Evans Gambit: What Each Rating Plays After 4.b4, and Why 5...Bc5 Costs Black

The Evans Gambit starts 1.e4 e5 2.Nf3 Nc6 3.Bc4 Bc5 4.b4, offering a pawn to speed up White's centre. Pages on it explain the theory well enough: the history, the named lines, the traps. What none of them shows is what players at each rating do once the pawn is offered. We ran the maia3-5m checkpoint, see how Maia chess actually works, on the position after 4.b4 and on the position after 4...Bxb4 5.c3, at seven rating bands from 1000 to 2200, and graded every legal move with Stockfish 18. Percentages on this page are Maia3's predicted probability that a player at that rating chooses the move. They are model predictions of human play, not game counts or win rates.

Both dataset positions have Black to move, so we converted every evaluation to White's side. A plus sign means White is better, and where Black is better we say so in words. The dataset tables come from Stockfish at depth 20. The line checks are separate local runs of ten seconds per line that reached depths 17 to 25, so a position can differ by several hundredths between the two methods.

What the Evans Gambit is, and where it came from

The opening is an attacking line of the Giuoco Piano, see the Italian Game page for the quiet setups. Wikipedia files it under ECO C51 for the non main lines and C52 for the main line, 4.b4 Bxb4 5.c3 Ba5. It is named after William Davies Evans, a Welsh sea captain, and the first game with it was Evans against McDonnell in London in 1827. That game used a slightly different move order, so "first played in 1827" is only loosely true. William Lewis published the first analysis in 1832.

The history has a few peaks. Anderssen's Evergreen Game against Dufresne in 1852 opened with it. Morphy and Chigorin played it and preferred the Morphy Attack, 9.Nc3, in the main line. Then Lasker's defence, returning the pawn under favourable circumstances, dealt the opening what Wikipedia calls a heavy blow. Sources differ on the date: masterinchess gives 1895, while Chess.com ties the defence to Lasker's 1897 game against Chigorin. ChessBase reports that Kasparov beat Anand in 25 moves at Riga in 1995 and never tried it again with White. Per ChessBase, Carlsen has played it in blitz and in 2024 Casablanca games, and Fischer played it only in a casual game against Fine. Treat those last two points as of October 2026.

First decision for Black: take the pawn?

After 4.b4, Stockfish ranks 4...Bxb4 first, with Black ahead by 0.21. The alternatives are all worse for Black: 4...Be7 gives White +0.28, 4...Bb6 gives +0.30, 4...Nxb4 gives +0.33 and 4...Bd4 gives +0.83. One tempting reply, 4...Nf6, is a blunder at +3.64. Here is what Maia3 predicts.

The likeliest move is the engine's best move at every band, so the gap is zero. Accepting is 61.0 percent at 1000 and climbs to 87.3 percent at 2200. Declining with 4...Bb6 is second everywhere, peaking at 20.4 percent at 1200 and falling to 9.2 at 2200. The small share of 4...Nxb4 and 4...Bd4 shrinks steadily with rating, which is what you would expect, since both cost a few tenths of a pawn or more.

One forum player pointed out that you rarely get to play the Evans at all, because most players answer 3.Bc4 with 3...Nf6, not 3...Bc5. That is a fair caution: this page only matters once Black has played the bishop to c5.

The retreat: where ratings split

After 4...Bxb4 5.c3 the bishop is attacked and must move. Stockfish ranks 5...Ba5 and 5...Be7 as a tie, with Black ahead by 0.23. 5...Bf8 gives White +0.30, 5...Bd6 gives +0.45, and 5...Bc5 gives +0.13. That makes 5...Bc5 0.36 pawns worse for Black than the best retreat. The reason is simple. After 5...Bc5 White plays 6.d4 with a gain of time, a point Wikipedia notes when it calls 5...Bc5 inferior to the other two.

This is the crossover the whole page is about. From 1000 to 1600 the likeliest human move is 5...Bc5, at 79.7, 73.7, 61.9 and 47.3 percent. At 1800 5...Ba5 overtakes it, 38.5 percent against 31.9, and by 2200 it reaches 51.1 percent while 5...Bc5 falls to 8.5. 5...Be7, the Anderssen Variation and the most frequent alternative to 5...Ba5 according to Wikipedia, grows from 2.5 percent at 1000 to 30.6 at 2200.

The dataset's own findability counts only 5...Ba5, which runs from 13.8 percent at 1000 to 51.1 at 2200. Since 5...Be7 scores identically, the fairer figure is the sum of the two: 16.3 percent at 1000 and 81.7 percent at 2200. Put plainly, about one in six players at 1000 finds a best retreat, and four in five at 2200 do.

The size of the mistake needs the same honesty. 0.36 pawns is real but modest. It is not a blunder, and the position after 5...Bc5 is still playable for Black. The data shows habit, not disaster: below 1800 the bishop goes back to c5 because that is where it came from.

Main line: 5...Ba5 6.d4

The main line confirms that the gambit gives White no objective edge. After 5...Ba5 6.d4 Stockfish shows Black ahead by 0.07 at depth 24, with 6...d6 the best reply and 6...exd4 level at 0.00. That is a tiny edge for Black, and also the clearest sign that the gambit's value is practical.

• 6...exd4 7.O-O. Black is ahead by 0.14 with 7...Nf6, and 7...Nge7 is close behind at 0.03 in Black's favour. The other try, 7.Qb3, leaves White at +0.06 after 7...Qe7. White has the move and the initiative here, but the engine does not find a clear plus.
• The Compromised Defence, 7...dxc3. As of October 2026, Wikipedia says it is not well regarded. Our check finds it sharp for both sides. After 7.O-O dxc3 only 8.Qb3 keeps White better, at +0.51. The natural looking 8.e5 falls to Black ahead by 0.98, and 8.Ba3 to Black ahead by 1.28. If you play White and Black plays this, 8.Qb3 is the move to know.
• Lasker style, 6...d6. Black returns the pawn for a simple position. After 6.d4 d6 the engine prefers 7.Qb3 at Black ahead by 0.13, then 7.O-O at Black ahead by 0.39 and 7.dxe5 at Black ahead by 0.56. This is why Lasker's idea hurt the gambit: every sensible White move leaves Black slightly better.

The named White attacks are a catalogue of ways to keep the initiative. Masterinchess lists the Fraser, Waller, Richardson, Levenfish and Steinitz attacks, and Wikipedia covers the Morphy Attack, 9.Nc3. You do not need them all. What you need is an idea of why White does not mind equality: the pieces come out quickly and the opponent's king stays in the centre for a while.

The Anderssen line: 5...Be7

The retreat to e7 is the quietest choice. After 5...Be7 6.d4 Na5, White's best is 7.Bd3 at +0.20, and 7.Nxe5 and 7.Be2 score 0.00 and Black ahead by 0.16. So the engine gives White at most a fifth of a pawn there. If you play White and your opponent chooses e7, expect a long game, not a quick attack.

The popular retreat: 5...Bc5

The line most players below 1800 will meet runs 5...Bc5 6.d4 exd4 7.O-O. Stockfish gives White +0.42 after 7...Bb6, and 7...d6 and 7...Na5 both give +0.37. White's small edge from the dataset has grown to about four tenths of a pawn once the centre opens. The 7...d6 branch is the Fraser Attack that masterinchess lists, 8.cxd4 Bb6 9.Nc3 Bg4 10.Qa4. The practical advice for White is to learn this line first, since it is what the data says your opponents are most likely to play up to 1600.

Evans Gambit declined

Black can refuse the pawn with 4...Bb6. As of October 2026, Wikipedia reports that most commentators call that weaker than accepting. Our numbers agree, but only mildly. Stockfish gives White +0.40 after 5.a4, +0.39 after 5.O-O and +0.14 after 5.c3. So there are two good moves for White and a quieter third.

• 5.a4 a6. This is the Showalter Variation when White follows with 6.Nc3. The engine prefers 6.d3 at +0.41, with White keeping a small, steady edge.
• 5.b5 Na5 6.Nxe5. This grabs a pawn but is not best. Black replies 6...Nh6 and stands better by 0.34. Greed does not pay here.

The declined line is a sensible practical choice for Black. It avoids the sharp theory, it keeps the bishop on a good diagonal, and it leaves White with a modest edge of about two fifths of a pawn. At 1200, where Maia3 puts 4...Bb6 at 20.4 percent, one player in five already chooses it. For another 1.e4 e5 gambit graded the same way, see the Danish Gambit page.

What to play as each colour

As White, learn one line well. If you are 1000 to 1600, the data says your opponent will most likely retreat 5...Bc5, so prepare 6.d4 exd4 7.O-O and the +0.42 position. Know 8.Qb3 against the Compromised Defence, and know that 5.a4 and 5.O-O are your tries against 4...Bb6. Beyond 1800, expect 5...Ba5 or 5...Be7 and a position that is close to level. Then the gambit becomes a question of how well you know the plans. If you prefer to learn ideas instead of move lists, why you should stop memorizing openings and learn plans fits this opening well.

As Black, take the pawn and retreat to a5 or e7. That is the engine's choice and the choice of strong players, and it saves the 0.36 pawns that Bc5 gives away. If you do not want the complications, 4...Bb6 is a fair way out. And if you are choosing an opening for your rating generally, our chess openings by rating dataset uses the same method across many positions.

Is the Evans Gambit sound?

Yes, and it is also roughly equal. Black is ahead by 0.21 after taking the pawn, ahead by 0.23 after the best retreat, and ahead by 0.07 after 6.d4 in the main line. Our deeper checks of the 5...Ba5 main line run from Black ahead by 0.14 to White ahead by 0.06, and White stays near +0.40 when Black declines. No engine number in this page shows White winning. What White gets is play: a lead in development, a centre, and a Black player who has to find the best retreat and the right answer to d4. The dataset says Black below 1800 often does not, and that costs about a third of a pawn. A cost that size will not decide a game on its own, but it is the margin White is playing for, and it is larger than anything the engine finds in the main line itself. The gambit works best as a way to reach positions where one side knows the plans and the other is guessing.

That is the honest case for the gambit. Wikipedia sums up Lasker's point, that Black can return the pawn under favourable circumstances. Fine thought Black has the harder defence, which masterinchess reports. Both can be true: the engine sees equality, and humans find it hard to hold.

Where other pages are better

This page is narrow on purpose. The dataset holds two positions, the one after 4.b4 and the one after 4...Bxb4 5.c3, and nothing deeper. The line checks are quick local runs, not an opening database, and we have no game results by rating. Wikipedia is the better reference for the full list of named variations and the history, ChessBase for the story of the famous games, and masterinchess for move sequences. None of the pages we opened splits Black's replies by rating, or shows that the likeliest retreat below 1800 is the weaker one, which is the gap this page fills.

What to take from it

The Evans Gambit is accepted by 61.0 percent of 1000 rated players and 87.3 percent at 2200, and that is also the engine's choice. The decision that separates ratings is the retreat: 5...Bc5 is the likeliest move through 1600 at 79.7 down to 47.3 percent, and it gives away 0.36 pawns against 5...Ba5 or 5...Be7, which together reach 81.7 percent at 2200. Declining with 4...Bb6 is playable and leaves White about +0.40. In the Compromised Defence only 8.Qb3 keeps White better. Maia3 is a model of human choice from the maia3-5m checkpoint, not a count of real games, and every percentage above should be read that way.