Chess French Defense: What Every Rating Actually Plays

On a chess.com forum thread about the French Defense, one poster calls 3...c5 against 3.Nc3 "way too rare to have a mainline," while another cites a single 1999 game as giving White a large advantage. Both are reacting to the same real habit: a move a large share of players actually reach for at that point, whether or not any specific guide recommends it. Chessdrive's own Maia3 data measures that habit directly. We ran the maia3-5m checkpoint, see how Maia chess actually works, against every legal reply to 1.e4 e6 2.d4 d5 3.Nc3 at seven rating bands from 1000 to 2200, the same method behind our chess openings by rating dataset, then graded every reply with Stockfish 18 at depth 20, see what the Stockfish engine is. Two positions only, one move order, 3.Nc3 specifically. Nothing here measures the Advance, the Exchange or the Tarrasch.

What the French Defense actually is

The French Defense begins 1.e4 e6, most commonly followed by 2.d4 d5, and covers the whole ECO range from C00 to C19. Per Wikipedia, the name comes from a correspondence match played between London and Paris from 1834 to 1836; the Paris team's use of the defense there was promoted by Jacques Chamouillet. White's third move splits the opening into four systems. 3.Nc3, the one this article measures, itself splits three ways depending on Black's reply: 3...Bb4 is the Winawer, 3...Nf6 is the Classical, and 3...dxe4 4.Nxe4 is the Rubinstein. 3.Nd2 is the Tarrasch Variation. 3.e5 is the Advance Variation. 3.exd5 is the Exchange Variation. Chess.com's own French Defense overview, as read in September 2026, reports the Tarrasch scoring White wins of 42 percent, draws of 30 percent and losses of 28 percent in its sample, and describes the Exchange as mostly heading to draws, with "Black having a slight edge... in decisive games." Wikipedia's history section names Botvinnik, Korchnoi, Rubinstein, Nimzowitsch, Petrosian, Psakhis, Uhlmann and Vaganian among the opening's notable contributors. That is a history, not a claim about who plays it now.

The bad bishop and the two pawn breaks

The trade-off in the whole opening is that pawn on e6. Wikipedia describes it as impeding the development of Black's own c8 bishop, "known by many players as the French bishop," and its go-to illustration is Tarrasch against Teichmann, San Sebastian 1912, where that bishop ends up boxed in by Black's own pawns on a6, b5, d5, e6 and f7. Black's standard answer is not to fix the bishop but to fight for the center instead, with one of two pawn breaks: ...c5, playing directly against White's center, or ...f6, used when the c5 break on its own is not enough. Both breaks show up inside the two positions measured below, just not always at the rating a repertoire book would expect.

Position A: after 3.Nc3, what Black actually reaches for

After 1.e4 e6 2.d4 d5 3.Nc3, Black has 34 legal moves. Stockfish 18 at depth 20 keeps the top two nearly tied: 3...dxe4, the Rubinstein, is best at -0.41, and 3...Nf6, the Classical, sits a hair behind at -0.48. 3...Bb4, the Winawer, is only the eighth choice at -0.76. Far down the list, at rank 18 of 34, sits 3...c5, at -1.37, about a pawn worse than dxe4.

At 1000, c5 is the single most likely move on the board, at 38.0 percent, more than double dxe4's 18.0 percent. It peaks at 42.1 percent at 1200 and is still narrowly the most likely move at 1600, at 25.3 percent against 24.2 for Nf6 and 23.8 for dxe4, then keeps falling to 2.4 percent by 2200. Nf6 runs the opposite direction: third choice at 1000, it overtakes both c5 and dxe4 by 1800 and finishes as the single most common reply at 2200, at 39.8 percent. Bb4, the Winawer, climbs the whole way too, from a distant 10.4 percent at 1000 to 37.7 percent at 2200, a close second to Nf6 there. dxe4 itself is the odd line out: it rises through the middle bands, peaks at 24.0 percent at 1800, then drops back to 15.5 percent at 2200, lower than where it started. The engine's best move does not automatically get more findable as rating rises. Here it does, then it does not.

The mean eval gap row tells the same story from the other side of the ledger. Averaged across both positions this dataset measured, it holds flat at 53.5 centipawns, close to half a pawn, for every band from 1000 through 1600, then drops to 9 centipawns from 1800 on, a fall of 44.5 centipawns. That drop is entirely this position's own doing: through 1600 the likeliest human move is c5, 96 centipawns worse than Stockfish's best, while from 1800 on the likeliest move is Nf6, only 7 centipawns off it. The second position's gap, covered below, is 11 centipawns at every band. Half a pawn of engine preference at move three is real, but it is not the kind of margin that decides a club game by itself. The other thirty or forty moves of the game still have to be played.

None of this makes 3...dxe4 a bad choice, even where it is not the most popular one. Chess.com's own page on the Rubinstein Variation describes the capture on e4 as a way to "clear the center and allow fairly smooth development," and lays out its case plainly: safe, simple, and it avoids risk, at the cost of a small space disadvantage, fewer chances to play for a win, and a generally passive character. A secondhand summary of a New In Chess piece on the same line goes further, describing it as trading the center for what it calls a rock-hard, low-risk structure with no real weaknesses for White to target. That specific framing comes from a search snippet rather than the full article, so treat it as one source's pitch rather than settled fact. Chess.com's Rubinstein page does not mention 3...c5 as an alternative at all, even though it is the move Maia3 predicts for nearly four in ten players at 1000. It is the same kind of trade-off as a flexible, low-theory setup against 1.e4 more generally, the subject of our Caro-Kann Defense findability data: give up a little activity for a structure with fewer ways to go wrong.

Why 3...c5 is a low-rated favorite, not a recommended one

The clearest case against 3...c5 that we found sits in an old chess.com forum thread, not on the overview pages. One poster, aggressivesociopath, gives a concrete line for White: 4.exd5 exd5 5.dxc5 d4 6.Bb5+ Nc6 7.Qe2+ Be6 8.Ne4 Nf6 9.Bg5 Qa5+ 10.Bd2 Qd8 11.Ng5 Qd5 12.Nxe6 fxe6 13.Nf3 Bxc5 14.O-O, citing the game Rausis against Tarira, Lisbon 1999, and calling the result "a large advantage to White." A second poster on the same thread, jphillips, argues the deeper reason c5 belongs elsewhere: "I think it's only ok on the third move in the Advance and the Tarrasch." Played against 3.Nc3 specifically, that poster argues, c5 leaves Black with an isolated d-pawn and a worse version of the Tarrasch's own structure, which is why, in that poster's words, the line is "way too rare to have a mainline." Not every source that has looked at this exact position agrees on the precise refutation, so treat the cited line as one strong try rather than a proven refutation. What the two posters agree on is the direction, and our own numbers fit it: the more theory a band is likely to know, the less often the model predicts c5.

Position B: after 3...c5 4.Nf3

Follow the most popular human reply at the reference rating used to build this dataset, 1400, two plies deeper and the board reaches 1.e4 e6 2.d4 d5 3.Nc3 c5 4.Nf3, a position with one foot in another opening: the board after 3...c5 can also arise from a Sicilian move order, 1.e4 c5 2.Nc3 e6 3.d4 d5, which is one reason to read this next to our own Sicilian Defense data. Black has 34 legal moves here too. Stockfish's own order is close at the top: 4...cxd4 is best at -0.63, then Nf6 at -0.71, Nc6 at -0.74, c4 at -0.79 and dxe4 at -0.84.

Nc6 is the single most popular reply at every rating band measured, not cxd4, Stockfish's actual best move, and that gap closes steadily rather than suddenly: cxd4 climbs from 20.2 percent at 1000 to 34.3 percent at 2200, while Nc6 slides from 46.1 percent down to 37.3 percent, without ever losing the top spot. Nf6 dips through the middle bands, down to 7.4 percent at 1400, then recovers to 20.6 percent by 2200, a genuine U shape this dataset does not explain on its own. What does not move at all is the eval gap: because the likeliest human move stays Nc6 at every single band, the distance between it and cxd4, 11 centipawns, a bit over a tenth of a pawn, never changes from 1000 to 2200. Position A's gap fell from 96 centipawns to 7 over the same climb in rating. This one holds completely flat.

What the gap actually means

Maia3 is a model's prediction of what a human at a given rating is likely to play, built from the maia3-5m checkpoint. It is not a count of real games, and every percentage above should be read that way. The two positions here are also the entire scope of what this dataset measures: one move order, 3.Nc3 specifically, two positions two plies apart. It says nothing about the Advance, the Exchange or the Tarrasch, and nothing about any French line past Black's fourth move. If your own opponents open 3.e5 or 3.Nd2 instead, none of the tables above describe what they are likely to do there.

To see either position rendered on an actual board, move by move, at your own rating band, Chessdrive's interactive French Defense page lays out exactly that.