When and why is it good to deny the development of White's dark squared bishop in the Caro-Kann exchange variation?

Question

In the position after

[FEN ""] 
[startlfipped "10"]
[Startply "10"]

1. e4 c6 2. Nf3 d5 3. exd5 cxd5 4. d4 Nc6 5. Bd3 Bg4 6. O-O? Bxf3! (6...Qc7? 7. c3) 7. Qxf3 Nxe4 
      

White has played inaccurately because Black can get a better version of the mainline of the exchange variation, as White cannot easily develop the bishop to f4 after 6. c3 and 6. ..Qc7 (this is the engine recommendation and the recommendation in my book).

However, when White instead plays 6. O-O, the engine thinks that Black is already better (-0.6) after taking on f3 (6. ..Bxf3). 6. ..Qc7 however is a sizeable mistake that not only throws away the advantage but even puts Black on the backfoot who then already has a pretty worrisome position (0.8).

I don't understand why

1. 6. O-O is so bad for White and why ..Bxf3 is so strong in response
2. 6. ..Qc7 is such a massive mistake, even though it still denies the natural development of the dark-squared bishop and is the best move when White plays 6. c3.

How could this be explained?

Answer 1

I add another answer since the crucial point hasn't been addressed, methinks. Indeed, after 6...Qc7 7.c3 it's a simple transpose and any decent algorithm shouldn't depend on move order to the position.

But White plays 7.h3. Black now can't win d4 (d5 hangs as well) after Bxf3, so this merely cedes the bishop pair. And if 7...Bh5, 8.g4 Bg6 9.c4. (Computer are never afraid. Unsurprisingly, no human - on Lichess - ever tried this. Even I wouldn't. Evaluation is a slight advantage for White.)

Answer 2

The 6...Bxf3 tactic wins the d4 pawn by removing its defender. Sure, it gives up the two bishops to do so, and that's the reason why black's not a full point up in evaluation. But that's not enough compensation for even a human white, and the computer cares about it even less.

After 6... Qc7 the tactic no longer works (The d5 pawn is also hanging after ...Bxf3 Qxf3, and ...Nxc2 won't work if it doesn't come with check because the black king is exposed) and so white can forgo c3 and ends up having gained the tempo of castling in exchange, making this better than the other line.

Answer 3

In case anybody's wondering why after 6 0-0? Black doesn't keep the Bishop pair and play Nxd4 directly, which may seem even better: That falls into a trap familiar (with colors reversed) from the QGD, losing a piece to 7 Nxd4! because after Bxd1 8 Bb5+ Qd7 9 Bxd7+ and 10 Rxd1 (or first 9 Rxd1) White's still ahead N for P and should win the ending easily.

Answer 4

After 6.O-O Bxf3 7.Qxf3, Black plays Nxd4 and is up a central pawn. That's enough for the evaluations to flip that much.