Your best Anti-Sicilian depends on your ELO. 6.6M games show how.

You've played the Sicilian. Or against it. It's one of the most-played defenses to 1.e4 on Lichess, and Black scores slightly better than the average defense at every rating, which is why it survives at every level.
But the Sicilian isn't one opening. It's effectively six different openings depending on what White plays on move 2, and which of those six wins changes radically as you climb the rating ladder.
I pulled 6.6 million Lichess games across five ELO bands (0–800 through 1600–2000). The popularity distribution flips. The win rates flip in some lines. Some moves don't even appear below 1400. Here's the full picture, and what it means for how you should actually prepare the Sicilian at your level.

What the data actually shows

The dataset draws from the Lichess Opening Explorer at 1.e4 c5, partitioned by ELO band. Sample sizes are large enough that ranking noise isn't a concern: 39K games at 0–800, 220K at 1000, 700K at 1200, 1.6M at 1400, and 4.1M at 1600–2000. For each White 2nd move I tracked popularity, White win rate, Black win rate, and the Stockfish eval of the resulting position. Worth flagging up front: Stockfish evaluates the position, not the players. The interesting question isn't "what does the engine think" but "where do engine evals and human results diverge?" The Sicilian has clean examples of both.

Here's the popularity distribution by ELO:

The three biggest shifts between 0-800 and 1600-2000 ELO:

• 2.Nf3 popularity: +32.6 points (24.8% → 57.4%)
• 2.Bc4 White win rate: -9 points (50% → 41%)
• 2.d4 acceptance by Black: +53 points (37.6% → 90.6%)

Four moves migrate substantially. Two move in opposite directions. Let's walk through them.

2.Nf3 doubles in popularity, the win rate barely moves

The biggest shift in the entire Sicilian tree is the rise of 2.Nf3, the Open Sicilian. From 24.8% at 0–800 to 57.4% at 1600–2000. A +32.6 point climb, the largest of any 2nd move in the dataset.
What's striking is what doesn't change: White's win rate. Across all five ELO bands, the Open Sicilian gives White 47–49%, a range of just 2 points. Whatever you think about "Sicilian theory ocean," the practical edge it gives White at 800 ELO is essentially the same edge it gives at 2000.
What does change is Black's response distribution. Below 1200, 2...Nc6 dominates. At 1600–2000, 2...d6 takes over at 39.9% versus Nc6's 36.1%. The Hyperaccelerated Dragon (2...g6) appears at 5.1%, the O'Kelly Variation (2...a6) at 2.8%, both effectively invisible below 1200.
Read this as a theory adoption curve. Strong players migrate to the Open Sicilian because they've read the books, and Black's diversification reflects deeper preparation. But the practical result is the same. If you only take one thing from this section, it's that 2.Nf3 is the ELO-agnostic Sicilian move. It works at every level you'll ever reach, without changing the math.

The Bowdler collapses, the Alapin climbs

Now the contre-intuitive part. Two Anti-Sicilians move in opposite directions, and the contrast is the cleanest evidence in the data that popularity and correctness aren't the same thing.
The Bowdler Attack (2.Bc4) is the single biggest ELO trap in the Sicilian. White's win rate: 50% at 0–800, 47% at 1000, 45% at 1200, 43% at 1400, 41% at 1600–2000. A monotonic 9-point collapse, the worst of any popular 2nd move. Popularity stays around 18% until 1400, then drops to 11.4% as players discover the refutation. At 1600–2000, 41.3% of Black players play 2...e6, and White scores only 39% against it.
The Alapin (2.c3) does the reverse. Popularity rises monotonically from 2.6% to 5.8%. White win rate at 1600–2000: 48%, the highest of any popular Anti-Sicilian at that level. It beats 2.Nf3 (47%), 2.Nc3 (46%), 2.d4 (46%), 2.Bc4 (41%).
Why does it work practically? Because Black mostly doesn't know the theory. The engine-correct reply is 2...d5, which scores 51% for Black, but only 18.6% of strong Black players actually play it. The popular reply, 2...Nc6 (24% of choices), gives White 52%. The Alapin survives on information asymmetry.

41.3% of Black players at 1600-2000 ELO play 2...e6 against the Bowdler. White scores only 39%.

The Smith-Morra is a behavior trap, not a theory trap

The Smith-Morra Gambit (2.d4) is the third interesting case. Overall White win rates are unremarkable across ELO (43–49%), but underneath that average sits the cleanest behavioral shift in the entire Sicilian.
Acceptance rate of the gambit (Black plays 2...cxd4) by ELO: 37.6% at 0–800, 44.2% at 1000, 59.3% at 1200, 75.4% at 1400, 90.6% at 1600–2000. Strongest monotonic shift in the dataset.
At low ELO, most Black players decline. And declining costs real points. At 1000 ELO, after 2.d4 d6, White scores 52% to Black's 43%, with Stockfish at +1.2 (a clear positional edge for the pawn). At 1200, it's worse: W 55%. Black is essentially refusing a pawn and accepting a structurally worse position.
By 1600–2000, the mistake has mostly been eliminated. Once accepted, the gambit is roughly balanced. At 1600–2000 after 2...cxd4, Black actually scores 50% to White's 45%. This is a textbook example of the engine-vs-human gap: Stockfish rates the gambit at +0.1 (basically equal), but the practical White edge below 1400 comes entirely from Black's behavior, not from the position.

Declining the Smith-Morra. At 1000 ELO this scores 43% for Black against 52% for White, with Stockfish at +1.2.

The moves you forget exist: Closed Sicilian and McDonnell

Two lines complete the picture.
The Closed Sicilian (2.Nc3) is the only Anti-Sicilian with a non-linear result curve. White's win rate drops from 50% at 0–800 to 42% at 1400, then recovers to 46% at 1600–2000. Popularity declines steadily (13% to 7.8%). The recovery at the top suggests well-prepared strong players exploit opponents who don't know the Closed framework. Stockfish eval is +0.3 at every level. The theory is sound; the dip at 1400 is a practical handling issue, not a theoretical one.
The McDonnell Attack (2.f4) is the only Sicilian 2nd move that doesn't appear below 1400 in the dataset at all. It emerges at 2.0% at 1400 and grows to 3.2% at 1600–2000. Win rate is a stable 45% for White at both levels, with Stockfish at 0.0. A real surprise weapon, not a refuted novelty. If you want to dodge Open Sicilian prep at higher ELO, it's a legitimate option that most opponents haven't faced practically.

What to do with this

Here's the win rate side of the picture for reference before the takeaways (format: White Win % / Draw % / Black Win %):

If you're under 1400 ELO:

• As White, drop the Bowdler. It's the only line where your win rate already declines at your level.
• 2.Nf3 gives ~49% White across 800–1200 without requiring deep Najdorf prep.
• If you face the Smith-Morra as Black, take the pawn. Declining is the highest-cost common mistake in the dataset at this level.
• Don't fear 2.Qh5. It already loses at 1200 (W 40%, B 57%) and Stockfish hates it (–0.4 to –0.6).

If you're 1400+ ELO:

• The Alapin (2.c3) is the highest White win rate at your level (48%). Most opponents won't know 2...d5.
• If you stick with 2.Nf3, prep needs to cover diversification. At 1600–2000 you'll face d6, Nc6, e6, g6, and a6 with measurable frequency.
• 2.f4 (McDonnell) is a legitimate surprise weapon, stable at 45% White and absent from any opponent's practical experience below 1400.

If you'd rather have all of this on a wall as a decision tree, with branches sized by popularity and win rates / stockfish evaluation per node, I made a Sicilian poster for that.

The takeaway

Two patterns to hold onto. First, the Sicilian behaves like six different openings, and which one favors White depends almost entirely on the ELO of the players. The Bowdler works at 800 and fails at 2000. The Alapin does the reverse. Strong players migrate to 2.Nf3 not because it's better in absolute terms (win rate is flat at 47–49%), but because the deeper theory becomes accessible at their level.
Second, the gap between engine evaluation and human practical results is the whole story. Stockfish doesn't care who plays; humans do. Opening choice at amateur level rewards information asymmetry and behavioral exploits, not theoretical purity. The Smith-Morra wins below 1400 because Black declines. The Alapin wins above 1600 because Black doesn't know 2...d5.
Which Sicilian Anti are you actually playing, and does the data match your experience?

I run MyChessPosters, where I turn Lichess data into ELO-specific decision-tree posters. If you found this analysis useful, the full visual breakdowns for the Sicilian and other openings are here.