the engine would prefer. Across our full dataset of 14 openings, the Declined ranks 3rd or 4th in mean findability at every rating band; the Accepted ranks 13th at 1000, recovering to 7th from 1400 up, a swing our chess openings by rating dataset describes as a 51.5 centipawn mean gap collapsing to 2.5 as soon as rating moves past 1000, exactly the story above: the early position's gap falls from 43 to 5 and the late position's from 60 to 0, because the costly 1000 rated move is Nc6 in both.

Deeper into the Declined: Orthodox, Lasker, Tartakower, and the minority attack

Our own two positions stop at Black's 4th move, but the Declined's theory runs much deeper. One level further, after Black has castled and played ...Nbd7, White's plan splits four ways across 7,961 sampled chess.com games: 7.Rc1, the classic minority-attack preparing move, at 50.3 percent; 7.Qc2 at 25.9; 7.cxd5, entering the Exchange Variation, at 10.9; and 7.Bd3 at 10.6. None of those forks, or the Lasker, Tartakower and Cambridge Springs systems that branch off even earlier, are in our own dataset. The Exchange Variation, 4.cxd5 exd5, creates the symmetrical Carlsbad structure, and White's standard plan there, pushing the a and b pawns while Black still holds more queenside pawns, is the textbook minority attack; Karpov versus Lautier, Dortmund 1995, with its 27.b5!, is a model execution, and stop memorizing openings and learn the plan instead covers it directly. The Declined still shows up at the top: Carlsen and Caruana drew Game 7 of the 2018 World Championship in one, Carlsen later admitting "What I did was just way too soft...Castling is essentially just an admission that the position is equal."

The Netflix effect, and the one game Beth Harmon actually opens this way

Netflix's own numbers state the show was watched by 62 million member households in its first 28 days, its most-watched limited scripted series at the time. What happened to chess afterward is real but unreconciled across sources: reported sales increases range from Spin Master's "triple digits" to eBay's 215 percent to Goliath Games' 1,048 percent, and reported sign-up increases range from chess.com's own claimed 500 percent to other outlets' differing figures, none sharing a stated methodology. Treat it as directional evidence of a real boom, not a precise number to cite as fact.

The better-sourced detail is narrower. A scene-by-scene analysis at vandevliet.me finds the Harmon-versus-Borgov finale is the only featured game where Beth actually opens with the real Queen's Gambit, 1.d4 d5 2.c4; in the piece's own words, "in all the other games, Beth always played 1. e4." The specific continuation shown was lifted from a real 1993 game between Vasily Ivanchuk and Patrick Wolff. Carlsen himself reviewed the show at 5 out of 6, calling it "extremely enjoyable" and the chess "pretty well" done, while doubting the character arc: "I just found the whole thing a little too unrealistic. I just couldn't quite buy that after being six years without playing a tournament game in an orphanage and you would then be one of the best players in a few years."

Where competitors are genuinely better

None of the above means Chessdrive's own dataset is the only thing worth reading here. Modern-chess.com's guide to the Accepted is the most data-literate competitor page found researching this article: it states "White has 58% against it," names Sergey Rublevsky as the leading top-GM practitioner with 182 games, and gives real per-line engine evaluations, all with zero rating-banded data, exactly the gap this article fills. Chess.com's own paid lesson library, fragmented across at least seven pages and authored by GM Gregory Kaidanov, goes deeper into concrete lines than a single article can, at the cost of being scattered and paywalled.

None of this replaces real game outcomes either. Chess.com's own database, read in September 2026, reports White scoring 42.4 percent wins, 31.5 draws and 26.1 losses against the Declined; 38.1/35.7/26.3 against the Accepted; and 40.0/37.2/22.8 against the Slav, an opening our own dataset does not measure at all. Those are results, not move probabilities: White's edge is real and modest across every branch, exactly what an opening that is not a true gambit, and not a loser for Black either, should look like in practice.

What this data is, and is not

Maia3 predicts what a human at a given rating is likely to play. It is not a record of real games, and its own project description says as much, describing it as reproducing "common human biases" and making "many of the same mistakes that humans make." Every percentage in this article is a model of human tendency, not a box score.

The scope is narrower than "the Queen's Gambit" as a whole. This dataset measures exactly four positions: the Accepted's 3rd and 4th moves, and the Declined's 3rd and 4th moves. It says nothing about the Slav, the Albin or the Chigorin, and nothing about White's other 3rd-move tries after 2...dxc4, which together make up roughly a quarter of chess.com's sampled games. The pattern is not unique to this opening either: our own King's Indian Defense data found the leading move staying in first place while losing share as rating rises, and our Caro-Kann Defense data found one main line overtaking another partway up the range. The Accepted's own near-tie is a third shape neither of those two openings shows.

To see every position in this article laid out on an actual board, Chessdrive's interactive Queen's Gambit Declined page and its Queen's Gambit Accepted counterpart lay the same positions out on a board. If your own 1.d4 repertoire runs through a different setup entirely, our London System data measures the other common way to reach a 1.d4 middlegame without offering the c4 pawn at all. And if you want to check whether you found the engine's move in your own last Queen's Gambit game rather than guess from a table, Chessdrive's chess analysis board will tell you directly.