Chess Openings by Rating: Which Ones Are Actually Findable

No single one of the popular chess openings is best at every rating. Every serious source on the subject, including the coaches quoted further down this page, refuses to name one, and after building a first party dataset across 14 openings, 28 positions and seven rating bands, so are we. What we can say is narrower and more useful: which openings hand the average player at a given rating a position they can actually navigate correctly, and which ones only look good on a list because the opponent has to cooperate too.

The short version, before the data: the field splits into two tiers rather than a ranking, the gap between the easiest and hardest opening we tested shrinks from 53.5 centipawns at 1000 rating to 10.0 at 2200, and at least one opening near the top of every tier we measured, the Englund Gambit, is not actually a good opening, it is just simple to play correctly. All three of those claims need the fuller explanation below to not be misread.

Where the coaches actually disagree

IM John Bartholomew's own published resource guide is explicit: below 1400 rating, "Openings are least important at this stage, and you should only worry about developing a basic repertoire (under 10 moves)." From 1400 to 2000 he recommends "picking a single opening repertoire for each side" and learning theory "up to Move 10 to 15 for the main lines and key sidelines at your level." Only at 2000 and above does he suggest spending real extra time "deepening your opening understanding."

Other strong voices start earlier or go further the other way. A chess.com forum thread summarizing Daniel Naroditsky's advice reports him telling students they need at least some basic opening knowledge to progress, pitched at players under 1000 rating, noticeably earlier than Bartholomew's under 1400 minimal ask. Ben Finegold's stream persona, per community framing rather than a verified primary transcript, goes the other way entirely, treating opening choice as close to irrelevant even at 1500 with lines like "if you're 1500 and start the game h3 Rh2, it won't matter." These are not the same claim, and blending them into one "here's what the coaches say" paragraph would misrepresent all three.

The skeptical position shows up unprompted in the same chess.com forum threads that ask what opening a beginner should learn. One reply: "For a beginner, specific openings don't matter. Learn basic opening principles." Another, addressed to a 764 rated player asking about openings: "You are rated 764. That means you will never win or lose because of the opening but because of tactical blunders by you or your opponent." That is not a fringe view, and it is one we already agree with in the piece where it actually belongs: our own article on why puzzle rating and game rating don't match argues tactics recognition, not opening memorization, is what closes the gap between knowing a pattern exists and noticing it yourself mid-game. Nothing below contradicts that. Opening choice and tactical training are not competing for the same hour, and the data in this piece is about which openings are worth the smaller amount of time openings do deserve, not an argument that they deserve more of it than they currently get.

What we tested

We ran Stockfish 18 at depth 20 against maia3-5m, the smallest checkpoint Maia3 recommends for general use, run through maia3-js on our own pipeline on 29 August 2026. Every position had MultiPV set to its exact legal move count, so no legal move was ever left unscored by the engine. For background on what Stockfish's search and NNUE evaluation actually do, and their own honest limits, see our companion piece on what Stockfish actually is; for what Maia3 is and how it differs from a depth-limited engine, see our Maia chess reference.

We covered 14 openings: Italian Game, Ruy Lopez, Scotch Game, Sicilian Defence (Open), French Defence, Caro-Kann, Scandinavian, Queen's Gambit Declined, Queen's Gambit Accepted, London System, King's Indian Defence, Vienna Game, King's Gambit and Englund Gambit. Each opening got two positions: an "early" position right after its defining move sequence, and a "late" position two plies deeper, down the single most popular continuation as judged by Maia3 itself at a reference rating of 1400, the median of the seven bands we measured. That reference rating is a disclosed, reproducible choice, not the only defensible one; a different reference elo could pick a different two ply continuation for some openings, and the exact move and its probability at that reference rating is recorded per opening so the choice is auditable rather than hidden. We then queried both models at seven self-rated bands: 1000, 1200, 1400, 1600, 1800, 2000 and 2200.

Two numbers come out of every position and rating band. Findability is the probability Maia3 assigns to the specific move Stockfish calls best. Eval gap is the centipawn difference between Stockfish's best move and the move Maia3 actually rates most likely a human would play, so 0 means the typical player's move and the engine's move are the same move. We used this same metric on a single historical game in our piece on what chess analysis can and cannot tell you; this dataset applies it across a much wider set of opening positions instead of one finished game. Every FEN in this run was built two independent ways, once from the human readable move list and once from the exact move history handed to Maia3, and any mismatch would have aborted the run before writing output; none occurred. Every probability distribution was checked to sum to within 0.99 to 1.01 across all legal moves at every position and rating band, with a second independent query run automatically whenever the engine's own top move read exactly 0.0 probability.

Findable is not the same as good

The Englund Gambit, 1.d4 e5, is the clearest proof in this dataset that "easy to play correctly" and "objectively sound" are different questions. GM Boris Avrukh has called it "the worst possible reply to White's first move," and it is almost never seen in top level play because accurate play by White keeps a lasting advantage. Our own numbers agree completely: Stockfish already scores the position at plus 1.38 for White right after 1...e5, before White has even played the recapture, and that advantage barely moves, plus 1.33 two plies later down the most popular continuation.

And yet the Englund Gambit sits at a mean eval gap of exactly 0 at every single rating band we tested, 1000 through 2200, and its findability climbs from 58.0 percent at 1000 rating to 77.2 percent by 2200, among the highest in the entire set. Both things are true at once: White is already comfortably better, and the typical player at every rating we measured plays the correct, engine-approved move anyway. The opening scores well here because the position is simple to navigate, not because the opening is good. A very similar pattern shows up in the Stafford Gambit, which we did not include in this dataset but which Wikipedia's own article describes as "objectively unsound," a line that became genuinely viral through IM Eric Rosen's content specifically because an unprepared opponent needs "almost perfect moves" to avoid its traps. Findability measures how hard a position is to navigate. It never measures whether the position was worth reaching.

It's two tiers, not a ranked list

Between five and seven of the 14 openings tie at exactly 0 mean eval gap at any given rating, depending on the band. At 1000 rating it is seven: Ruy Lopez, Sicilian Defence (Open), Caro-Kann, Queen's Gambit Declined, King's Indian Defence, Vienna Game and Englund Gambit. By 1400 rating it narrows to five: Sicilian Defence (Open), Caro-Kann, Queen's Gambit Declined, King's Indian Defence and Englund Gambit, and that group of five stays intact, with membership shuffling slightly, through 2200. That is a two-tier structure, a group where the typical player's move already matches the engine and a tail where it does not, not a clean ordering from best to worst. Presenting this as a numbered leaderboard would claim a precision the data does not have; the order within a tied group of zeros is arbitrary.

Tier membership also moves faster than you might expect between adjacent bands. Vienna Game sits at 0 only at 1000 rating and jumps to a 3.0 centipawn gap by 1200, where it stays through 2200. Queen's Gambit Accepted swings hardest of anything in the set: a 51.5 centipawn mean gap at 1000 rating collapses to 2.5 by 1200, a single 200 point jump, because the likeliest human move in both of its tested positions shifts to match Stockfish's preference almost as soon as rating moves up. Below is every opening's ECO code and the exact line we tested, grouped by side, followed by the full mean eval gap table across all seven bands.

Mean eval gap in centipawns, averaged across each opening's early and late position, lower means more findable. Rows are grouped by side and kept in the same order at every rating rather than re-sorted by score, deliberately, so this table does not read as a leaderboard.

Scotch Game's mean gap comes almost entirely from its second position, and that split is worth naming directly. The early recapture exd4 is close to automatic already at 1000 rating, 48.5 percent findability, rising to 95.6 percent by 2200. The position two plies later is where the gap actually sits: a 55.0 centipawn gap running from 1000 through 1800 rating, because the likeliest human move there is the knight recapture Nxd4 while Stockfish already prefers the quieter Nf6, narrowing to 11.0 centipawns only once rating reaches 2000. A single averaged number per opening, without that early versus late breakdown, would have hidden exactly where the difficulty lives.

Findability does not rise cleanly with rating

It would be convenient to say stronger players find the objectively best move more often, in a straight line, and mostly that held: mean monotonicity across all 28 positions in this dataset was 0.690, meaning 69.0 percent of consecutive rating-band steps saw findability hold steady or rise. It was not 1.0, and the exceptions are not noise.

Both series run backwards for most of the range. In the Ruy Lopez early position, Stockfish's top choice is Nf6, and it is also the likeliest human move at 1000 and 1200 rating, hence the 0 eval gap at those two bands in the table above. From 1400 rating onward the likeliest human move stops being Nf6 at all: it becomes d6 at 1400 and 1600, then a6 from 1800 rating up, the move Wikipedia's own Ruy Lopez article says is played in roughly two-thirds of all Ruy Lopez games at any level. Findability of Nf6 specifically falls because stronger players are increasingly choosing a different, extremely well studied and nearly equally good move instead, not because they are getting worse at chess. The eval gap tells the same story in smaller print: 48 centipawns at 1400 to 1600 when players are still transitioning through d6, down to just 6 centipawns once a6 takes over from 1800 rating on, essentially a photo finish between two sound continuations.

London System's late position runs the same way for a similar reason. Stockfish's top move, Nf6, is also the likeliest human move at 1000 through 1400 rating, a 0 eval gap. From 1600 rating on, the likeliest move becomes Bf5 instead, a real developing move that carries only a 10 centipawn cost. Rating is not failing here. Players are converging on established theory that happens to sit slightly below the engine's single top pick in this exact position, which is a completely different phenomenon from a weaker player simply missing the best move.

The real finding: the gap between openings shrinks as rating climbs

The worst-scoring opening in the set changes hands three times as rating climbs. French Defence is worst through 1600 rating, at a 53.5 centipawn mean gap, which matches its own reputation for a cramped early position and a blocked light-squared bishop. Once French's gap collapses to 9.0 at 1800 rating, Scotch Game becomes the worst of the 14 at 27.5, purely because Scotch's own late position problem, the Nxd4 versus Nf6 gap described above, has not resolved yet at that band. By 2000 rating, Scotch has resolved too, and London System, sitting at a comparatively modest 10.0 the whole time, becomes the nominal worst simply because everything else has closed in around it.

Treat the compression as real but be careful with the causal story. Part of why the gap at 1000 rating looks larger than the gap at 2200 is simply that weaker players play less accurately everywhere on the board, not only in these specific opening positions; that is a general fact about rating, not a fact specific to openings. The defensible comparison is the spread between openings within a single rating band, which really does shrink from 53.5 centipawns at 1000 to 10.0 at 2200, not the raw size of any single number compared across bands.

This is where the coaching disagreement from the top of this page actually resolves, at least against our own data. At 2200 rating, the practical cost of choosing among these 14 openings is small, 10.0 centipawns at most, which leans toward Finegold's stream-persona position that opening choice barely matters, at least once a player is well clear of 1500. At 1000 rating the same choice carries up to 53.5 centipawns of real cost, which leans the other way, closer to Naroditsky's earlier and more insistent framing. Bartholomew's own tiered guidance, minimal below 1400 and a real repertoire from 1400 to 2000, sits almost exactly where our data crosses from a wide spread to a narrow one. None of this proves any one coach right in general; it is one dataset, on 14 openings, agreeing more with different people at different ratings.

For White, for Black: what the top tier actually contains

At 1000 rating, two of the six White openings we tested land in the zero gap tier, Ruy Lopez and Vienna Game, against five of the eight Black defenses, Sicilian Defence (Open), Caro-Kann, Queen's Gambit Declined, King's Indian Defence and Englund Gambit. By 2200 rating the split is starker: none of the six White openings in this set sit at a zero mean gap any longer, while the same five Black defenses still do, joined by Scandinavian Defence once its own gap closes at 1800 rating.

We would not lean hard on this as a general rule about White versus Black. It could reflect White's extra opening tempo giving Stockfish more nearly equal top choices to weigh against whatever a human actually plays in these specific lines, or it could just as easily be an artifact of which single continuation we picked to represent each of the six White openings. Two positions per opening is enough to notice a pattern like this, not enough to settle why it happens.

What existing opening pages already give you

It would be inaccurate to say nobody gives rating-aware guidance at all, and we are not going to claim that. Lichess's own /opening page already filters by rating from 1000 to 2500 and by speed, and shows the real popularity of each opening at the selected band, meaning it already answers what players at a given rating actually choose to play. Duolingo's chess blog goes further than most competitors with a single explicit line: it recommends players around 1000 to 1200 Elo study "one or two openings for each color," though it backs that recommendation with no win rate or difficulty data. Chess.com shows real win, draw and loss percentages for some individual sub-lines; its own Italian Game page states the main Giuoco Piano line ends "Black wins 31%, draws 33%, and loses 36%," a real number, with no rating filter applied to it. A genuine opening explorer with its own win and draw percentages exists at 365chess.com too, though its deeper per-opening pages are JavaScript rendered and returned 403 to a direct fetch when we checked, and it has no single interface for comparing results across rating bands.

The gap that is actually missing, and the one this dataset is built to close, is narrower than "no data exists." Nobody we found combines a rating filter with a genuine practical-findability judgment: is the objectively best continuation in a given opening one a player at a specific rating can actually find, not just one that gets played, or one that scores well in aggregate across every rating mixed together.

What Chessdrive does with openings

Chessdrive, the app we build, is the interested party in this section, and the dataset above came out of our own pipeline, not somebody else's paper, so read both accordingly. In the app, maia3-5m runs on-device as your opponent at a single continuous rating you set from 600 to 2600, and Stockfish runs separately, after a game closes, to review what happened rather than during play. Chessdrive also generates a starting opening repertoire for both colors from a short taste questionnaire, rather than handing every player the same fixed list regardless of what they actually enjoy playing.

What it does not do: there is no public, browsable opening explorer inside Chessdrive the way Lichess or 365chess offer one, and this specific per-opening findability breakdown is a piece of research we are publishing here, not a live in-app feature you can currently filter by rating yourself. Chessdrive is live on iOS and Android, priced at $5.99 a month or $39.99 a year on the US storefront, with a free trial to start on either plan. Store prices vary by region, so check your own storefront.

Limits of this data

Two positions per opening, one early and one late, is a real dataset but a small sample of any single opening's full character. Sicilian Defence (Open) alone covers roughly B20 through B99 in the ECO system, dozens of genuinely different structures depending on Black's fifth move choice; we tested one specific line within it, not the whole family. Treat every number here as describing the two positions we actually scored, not the opening as a philosophical whole.

Findability measures Maia3's model of how a player at a given rating would move, built from training on real human games, not a live survey of real players making real decisions at a real board. The two are correlated, that is the entire premise of the Maia project, but they are not identical, and Maia3's own published accuracy against held-out human games tops out at 57.1 percent, well short of certainty. The reference rating of 1400, used to pick each opening's "most popular" two-ply continuation, is a disclosed choice rather than the only defensible one; a different reference rating could have picked a different line for a handful of these openings, and the article above would read somewhat differently as a result.

Published 29 August 2026. Last verified: 29 August 2026, against the sources cited above and against Chessdrive's own generated Stockfish 18 and Maia3 output described in this article. Coaching advice, forum sentiment and third party opening pages can change; confirm current details on each source's own page before relying on them.