Lichess Analysis: What's Free, What's Capped, and What the Report Can't Score

Type "lichess analysis" into a search bar and the tool itself is already the top result. That's the odd thing about this particular search: the product being asked about is also the best answer to the question, so almost nothing gets written about it that isn't Lichess's own page. This piece tries to be the missing companion: a plain, sourced account of what the site's free analysis actually includes, the one place it draws a real line, and a specific explanation, backed by our own data rather than a borrowed opinion, of what its output still can't tell you about your own game.

What Lichess actually gives away, free

Start with credit where it's earned, because most comparisons skip straight to the complaint. Lichess is a registered French nonprofit under the 1901 association law, and its own blog states the funding model without hedging: "All operating costs, development, and content are funded solely by user donations." The same post makes a point of naming what it's choosing not to do: it could run ads or a subscription and make more money, but "we won't," because unlike a commercial competitor, "all of that income goes back into the organisation." That's not marketing copy lifted from a features page. It's a stated policy, and the site's finances are public at lichess.org/costs for anyone who wants to check the claim against the numbers rather than the press release.

What that funding model buys a free account is genuinely large. The analysis board itself, running Stockfish 16+ NNUE locally in the browser plus unlimited cloud engine evaluation, carries no cap at all, for a free account or a donating Patron alike. The Opening Explorer, both the global database and a personal one built from your own games, covers what Lichess's own features page describes as "6,000,000,000 games," also free and unlimited for both tiers. Studies, Lichess's shareable annotated boards, Chess Insights, a detailed breakdown of your own play by phase, opening and time control, and the Practice section are all listed the same way: unlimited, for everyone. Practice alone is a real curriculum rather than a single drill: seven Checkmates courses, eight Fundamental Tactics lessons covering pins, forks, skewers and discovered attacks, ten Advanced Tactics courses on ideas like zugzwang and deflection, three Pawn Endgame courses and four Rook Endgame courses, all playable without creating an account. Signing up only saves your progress through them.

None of that is a trial, a teaser, or a funnel toward a paid tier, because there is no paid tier to funnel toward. A Lichess Patron subscription exists, but per the site's own features page it buys a handful of cosmetic and identity extras, not more analysis. That's an unusual thing to be able to say about any software in 2026, and it's worth saying plainly before getting to the one place this generosity actually has an edge. For how to read what the board itself is showing once you open it, see our piece on how to read an analysis board.

The one real limit: 40 deep reports a day

Here's the part that's easy to get wrong, and worth stating precisely, because a lot of explainers, several earlier pieces on this site included, have gotten it wrong before: Lichess is not free of any quota anywhere. The specific thing that's capped is the deep, post game computer report, the "Learn from your mistakes" grade that assigns an accuracy percentage, an average centipawn loss figure, and inaccuracy, mistake and blunder labels to each of your moves, run server side on Stockfish 18. Lichess's own features page lists that report at 40 games a day, and, this is the detail that actually matters, the limit is identical for a free account and a donating Patron. Paying Lichess does not buy a bigger number.

A Lichess forum reply from a CM flaired account confirms the figure in plain language: "It's 40 a day currently. The number can be found here." Another poster in the same thread explains why the cap exists at all, and it reads as a reasonable trade off rather than a defensive one: "They have to enforce limits, because anything that's unlimited gets abused. And the analysis is server side analysis, which becomes very costly." Running Stockfish 18 to real depth on every submitted game, on a site whose own Android listing claims more than five million games played every day, is not free to Lichess even though it's free to you. Forty full reports a day, with no account required to use the live analysis board at all, is a generous line to draw for something that costs real server time per request.

It's also not a fixed number frozen in some old document. One forum poster tracking the figure over time wrote: "it keeps rising, few months back i remember it was only 30. also looks like they removed the weekly limit altogether," Read that as one attentive user's observation rather than an announced policy change, but it points the same way: this quota has been loosening rather than tightening, and the same poster believes a separate weekly cap that used to sit alongside the daily one is no longer being applied. Treat 40 as the number as of September 2026 rather than a permanent ceiling. It has moved before, and nothing here suggests it won't again.

Why calling this "raw Stockfish" undersells it

It would be easy, and wrong, to describe Lichess's report as a centipawn number stapled onto your game. The site's own accuracy page publishes the actual pipeline, and it's more considered than that. Centipawns get converted into a win percentage first, because a swing of 300 centipawns means something completely different depending on the position, and Lichess's own documentation makes exactly this point against itself: "Losing 300 centipawns in an equal position is a major blunder. But losing 300 centipawns when the game is already won or lost makes almost no difference and is largely irrelevant." The game level accuracy score isn't a plain average of that win percentage swing either. Lichess describes it as built from sliding windows and both a harmonic and an arithmetic mean, explicitly not the same thing as average move accuracy.

There's more processing sitting on top of the score itself. "Learn from your mistakes," a feature Lichess shipped back in December 2016 and still runs today, takes each flagged inaccuracy, mistake or blunder and lets you replay it one move at a time: your actual move shows in red, and instead of handing you the engine's answer immediately, it asks you to find a better one yourself first. It also checks candidate mistakes against a two million game masters database before flagging them, specifically so it doesn't scold you for playing known, sound opening theory. That's a real teaching layer, not a bare number, and it runs on every analyzed game on the site, unlimited, with no ads attached.

None of that is nothing, and a fair comparison has to say so. What none of it does, and this is where the real gap actually sits, is take your rating into account at any point in the pipeline, the specific limitation covered in full in our piece on what the engine can't tell you.

What real users report when the report doesn't match the game

Search Lichess's own forums for "0 blunders" or "high accuracy" and the same complaint shows up repeatedly, from players who trust the tool enough to be confused by it rather than dismiss it outright. One thread, started by a user called nicokin89, posted a 56 move game between what the poster called "both below average players" that Lichess scored at zero inaccuracies, zero mistakes, zero blunders and 9 average centipawn loss. In their own words: "This is nothing of a great game as we are both below average players but this is according to lichess engine a perfect game." They asked, half seriously, whether a score like that could be mistaken for cheating.

The explanation, from a reply by user ProfDrHack, is the useful part: "Lichess internally caps the evaluation score once you reach a high enough value. You spent two thirds of the game on this plateau, not losing any centipawns in this period, thus your average centipawn loss (acpl) is very low." Once a position is decisively won, there's nothing left to lose by the metric the score is built from, so a long stretch of merely adequate moves inside an already decided game reads back as flawless. A second, independent thread makes the identical point from the other side of the same coin. User DeiGratiaLudus posted a 96 percent accuracy game that "felt, in comparison to many others of mine, VERY easy," and asked whether the number reflected real skill or just a weak opponent. The reply from Testrider names the limitation directly: "Accuracy is not really a measure of move quality, just of how badly you gave away your chance to win the entire game." Both threads describe the same mechanism from opposite directions, and Lichess's own documentation half admits it already, in that same line about a 300 centipawn swing meaning almost nothing once a game is decided.

A second, smaller failure mode is just findability of the feature itself rather than a flaw in what it measures. Chess Insights, the free, unlimited, per player breakdown of your play by phase, opening and time control, is real and genuinely useful. It's also apparently easy enough to miss that Lichess's own forums carry multiple threads with titles as plain as "Where can I find the 'Chess Insights'?" That's a navigation problem, not a product one, but it means a chunk of the platform's most personalized free data quietly goes unused by people who'd want it if they knew where to look. The same friction shows up on mobile: an open GitHub issue on the official Lichess mobile app repository, titled "Provide an easier way to request computer analysis after playing a game," is a live, acknowledged admission that requesting analysis on the app can take more taps than it should, on top of everything the desktop board already does well.

Lichess versus chess.com: a real debate that isn't the one this article is making

A separate, genuinely unresolved argument runs through Lichess's own "lichess vs chesscom analysis" forum thread, worth naming precisely so it doesn't get folded into anything above. User petri999 makes the broad point first: "There is no objective definition of inaccuracy, mistake and blunder." User DStamateleios adds a specific technical wrinkle, that "chess.com runs in lower depths unless you manually change that in the settings of game review," and argues Lichess has become "a softer judge" once a position is already winning, essentially the same plateau effect described above, treated here as a grading philosophy rather than a formula quirk. User pet59 pushes back from the opposite direction, arguing that leniency is itself the problem: "If the analysis does not tell you there was a better move, you won't (well we amateurs at least) look at it and learn nothing."

That's a real disagreement about how strict a report should be, and it has nothing to do with rating awareness specifically. A stricter or looser threshold would still apply the exact same standard to a 900 and a 2200 alike. It's also not obviously in either platform's favor: chess.com's own free tier caps its game review at one report a day, well short of Lichess's 40, so the two aren't even trading strictness for generosity in a way that clearly favors one side. Worth knowing the debate exists. Worth not confusing it with the specific gap the rest of this piece is about.

The one thing no version of this report asks: your rating

Every piece of processing described above, the win percentage conversion, the volatility weighted accuracy score, the masters database check, the replay trainer, runs the same way regardless of who played the game. A 900 and a 2200 get scored by the identical formula, and a missed best move gets the identical label, inaccuracy, mistake or blunder, whether or not a player at that specific rating had any real chance of finding it. Findability by rating isn't a wrong answer Lichess gives. It's a question its pipeline never asks in the first place, and that's true of every free analysis tool this research turned up, not a knock unique to Lichess.

We tested how large that blind spot actually is using our own data, not a borrowed claim. Chessdrive built a dataset in August 2026 running Maia3, a model trained to predict what a human of a chosen rating would actually play rather than what an engine considers correct, across 14 named openings at seven self rated Elo bands, checked against Stockfish 18 at depth 20. We call the resulting number findability, the probability Maia3 assigns to the exact move Stockfish scores as best. Full detail on Maia3 itself, including the three generations and how it differs from an engine, is in our own Maia, the human move prediction bots reference. At the 1000 rating band, mean findability across the openings we tested ranges from 10.66 percent for the Queen's Gambit Accepted up to 62.65 percent for the Ruy Lopez, more than 50 percentage points apart between two openings at the exact same rating. A Lichess report has no way to reflect that difference: a missed best move in either opening gets the same word attached to it, even though one of those moves was, at that rating, far harder to find than the other.

Rating doesn't move every opening's findability in the same direction either, and the two extremes make the point better than an average could. The King's Gambit barely changes across the entire range we tested: 25.52 percent findability at 1000 rating, climbing to a high of 29.89 percent by 1600, then settling back to 26.15 percent at 2200, a net difference of under a single point across twelve hundred rating points of climbing strength. The Queen's Gambit Accepted does the opposite. It very nearly triples over the same range, from 10.66 percent at 1000 rating to 30.08 percent at 2200.

Both are genuine facts about the exact same pipeline: Maia3 against Stockfish 18 at depth 20, run across the full move history from the standard starting position, at seven rating bands from 1000 to 2200. A rating conditioned report would treat a missed King's Gambit best move very differently from a missed Queen's Gambit Accepted one, because the two openings tell you almost nothing and quite a lot, respectively, about how much stronger play should be expected to close that specific gap over time. A report that only knows centipawns, win percentage and a fixed label vocabulary has no way to draw that distinction, on Lichess or anywhere else this research found.

The same asymmetry shows up inside a single game, not only across openings. In the finishing sequence of the 1858 Opera Game, Morphy against the Duke of Brunswick and Count Isouard, scored in a separate dataset against Stockfish 17 at depth 20 rather than the Stockfish 18 run above, the engine finds a forced mate in two with Qb8+. At the 1000 rating band, Maia3 gives more probability to a different move, Qxe6+, a natural looking check that actually throws the mate away entirely, Stockfish scores it at -1.25, than to the mate itself: 19.36 percent for the losing move against 16.23 percent for the winning one. A report that labels Qxe6+ a mistake or a blunder is technically correct and rating blind at the same time. At 1000, that "mistake" was, in this exact position, the single most probable move a player at that level would actually make. For the fuller breakdown of that position, including how the balance shifts again by 2200 once a third, quieter move takes over most of the probability, see our piece on what Stockfish's evaluation actually means.

So what does this actually change

None of this argues for skipping Lichess's own report, which is worth using precisely because it's free, generous, and about as transparent as chess software gets. It argues for reading a "mistake" label as a statement about a position, not a statement about whether you personally should have found the better move. The two are related, but they're not the same claim, and only one of them is currently something any free analysis pipeline, Lichess's included, actually measures.

That's the specific piece Chessdrive's own review tries to add on top of, not instead of, an engine's judgment: scoring each flagged moment for whether a player at your own rating would plausibly have found the alternative, using the same Maia3 against Stockfish comparison behind the numbers above. It won't replace Lichess's Opening Explorer or its Insights page, both of which are already free and already good. It answers a narrower, different question: not just what the better move was, but whether missing it says something about this one game, or about a pattern your rating band tends to miss more broadly. For how chess.com, Lichess and Chessdrive each build that judgment differently, see our full comparison in how chess.com, lichess and Chessdrive review your games, and for what a rating number itself does and doesn't predict, see what your chess rating actually predicts.