Play Chess Against Computer: How to Actually Choose

The honest answer to play chess against computer is three different products, not one: Lichess's free numbered levels, Chess.com's roster of over a hundred named bots, or a packaged app such as Chessdrive that trains its opponent on real human games instead of capping an engine's search depth. Which one is actually right for you depends on what you want out of the game, and that is what this page is built to answer, rather than adding a fourth vague list of engines to the pile. It also covers something almost nothing ranking for this term explains: what a "difficulty" setting is actually doing to the engine underneath it, and why that mechanism is the reason a computer opponent so often plays five good moves and then hangs a piece for no reason.

Where to actually play chess against a computer today

Three real options exist right now, and each is worth knowing about before you pick.

• Lichess.org: free, with no account needed to play a person or a computer, and no paid tier anywhere on the site at all; it is a non-profit funded by donations. The numbered Levels 1 to 8 are Lichess's own built-in strength presets. Lichess also runs free official Maia-1 bots, BOT maia1, BOT maia5 and BOT maia9, trained to predict what a human would actually play rather than search for the objectively best move, but they sit outside the main "Play with the Computer" screen entirely. You have to already know to search for them by name, which is a real discoverability gap even though the bots themselves cost nothing.
• Chess.com: also free to open, with a roster of more than 100 named bot personalities (Martin, Nelson, Isabel, Aaron, Mittens at roughly 3200 Elo, plus celebrity tie-ins like MrBeast and Hikaru Nakamura), all running on the Komodo engine. Some are free to Basic members; others carry a lock icon and require a paid Premium membership. Games against any bot are always unrated.
• Chessdrive: a downloaded app rather than a browser tab, live on iOS and Android. Instead of picking a bot from a menu, you set one continuous rating from 600 to 2600 and the app runs the real maia3-5m checkpoint on-device as your opponent at that rating. Stockfish runs separately, after the game closes, to review what happened. Chessdrive is the app we build, so treat everything it says about itself here as the interested party's own claim.

Said plainly, before anything else on this page: if the goal is simply the widest possible choice of free opponents, Lichess and Chess.com both beat Chessdrive on raw selection. Lichess wins on price and honesty, since none of it is behind a paywall. Chess.com wins on roster size and named personalities. Chessdrive's claim is narrower, and the rest of this page explains exactly what that narrower claim actually is.

If what you actually want is to play chess vs computer online with nothing installed at all, Lichess and Chess.com both open directly in a browser tab and cost nothing to try. If what you want is chess computer free with zero catches anywhere in the product, that description fits Lichess specifically, not Chess.com's full roster and not Chessdrive past the trial.

What a difficulty setting actually changes

Every one of these products answers the same underlying question: how do you make an engine built to find the objectively best move play worse on purpose. The mechanism is rarely what people assume, and it is documented if you go looking for it.

Stockfish itself ships two separate ways to do this, both configurable UCI options rather than one single dial. Skill Level is a spin control, default 20, running 0 to 20. Turning it down does not simply cut search depth: it enables MultiPV internally and, with a probability tied to the Skill Level value, has the engine choose a weaker move instead of its top choice. Separately, UCI_LimitStrength is a boolean, default false; switch it on and Stockfish targets a specific Elo set by UCI_Elo instead, a spin option running 1320 to 3190. That Elo is calibrated against CCRL 40/4 at a 120 second plus 1 second time control, not against Lichess's or Chess.com's own rating pools. A Stockfish set to "1500" this way is not the same 1500 as a 1500-rated Lichess account; the two numbers come from different measuring sticks entirely. We cover Stockfish's full feature set, its license and what its evaluation score does and doesn't tell you in a separate reference on what Stockfish actually is.

Lichess's own numbered Levels 1 to 8 are not one lever, they are three, capped together at once: a Skill Level value, a search depth cap, and a movetime cap. Fairy-Stockfish maintainer Fabian Fichter, quoting the underlying fishnet worker source directly, confirmed that Level 8 runs Skill Level 20, the maximum, depth 22, and a movetime of 1000 milliseconds, effectively full strength. Fichter also cautioned, writing on the computer chess forum TalkChess, that any Elo number attached to a given level is "just ballpark numbers for orientation," not an officially normalized rating. Reading the fishnet worker source directly settles what the lower levels actually do, and one detail there is worth correcting because it is widely misstated: the depth cap does not keep falling as you go down. It sits at a constant depth 5 across Levels 1 through 5. What actually changes at the bottom of the range is the skill setting, dropping to roughly negative nine, and the movetime, as short as 50 milliseconds. So a Level 1 opponent is not thinking one move ahead. It is searching five plies deep and then deliberately picking a worse move than the one it found, which is precisely the mechanism that makes weak engine play feel unnatural rather than merely weak.

Chess.com does not run one dial at all. A chess.com staff-flaired forum account, Martin_Stahl, confirmed directly that "the bots are based on Komodo with various setting changes and custom opening books," and the site's own help center states plainly that "all bots, as well as practice vs computer, are powered by Komodo." Each named bot is closer to an authored personality than an engine reused at every level. Chess.com's documentation about which Stockfish version handles what, separate from the bots, is also honestly a mess worth naming rather than smoothing over: one help article states "Analysis uses Stockfish 15 by default," while a separate July 2026 forum thread describes Game Review running Stockfish 16 on chess.com's own servers for stability, with the live Analysis board pulling newer builds, including Stockfish 18, client-side. Chess.com's own pages disagree with each other about which version does what. That is chess.com's own copy contradicting itself, not a claim we are making about them.

Why a strength-limited engine feels wrong to play

Capping skill level or search depth produces a specific, recognizable failure shape: the engine plays several accurate, book-shaped moves and then, without warning, hangs a piece or misses something a much weaker player would have felt was there. Players describe this in strikingly similar language across different sites and different bots. "They play at a grandmaster level for a few moves and then boom, leave your queen hanging," one puts it. Another: "A bot will do things like move their king for no reason at all." A third, more cynically but not inaccurately: "Most humans don't enjoy being beaten in every game, so they had to throw in a few lines of code to make the bots randomly drop pieces."

CSSLab, the University of Toronto lab behind the Maia research project, measured this rather than just describing it. Their own writeup states it plainly: "each curve is strictly increasing, meaning that even depth-1 Stockfish does a better job at matching 1900-rated human moves than it does matching 1100-rated human moves." An artificially weakened engine's play, in other words, is closer to full-strength Stockfish with noise mixed in than it is to how a genuinely weaker human actually thinks through a position. That is the mechanism behind the forum complaints above, and it is not that these products are careless: capping search depth or skill was never built to reproduce a human error pattern in the first place. We cover that research in full, including how a human-trained model fixes it, in our companion piece on why chess bots play five good moves and then hang a queen, and on the model behind the fix in our own reference on Maia chess; neither argument is repeated here.

The strongest evidence for this is not a forum thread, it is a controlled experiment. Eisma, Koerts and de Winter, publishing "Turing tests in chess: An experiment revealing the role of human subjectivity" in Computers in Human Behavior Reports, volume 16, article 100496, 2024, had 24 chess players play blitz games against a human, against Maia, against a downgraded Stockfish, and against full-strength Stockfish, then guess who or what they had just played. Full-strength Stockfish was usually correctly identified as an engine. Maia was often mistaken for a human. The downgraded Stockfish fared worst of all: the paper reports that "the moves of the downgraded Stockfish were relatively often labeled as strange by the participants." The authors' own framing of what a chess Turing test actually measures is worth quoting directly: it is "a test of whether the chess computer can devise suboptimal moves that correspond to human moves, and not necessarily a test of computer intelligence." On that evidence, a weakened search engine is bad at exactly that, even when the raw difficulty setting is tuned correctly.

Which one to actually pick

None of this makes one option correct for everyone. It changes what you should actually pick for, based on what you are optimizing for when you sit down to play chess against an AI opponent.

• Widest free choice, no strings attached: Lichess. It carries no paid tier at all, and once you know to look for them, its official Maia bots, BOT maia1, BOT maia5 and BOT maia9, are free human-move predictors sitting one search box away from the numbered levels. Nobody beats Lichess on price or on the breadth of what is genuinely free.
• Biggest roster of named personalities: Chess.com. A hundred-plus authored bots is more variety than either Lichess or Chessdrive offers, and the social features and player base around it are real advantages neither of the other two can match. The tradeoff is that most of that roster sits behind a paid membership.
• One continuous rating, human-shaped mistakes, and a review layer in the same flow: this is the narrower claim Chessdrive actually makes. Instead of a bot number or a named personality, you set a single Elo from 600 to 2600 and Chessdrive runs the real maia3-5m checkpoint on-device at that rating, so the opponent's mistakes come from training on real human games rather than from a search-depth cap. Stockfish reviews the finished game separately afterward. Chessdrive does not try to out-roster Chess.com or out-price Lichess: it costs $5.99 a month or $39.99 a year on the US storefront, with regional prices varying and a free trial to start, and it is an early product with a much smaller track record than either of the other two. If the biggest possible bot library is what you actually want, that is a real reason to pick something else instead.

If your priority is specifically playing free with the widest range of human-feeling opponents and you don't mind hunting for the right bot name, that is Lichess's Maia bots, not any paid product, ours included. We would rather say that plainly than pretend otherwise.

For the fuller picture of what "free" looks like across the whole space, beyond just computer opponents, our companion guide on how to play chess online free covers two-player games, offline play, and the rules themselves in one place.

Last verified: 29 August 2026, against the sources cited above. Engine versions, bot rosters and pricing tiers change without notice on products we don't control; confirm current details on each platform's own page before relying on them.