Best Chess Moves: What a Calculator Won't Tell You

Two different people type "best chess moves" into a search bar, and they want two different things. One wants a tool: paste in a position, get the move, right now. The other wants a list: which openings are actually good, ranked by something more than a stranger's opinion. This page answers both, in that order, then adds the one thing neither a calculator nor a listicle can produce, because it needs a dataset the calculators and listicles were never built to hold.

The calculator question, answered honestly first

If a live game is open in another tab, stop here before opening a second one. Chess.com's Fair Play policy, as read on 2026-09-07, is unambiguous: "Do not use chess engines, software of any kind, bots, plugins, browser extensions, or any tools that analyze positions during play." The same policy closes a narrower loophole too, banning lookup tools even outside live play: "Do not use tablebases or any other resources that show the best move (in both Online and Daily chess)." The site states it can detect this, and the consequences are not a warning: a violation can lead to account closure, a public label as a Fair Play violator, a lifetime ban, and disclosure "to relevant authorities, other websites, or chess governing bodies," with the platform reserving discretion to take "any action" it judges appropriate. None of that applies to a finished game. Running a calculator once a game has ended is the entire point of post-game review, and it is what the rest of this section is for.

Three tools show up repeatedly under this exact search, and they are worth naming plainly rather than dancing around. Lichess's analysis board costs nothing and never has a paid tier to upgrade into: its own Patron page, checked 2026-09-07, states, "Lichess is entirely free, forever, and for everyone. That's a promise," and confirms directly that no feature is reserved for paying users, only cosmetic profile flair. Chessigma's Next Move tool, as of the same date, runs Stockfish 17 entirely inside the browser via WebAssembly, sending no data to a server, described on its own page as "Completely free, with no move limits and no account required," analyzing to a depth of 20 or higher depending on the visitor's device and returning both a move and an evaluation such as +0.4 or -2.5. Nextchessmove.com markets itself as "the strongest online chess calculator," letting a visitor drag pieces into a position and click to get a suggestion, but its landing page does not say which engine sits behind it or what it costs; that is a gap this research could not close, not a hidden feature worth guessing at.

None of the three explain why the move handed back is not automatically right for a human to play against another human, which is the question the rest of this page is for. For the mechanics of either board rather than a repeat here, see Lichess's free Stockfish analysis board and how a chess analysis board works. Post-game review also sits inside Chess.com's own premium tiers, at a cost, with the tiers as published on 2026-09-07: free Basic gives "1 full game review analysis per day," Platinum adds "UNLIMITED game review and analysis," and Diamond adds coach explanations on every move.

The opening question, and why the ranking lists dodge it

Search "best opening moves in chess" and the results are not short on confidence about tone, only on data. World Chess Shop recommends five openings for teaching value, calling them "real chess, not just tricks or traps," with no engine score or database citation on the page. House of Staunton runs a ten-opening list using only qualitative labels like "theoretically sharp" and "solid," again with no evaluation number or win-rate figure anywhere. Chess.com's own flagship article on the topic, the highest-authority page in this space, recommends five openings and declines to pick a winner among them, closing instead with a joke: "the best opening of all is 'The Control-the-Center-with-Active-Pieces-and-Play-Carefully Opening!'" Three publishers, three lists, zero engine evaluations, zero master-game frequencies, and one flat refusal to answer the question the article's own title asks. That refusal is close to the right instinct stated badly: there genuinely isn't one best opening move, and a hedge is more honest than an invented number. Where those pages fall short is that none of them show any actual measurement to hedge from.

We have one, though it needs an honest boundary drawn around it first. Lichess's public opening explorer, which reports what percentage of real games at each rating actually open with 1.e4 versus 1.d4 versus anything else, was unreachable while this article was researched: every attempt to query it, across three separate methods, returned an authorization error rather than data. So this page will not print a master-game frequency for any first move, because none was actually verified this week. What we do have is narrower and, for a player trying to improve rather than trying to win a trivia question, arguably more useful: a measurement of which specific opening positions are close calls and which are lopsided, and which ones a player at a given rating tends to actually find over the board.

That measurement comes from a Stockfish 18 run frozen on 2026-08-29 and re-read for this article on 2026-09-07, nothing recomputed by an engine in between, just the same figures read back out. Fourteen openings, two positions each, one early and one two plies deeper, 28 positions total. At each one, Stockfish scored every single legal move at depth 20, and Maia3, the human move prediction model, at seven rating bands from 1000 to 2200 supplied the probability a player at that rating would actually play each option. Every position's board was independently rebuilt two separate ways with matching results, and every probability distribution summed to within 0.99 and 1.01 of 1.0, as a distribution should; the file this article draws from ran nine of its own validation checks, and all nine passed.

The first finding cuts against the idea that any opening move is simply correct: across all 28 positions, the gap between Stockfish's best move and its second-best move has a median of just 9.5 centipawns, barely a tenth of a pawn, though the average is pulled up to 31.39 by one outlier. Seven of the 28 positions sit within 5 centipawns of a tie. Fourteen, exactly half, sit within 10. Twenty-one of 28 are within 20 centipawns, a gap most players would never notice across the board. Only one position out of 28 crosses 100 centipawns, a full pawn of difference between the best move and the next best.

Some of the closest calls: in the Queen's Gambit Accepted's later position, e6 and a6 score exactly the same, -0.26 apiece, a true tie. In the Italian Game's early position, Nf6 at -0.21 beats Bc5 at -0.22 by a single centipawn. The Italian's later position repeats the pattern almost exactly, d6 over a6 by one centipawn. The Scandinavian's later position has Nf6 over e6 by one centipawn too. The Queen's Gambit Declined's early position, Nf6 over Bb4, is the widest of this tight group and it is still only 5 centipawns.

The wide gaps tell a different, more useful story, because in this dataset they are not close judgment calls, they are outright mistakes. The single widest margin in the entire set, 424 centipawns, sits in the Ruy Lopez's later position, after 1.e4 e5 2.Nf3 Nc6 3.Bb5 a6 4.Bxc6: bxc6 scores -0.32 while the alternative, Bd7, scores -4.56, because Bd7 there simply loses material outright. It is the only position in the whole dataset over a full pawn, and it is not a stylistic preference, it is a blunder waiting to happen. The next widest are the Englund Gambit's early position, dxe5 at +1.38 over Nc3 at +0.39 by 99 centipawns; the Scotch Game's early position, exd4 over Nxd4 by 48; the Vienna Game's later position, Nf6 over Bc5 by 47; and the Englund Gambit's later position, Nf3 over Bf4 by 44.

Put those two lists side by side and a rough rule appears: a genuine tie means the "best" label is mostly cosmetic, since playing either near-tied option loses essentially nothing, while a wide gap usually marks a real error, like handing over a piece, not a subtler stylistic preference. Neither the calculators nor the listicles distinguish between these two cases, because neither measures a margin at all; they either name one move with no context, or avoid naming one entirely.

Eleven of the dataset's 14 openings appear above by name, and every one of the 14 has its own page with the full per-rating breakdown, board diagrams and the exact move sequence behind each number: browse them all from the best opening moves ranked by rating, read the Caro-Kann Defense crossover at 1800 for the one opening whose deeper position runs backwards with rating, or go straight to the Italian Game, the Ruy Lopez, the Sicilian Defence, the French Defence, the Caro-Kann, the Queen's Gambit Declined or the London System, or browse the full set at the openings hub. We do not reprint their per-opening findability tables here.

What "best" actually means to an engine

Both halves of this search assume "best move" is a single, stable idea. It isn't. Stockfish's own FAQ defines search depth as "how far a chess engine searches in half-moves, or plies," and cautions that a reported depth "is a useful guide but not a strict guarantee that you've examined every move to that ply," because the engine prunes some branches, extends others, and reduces the rest unevenly rather than exploring every line equally. That is the documented reason two engines, or the same engine at two different depths, can legitimately disagree about which move is best in an identical position. "Best" is not a fixed property of the position; it is the engine's current, depth-limited estimate, and a deeper search can move that estimate.

The number attached to that estimate is also less literal than it looks. The familiar evaluation bar reading +0.4 or -2.5 is not a simple count of material or a simple point of position quality. Per the same FAQ, modern Stockfish's evaluation is normalized to a win-probability model, describing "a 1.0 pawn advantage being a 0.5 (that is 50%) win probability." A +1.0 does not mean "one extra pawn's worth of advantage" in any literal sense; it corresponds to roughly a 50 percent win probability under the engine's own model, a number about outcomes, not material. Chess.com's own move labels layer a further wrinkle on top: what "best," "great" and "brilliant" actually mean there, and why players report them behaving inconsistently, is covered in the FAQ below.

What our own numbers say about the gap between an engine's best move and a human's

Here is the finding this page exists to report, and it runs against the instinct that a higher rating should mean agreeing with the engine more often. Across the same 28 positions, the share of the time a player's single most likely move matches Stockfish's outright top choice actually falls as rating climbs: 19 of 28 positions at 1000 rating, still 19 at 1200, down to 17 at 1400, and settling at 16 of 28 from 1600 rating all the way through 2200. A 2200 rated player's likeliest move agrees with the engine's single best move less often than a 1000 rated player's does.

Read that number alone and it sounds like stronger players are getting worse at finding the best move. They aren't. What actually improves with rating is something closer to damage control. When a player's likeliest move is not the engine's literal top choice, the average cost of that disagreement collapses as rating rises: 39.56 centipawns at 1000, 32.78 at 1200, 31.27 at 1400, 27.83 at 1600, then a sharp drop to 11.08 at 1800, 7.42 at 2000, and 7.42 again at 2200. And the number of positions where the likeliest move at a given rating actually costs something real, 50 centipawns or more, shrinks from 3 out of 28 at 1000, 1200 and 1400, down to 3 again at 1600, then 1 at 1800, and 0 at both 2000 and 2200. No band in the dataset ever reaches a full pawn of cost, 100 centipawns, at any rating.

A related number tells the same story from a different angle. The share of positions where a player's likeliest move lands somewhere in the engine's top three choices, not necessarily its single favorite, climbs steadily: 22 of 28 at 1000, 24 at 1200, 23 at 1400, 24 at 1600, 27 at 1800, and a full 28 of 28, every single position, at both 2000 and 2200. By 2000 rating, the likeliest human move in this dataset is never outside the engine's own top three options. Mean findability, the probability Maia3 assigns to the exact move Stockfish calls best, rises in the same steady direction: 0.3523 at 1000, 0.3835 at 1200, 0.4043 at 1400, 0.4225 at 1600, 0.4421 at 1800, 0.4678 at 2000, and 0.4995 at 2200. Put the pieces together and the shape is not "stronger players find the single best move more often." It's "stronger players stop playing the bad ones." A 1000 rated player and a 2200 rated player disagree with the engine's exact top pick at close to the same rate. What separates them is that the 1000's disagreements are real mistakes and the 2200's are near misses.

Two positions carry that pattern in miniature. In the French Defence's early position, after 1.e4 e6 2.d4 d5 3.Nc3, the likeliest move at 1000 rating is c5, costing 96 centipawns against the engine's actual best, dxe4, the single largest gap anywhere in the dataset at that band. By 2200, the likeliest move in that same position is Nf6, costing only 7. In the Scandinavian, after 1.e4 d5 2.exd5 Qxd5 3.Nc3, the engine's best move is Qa5 at -0.63; at 1000 rating the likeliest move is instead Qe6+, costing 46 centipawns, and findability for the correct move Qa5 is just 0.1771. By 2200, Qa5 is itself the likeliest move, and its findability has risen to 0.5347.

Not every opening moves this cleanly. In the Ruy Lopez's later position, the one with the 424 centipawn gap above, the correct move bxc6 is already the likeliest human move at every band we tested, findability starting at 0.9711 at 1000 and climbing to 0.9912 at 2200: when an alternative is a flat-out blunder, players at every level tend to avoid it. Fourteen of the 28 positions show the engine's best move as the likeliest human choice at every band from 1000 to 2200. Seven never show that agreement at any band, including the Italian Game's later position, the Scotch Game's later position, both French Defence positions, the Queen's Gambit Accepted's early position, the London System's early position, and the King's Gambit's later position. The remaining seven flip somewhere in between.

The strongest players in this dataset are not the ones who best divine the engine's exact preference. They are the ones who have narrowed their own list of realistic candidate moves down to the ones that don't cost anything, whether or not that happens to be the engine's literal favorite. That lines up with something a chess writer said a century before any of these engines existed. Eugene Znosko-Borovsky, quoted in a Chess.com article on chess quotes, put it this way: "It is not a move, even the best move, that you must seek, but a realizable plan." A Chess.com forum poster named bigbadsquid arrived at close to the same idea from the opposite direction, watching an engine's number rather than reading old books, in a thread asking whether the best move is always the best move: "When the position is either winning...or losing, then the best move suggested by an engine is often not the best practical move against a human... As a human you want to do the opposite: simplify when winning and complicate when losing." Another poster in the same thread, woton, made the narrower version of the same point: "Sometimes, the computer calculated (cc) best move is not the best move for lower rated players."

How to actually find a good move without a calculator

None of the above argues against thinking clearly, only against outsourcing it to a tool you cannot legally use mid game anyway. The old, boring process still works, because it maps onto what strong players in this dataset are actually doing: narrowing candidates rather than divining one perfect answer. List two or three candidate moves before calculating any of them in depth. Check every capture, check and threat for both sides, since missing one of those is where the real cost above comes from, not from picking the second-best of two nearly equal quiet moves. Ask what plan a move serves, not just whether it's safe this instant; a move that loses nothing but connects to no follow-up idea is still a wasted move. Our companion piece on what engine analysis can't tell you goes through this process in more depth.

Where Chessdrive fits

We build a chess app, so read this paragraph as the interested party talking about its own product, not as a neutral recommendation. Chessdrive is not a best-move calculator and does not try to be one. In the app, the real maia3-5m model runs on-device as your opponent, set at a single rating you choose anywhere from 600 to 2600, so practice happens against something closer to a human at that level than a strength-limited version of a perfect engine. Stockfish reviews the finished game separately, after it ends, the same post-game use case described above rather than a live-game aid. A short questionnaire also generates a starting opening repertoire for both colors, rather than asking a new player to pick blind from a list like the ones described earlier in this article. It 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; prices vary by region.