Solitaire Win Rates: A Researched Database

Most of the solitaire sites that publish win-rate numbers pull those numbers out of the air. A lot of the figures that circulate online — 99% for FreeCell, 43% for Klondike, 20% for Spider — trace back to a single uncited blog post from the mid-2000s that someone else later paraphrased. The number propagates because nobody checks the receipt. We think that is backwards. Win rates are exactly the kind of claim that deserves a citation, a sample size, and an error bar, because the same game can produce wildly different win-rate figures depending on how it was measured.

This page is our answer to the citation problem. Every entry below records where the number came from, how many deals were examined to produce it, what methodology was used, and what we do and do not know. The research library is maintained by the Solitaire Stack Research Desk and updates as new solver runs complete or new academic work lands in our reading queue. If a row says "estimate," it means the figure is a community best-guess we have not yet reproduced in our own simulator — and we say so rather than rounding up into false confidence.

We publish this as a public, cross-referenced database because solitaire players deserve better than marketing-copy numbers. When a site tells you "85% of Spider 2-suit deals are winnable," you should be able to ask: says who, from how many deals, under what assumptions, and get an answer. The database below is our best attempt at providing that answer for fifteen of the most-played solitaire variants, and at being honest about the handful of games where our answer is still provisional.

A word on scope. This page is about solvability — the fraction of randomly-dealt games that a strong player, or a solver, can win. It is not the same thing as a player's personal win rate, which depends on their own sequencing decisions, whether they allow undo, and how patient they are with long branches. Most solitaire apps will happily tell you your personal win rate in their statistics screen; this page tells you the ceiling that personal rate is bumping into. The gap between the two is where strategy articles earn their keep, and it is the specific gap each individual game page on the network is about.

Every row in the database carries a methodology tag. The tag tells you how much weight to put on the number, because not all win-rate research is the same thing. Exhaustive analysis means a solver was run against a closed set of deals and the answer is a count, not a sample: for the 32,000 Microsoft-numbered FreeCell deals, for example, we know exactly which deals are winnable because every one has been solved. There is no error bar because there is no sampling — the answer is a fact about that specific set.

Monte Carlo means a solver was run against a large random sample of deals and the win rate is reported as a percentage with a confidence interval. A Monte Carlo figure with N=10,000 deals has a 95% confidence half-width of roughly 1 percentage point; at N=1,000,000 the half-width drops to about 0.1 points. Sample size matters because tight claims require many deals, and a number quoted without N is almost impossible to evaluate. When we run our own Monte Carlo sweeps we disclose the solver heuristic alongside the number, because the same game plays differently under a greedy strategy versus a lookahead.

Published research means a peer-reviewed or widely-cited academic paper generated the number — Bjarnason, Fern & Tadepalli's 2007 ICAPS paper on Klondike is the canonical example. Community data means the figure aggregates reported player statistics from large deployed clients (the original Microsoft Spider telemetry, modern solitaire apps, long-running forum threads) — useful but noisy. Estimate is the humility tag: we believe the figure is in the right neighbourhood, but we have not yet produced or found rigorous backing, and the row should be treated as provisional until a simulation or citation upgrades it.

Confidence intervals, when reported, use the standard 95% level. For exhaustive entries the interval is a deterministic range reflecting rounding precision, not statistical uncertainty; for Monte Carlo entries the interval reflects the binomial sampling distribution at the stated N. We prefer to publish a wide honest interval over a tight fake one. A claim of "winnable 92.3% of the time" with no N and no interval is not a finding — it is a vibe.

One more methodological note. Every solvability number in this database assumes unlimited undo and full-information play — the solver sees the whole layout, can back out of any move, and optimises without time pressure. That is almost never how a human plays. Human sessions are one-shot, time-boxed, and operate with the tableau hidden (for games where face-down cards exist) or simply held in working memory (for games where every card is visible from the start, like FreeCell). When you see a solvability figure of 82% for Klondike and a reported player win rate of 35%, the gap is not a mistake — it is the specific cost of playing the game as a human, under real constraints. We disclose this so readers do not walk away thinking that personal win rates two-thirds lower than the solvability ceiling mean they are bad at solitaire. They almost certainly do not.

The table below is the canonical view of every game we have a cited win-rate figure for. Rows are sorted high-to-low by win rate; the methodology column tells you which class of evidence supports the figure. Click into any game page for the full treatment of rules, strategy, and the history behind the number — this page is the index, not the essay.

Two things to watch for. First, compare the win rate column to the sample size: a 90% figure at N=1,000 is much weaker than a 90% figure at N=1,000,000, even though they read the same on the page. Second, remember that these are solvability numbers under strong play. Human win rates — the rate at which an average player actually wins — sit meaningfully lower for every game on this list, because humans cannot see the face-down cards or compute every branch. The gap between the solvable ceiling and the human floor is where strategy articles earn their keep.

The dataset is also published under Creative Commons Attribution 4.0. If you want to cite these numbers in your own work, please do — link back to this page and the primary sources listed in each row, and we will happily help you interpret the methodology.

Dataset generated 2026-04-05 · 17 entries · methodology mix: 2 exhaustive, 3 published research, 9 community data, 3 estimate.

A common mistake when reading the database is to equate low win rate with high difficulty. The two are related but not the same. Clock Solitaire sits near a 1% win rate and is trivial to play — no player decisions exist, so the game is entirely a lottery. FreeCell sits near 100% solvability and is genuinely hard to play well, because the work of winning the 99.99% of deals that are solvable requires careful sequencing. Difficulty lives in the gap between what the game is willing to hand you and what your own sequencing can unlock.

The cleaner mental model is to decompose each game into a luck ceiling and a skill gradient. The luck ceiling is the solvability figure in the database — the fraction of deals that a perfect player could win. The skill gradient is the slope of player win rates as playing strength improves: for FreeCell the gradient is steep (novice win rate ~60%, expert win rate ~99.99%), while for Clock the gradient is flat because no skill applies. When we rank games for difficulty, we weight the gradient heavily; the luck ceiling just tells you whether the destination is reachable at all.

This is also why averaging win rates across difficulties is misleading. Spider's one-suit / two-suit / four-suit win rates are three separate curves; a headline "Spider wins ~50% of the time" hides the fact that nobody actually plays the average. Each difficulty level gets its own row in the database, and when a game has a redeal toggle that materially moves the number (Pyramid, Canfield) we note the assumed ruleset in the source citation. The same logic applies to FreeCell's restricted-cell variants: 4-cell classic, 3-cell, 2-cell, and 1-cell FreeCell are effectively four different games, and collapsing them into a single "FreeCell" row would hide exactly the information a reader is trying to use when they pick a difficulty.

The game pages themselves go deeper on this. Each individual variant page on the network carries its own skill-vs-luck breakdown, a tactical checklist for the specific sequencing problems that game poses, and an honest assessment of how much of the session outcome the player actually controls. The database on this page is the index into those essays: read a row, follow the link, and you get the full treatment. The database exists to stop readers from having to hunt for the citation, not to replace the strategy writing itself.

FreeCell is the rare solitaire game with a clean exhaustive result. Starting in 1994, Don Woods and Michael Keller ran solvers against the 32,000 deals numbered by the original Microsoft Windows FreeCell client; the effort became known as the Internet FreeCell Project. After years of community runs, the result was unambiguous: only deal #11,982 is unsolvable under standard rules with unlimited undo. One deal out of 32,000 means the solvability rate on that closed set is 99.996875%. This is not a sample — it is a count, and it is why the FreeCell row carries the exhaustive tag.

The broader "99.9987% of random deals are solvable" figure that circulates for FreeCell is a separate Monte-Carlo result, derived from fc-solve sweeps on random shuffles outside the Microsoft numbering. The two numbers agree at the 99.99% level because the Microsoft deals are a reasonably unbiased sample, but they answer different questions: the exhaustive count tells you about a specific 32,000-deal library, and the random-sweep figure tells you about the space of all possible shuffles. We report both in the database and explain the distinction on the FreeCell solvability page.

What does 99.9987% mean for a player? Practically, that you should never blame the deal. If you lose a random FreeCell game, the near-certainty is that the deal was winnable and your sequencing missed the path. The solvable ceiling is also the reason FreeCell has never had a mainstream "too hard" reputation — the game rewards careful play and almost always has an answer, which makes it feel fair even when a specific deal takes an hour.

There are a few other solitaire variants that admit exhaustive or near-exhaustive analysis: small-footprint games like Accordion, Clock, and Aces Up have state spaces small enough that a solver can traverse them fully on modest hardware. The trouble is that "small state space" correlates with "few meaningful player decisions," which is why those games sit at the low-skill end of the ranking. Games that reward skill tend to have branching factors that explode — Spider and Klondike specifically — and those are the games that force us into Monte Carlo sampling rather than exhaustive counts. The pattern generalises: if you can solve a solitaire game exhaustively on a laptop, you are probably playing a game with limited strategic depth. FreeCell is the unusual case where exhaustive work is possible on a closed deal library while the game itself remains deep, and that combination is why it is the most-studied solitaire variant in the academic literature.

Klondike is the canonical Monte-Carlo case in the solitaire literature. The state space for standard 52-card Klondike is too large for exhaustive search — the combinatorics of stock cycling and face-down cards blow past what any solver can enumerate — so the win-rate question has to be answered by sampling. Bjarnason, Fern & Tadepalli (2007), in their ICAPS paper "Lower Bounding Klondike Solitaire with Monte-Carlo Planning," ran planners across a large sample of deals and reported an upper bound of roughly 82% solvable under "thoughtful play" — their name for a solver that assumes full information about face-down cards. It is this figure that grounds the Klondike row in the database.

Thoughtful play is an optimistic bound, not the number a human player would hit. Real Klondike players cannot see face-down cards when they choose their moves; they have to commit without the information that the thoughtful-play solver gets for free. The result is a ~40 percentage-point gap between the 82% thoughtful-play ceiling and the ~30–40% rate at which engaged human players actually win. That gap is the entire reason Klondike feels hard: the game is deeply solvable in principle and mostly unsolvable in practice, because practical play is information-constrained.

Monte-Carlo results also need confidence intervals, and here is where sample size earns its keep. A Monte-Carlo run at N=10,000 deals gives you a 95% half-width of roughly 1 percentage point at the 50% win-rate level; halve the margin and you need four times the sample; halve it again and you need sixteen times. The Bjarnason paper's ~1,000,000-deal sweeps land at half-widths measured in tenths of a point, which is why their figure is the one everyone cites. When a site tells you "Klondike Draw-3 wins 21% of the time" with no N, treat the precision as theatre — the claim might be accurate, but the evidence presented does not support the precision quoted.

The draw-3 vs draw-1 story also lives on the Monte-Carlo frontier. Follow-up work (Blake & Gent, 2013, and subsequent community solvers) suggests that draw-3 with unlimited redeals is slightly more solvable than draw-1, because cycling the stock reveals more cards in total. Humans do worse at draw-3 despite the higher ceiling because the memorisation load rises — you have to remember which stock cards you skipped on each cycle. The database reports both ceilings separately so readers do not collapse them into a single number.

Other Monte-Carlo results we watch, but treat cautiously until we reproduce them ourselves, include community solver statistics from the Spider and TriPeaks communities and the long-running win-rate logs kept by the Pagat Solitaire mailing-list crowd. These are the sorts of sources that cite each other without always surfacing a primary run, which is how vague numbers propagate. Our project plan is to rebuild each row with our own sweep at N ≥ 100,000 per variant and a disclosed heuristic, then publish both the aggregate figure and the raw per-deal results. Until that work ships, the database is transparent about which rows are ours and which are someone else's.

Roughly two-thirds of the database sits in the estimate and community data tiers. These are games with no peer-reviewed solver paper behind them and no first-party simulation from us yet. Pyramid, TriPeaks, Golf, Canfield, Forty Thieves, Yukon, Seahaven, Baker's Game, Eight Off, and the three Spider difficulties all live here. The figures are honest — drawn from community telemetry, older desktop-client statistics, and solver write-ups across the solitaire enthusiast community — but they are not yet figures we have produced ourselves or traced to a citable academic source.

What do we know about these games? Quite a lot, directionally. Yukon is ~85% solvable because all cards are face-up from deal one, which collapses it into an information-complete puzzle similar to FreeCell. Eight Off is near-universally solvable because it hands the player eight free cells rather than FreeCell's four, and those extra reserves dissolve nearly every tableau jam. Forty Thieves is brutal — somewhere in the 10-20% range — because its strict same-suit descending rule punishes sequencing errors that FreeCell and Klondike would forgive. Pyramid's win rate depends heavily on the redeal allowance: single-pass Pyramid sits around 1-3%, while three-redeal Pyramid climbs to ~6%.

We publish these as estimates rather than silences because a directionally-correct number with an error bar is more useful than no number. When we move one of these rows from "estimate" to "monte-carlo," it means we have run the simulator ourselves, disclosed the solver heuristic, and collected enough deals to shrink the confidence interval to something we trust. Until then, the tag warns you to discount the precision. A row that reads "65% (estimate, community_data)" should be parsed as "around two-thirds, from telemetry we did not generate." It is still useful — it tells you the game is meaningfully skill-dependent and that a typical session can be won — but the second decimal place is not trustworthy, and we will not pretend it is.

Search for "Klondike win rate" and you will see figures spanning 15% to 82%. Search for Spider 2-suit and you will see 30% to 70%. The spread is not random — it comes from different sites measuring different things and labelling them all the same way. Some sites report thoughtful-play ceilings; others report human-player telemetry; a few report the win rate of a specific greedy heuristic run by their own solver without disclosing which one. All three are valid numbers, and none of them are interchangeable.

When you encounter a competitor win-rate claim, we recommend asking four questions. First, what population was sampled — random deals, a specific numbered library (like the Microsoft 32,000), or live player games? Second, who was playing — a full-information solver, a specific heuristic, or real humans? Third, what was N — and is it large enough to support the precision quoted? Fourth, where is the primary source — is the number traceable to a paper, a solver run, or a blog post that does not cite anything at all? A claim that survives all four questions is usable; a claim that fails any of them should be discounted.

We practise what we preach by exposing the primary source for every row. When we cannot trace a number to a paper or a first-party solver run, the row carries the estimate tag and says so. We would rather publish "we do not know precisely yet" than dress up a guess as a finding. That is the quiet gap between a citation database and a marketing page, and it is the gap this project exists to close.

The best-known Klondike win-rate disagreement is a good worked example. Some sources cite 82%, others cite 43%, others cite 21%. None of those numbers are wrong in isolation, and they are not even talking about the same question. 82% is the thoughtful-play upper bound from Bjarnason et al., where the solver sees face-down cards. 43% shows up in community write-ups as an approximation of strong human draw-1 play with unlimited redeals. 21% tends to appear for draw-3 human play under casino-style scoring. All three are useful; none are interchangeable; and a site that reports one of them without clarifying which question it answers is not being helpful to its readers. We try to make which-question explicit for every entry in the database.

The practical takeaway from the database is simple: you are probably not bad at Spider. Four-suit Spider's ceiling is 5-15% under community telemetry, and that is the solvable ceiling, not the human rate. A real player who wins 8% of four-suit deals is playing within normal bounds for the game. The same correction applies to Pyramid (1-3% under single-pass rules), Golf (5-12%), and Forty Thieves (10-20%). If your Spider 4-suit win rate is 10%, the database says you are exactly where the game allows — you are not missing a trick that would push you to 60%, because 60% does not exist for that ruleset.

The same database also tells you where the skill gradient is worth climbing. FreeCell, Yukon, Eight Off, and Spider 1-suit all have solvable ceilings above 85%, which means effort on those games compounds — a 10-percentage-point improvement in your sequencing actually reaches. For the high-luck games (Clock, Pyramid single-pass, four-suit Spider), improvement returns are flatter and the right expectation is that most sessions end in a loss by design.

The last practical point is about undo. Every solvability figure in the database assumes undo is available, because solvers assume backtracking. If you play no-undo FreeCell — the original Microsoft ruleset — your personal win rate will sit meaningfully below the solvability ceiling, because one wrong sequence kills the whole deal. The gap between undo-allowed and no-undo play is a small but real adjustment to make when you compare your own stats to the database. Undo is not cheating; it is a tool that lets the game play at the level its solvability ceiling implies. No-undo is a harder discipline that most players should treat as a separate mode with its own lower expected win rate.

The database has obvious holes. The Spider rows are all community-data tier; we would like to replace them with our own Monte-Carlo sweeps at N ≥ 100,000 per difficulty, using a disclosed solver heuristic, and with matched confidence intervals. Pyramid needs separate rows for single-pass, two-redeal, and three-redeal rulesets, because the spread is wide. Forty Thieves has enough rule variants (Josephine, Indian, Limited) that each deserves its own row rather than being flattened into one band.

We are also interested in the exotic variants we currently cover at the rules level but not at the win-rate level: Scorpion, Bristol, Accordion, La Belle Lucie, Cruel, Bakers Dozen, and the restricted-cell FreeCell variants (1-cell, 2-cell, 3-cell). Restricted-cell FreeCell in particular is a natural extension of Keller's work — we know the win rate collapses as free cells are removed, but the shape of the drop has not been published in a way we find citable. A clean result here would give players a calibrated menu of difficulty steps between Baker's Game and classic FreeCell, and it is exactly the kind of project our simulator roadmap was built for. When we close one of these research questions we will update this page and note the date in the update log. If you have a source we should be reading, write to the Research Desk — we take reader corrections seriously and will re-run the numbers in public when a better primary source turns up.