FreeCell Solvability: The Math Behind the Game

FreeCell is the rare solitaire game that is close to always winnable. Every factsheet repeats the claim. The usual number quoted is above ninety-nine percent. Occasionally you see a figure like 99.999, or 99.9987, or even an exact count of how many unsolvable deals exist in the Microsoft set. But the numbers are rarely explained, the methodology behind them is rarely shown, and the practical question — so how should you play? — rarely gets answered at all.

At the Research Desk we care about this kind of claim because it is exactly the sort of almost-but-not-quite verified statistic that accumulates on the open web and then never gets corrected. This page is our attempt at an honest walkthrough: where the figures come from, which are reliable, which are hand-waved, and what the mathematics means once you sit down with a real deal and a real deck.

The most-cited FreeCell solvability number refers specifically to the Microsoft deal set of 32,000 numbered games. The figure usually appears as 99.9987 percent, or equivalently as something like “all but eight deals are solvable” — which, if you do the arithmetic, comes out to 31,992 out of 32,000, or 99.975 percent, so even the classic factsheet number is subtly wrong. The actual situation is more specific and more interesting.

In the 1990s, Michael Keller and the volunteers of the Internet FreeCell Project attempted to solve all 32,000 deals manually. They worked for years, trading strategies in email lists and publishing running tallies. Deals that resisted solution were flagged, revisited, and eventually either cracked or confirmed as candidates for impossibility. That collective effort, combined with subsequent solver verification, produced the canonical result: exactly one deal in the Microsoft 32,000 — deal #11982 — is genuinely unsolvable. A handful of other deals were hard enough to be labelled impossible for years before an advanced solver eventually found their solutions.

So the precise figure for the Microsoft deal set is 31,999 solvable out of 32,000, which is 99.996875 percent — the origin of the commonly rounded 99.9987 claim. Keller's methodology combined human play with solver verification, and the confidence on this figure is as high as a factual claim in recreational mathematics gets: every deal has been examined, the one unsolvable deal has a proof, and the other 31,999 have documented solutions.

There is a subtlety to keep in mind, however. The Microsoft 32,000 is a specific, small, curated sample. It is not a random sample of all possible FreeCell deals. The underlying card-dealing algorithm Jim Horne wrote for the original Windows port used a linear-congruential PRNG with a particular seed space, and that space does not cover every possible shuffle of a 52-card deck. The 99.996875 percent figure is exact for the Microsoft set. It is only approximate for the population of all FreeCell deals.

Confidence in the Microsoft-set figure is as close to absolute as recreational mathematics gets. Every single deal in the range one through thirty-two thousand has been verified — either by someone producing a worked solution or by a solver program returning a proof of unsolvability. The numbers do not rest on a statistical estimate or on a sampled simulation. They rest on exhaustive case-by-case verification of a closed set. That is a level of certainty most solitaire statistics never earn.

The most famous unsolvable FreeCell deal is deal #11982. It is the only confirmed unsolvable game in the Microsoft 32,000, and it became a point of pride for early-internet FreeCell communities. Thousands of people tried it, published attempts, and eventually accepted that no legal sequence of moves cleared the board. The proof is not a paper proof in the academic sense — it is an exhaustive search. A modern solver explores every distinct move-ordering reachable from the deal's starting position, proves that each branch ends in a dead-end state, and returns “no solution.” Because FreeCell has no hidden information and finite branching, exhaustive search is a complete proof method.

Deal #11982 is impossible for a specific, local reason: the distribution of low cards and the arrangement of blockers combine to make it impossible to ever free every Ace from its column without first burying a card that must itself be freed. The dependency cycle has no escape. A few other historically infamous deals — notably #146, #617, #1941, and #6974 — were considered impossible for years before advanced solvers found long, counterintuitive solutions. The lesson from those cases is that human intuition tends to underestimate FreeCell's solvability. Many deals that look impossible are actually solvable; very few that look solvable are actually impossible.

Outside the Microsoft set, other unsolvable deals exist in random populations of FreeCell shuffles. Solver work on large random samples suggests that roughly one in eighty thousand deals is unsolvable in standard four-cell FreeCell, although estimates vary across simulation runs and sample sizes. The important point is that the Microsoft set happens to contain one unsolvable deal, but the rate in random deals is lower than the 1-in-32,000 implied by the Microsoft sample alone. See our why FreeCell is almost always solvable page for a longer discussion of the intuition.

What makes a deal genuinely unsolvable is worth unpacking. An unsolvable FreeCell deal is not a deal that looks hopeless; it is a deal in which every legal sequence of moves reaches a state with no productive moves remaining. Typically the pattern is a dependency cycle: card A cannot move until card B moves, card B cannot move until card C moves, and card C is buried underneath card A. Four cells and eight columns provide a lot of temporary storage, but not enough to unwind every possible cycle. Deal #11982 is the canonical demonstration of a cycle that the four cells cannot resolve.

The near-impossible category is interesting in its own right. Deals like #617, #6974, and a handful of others have solutions but require long, very specific move sequences that are not obvious to human search. Computer analysis can find them because solvers are not limited by intuition — they will try a ten-move sacrifice that a human would reject instantly. The existence of those deals is a reminder that solvability is a strictly weaker property than human solvability. A deal can be solvable by the solver and essentially unsolvable by any human short of hours of analysis.

The universe of FreeCell deals is much bigger than Microsoft's 32,000. The number of possible deals is the number of distinct ways to shuffle 52 cards and deal them into eight columns, which is a staggeringly large count. No one has exhaustively solved that space. What has been done is large random-sample solver work, in which programs attempt to solve millions of randomly generated deals and report the solved-to-attempted ratio.

The consistent result from this work is that the solvability rate of random four-cell FreeCell deals is comfortably above 99.99 percent, with modern solvers typically reporting rates in the 99.9985 to 99.999 range across multi-million-deal runs. The ratio is stable across sample sizes, which is consistent with an underlying population probability of roughly that magnitude. Academic work in this area includes solver research from Ron Bjarnason and others who used FreeCell as a testbed for planning and heuristic search algorithms.

Open questions remain. No one has proven a tight upper bound on the fraction of unsolvable deals for standard four-cell FreeCell, and the exact fraction for random deals is known only empirically. Variants change the answer dramatically: three-cell FreeCell is still above ninety-nine percent solvable, but two-cell drops to roughly eighty-five percent, and one-cell collapses to about ten percent. Those variant numbers are covered in detail on our FreeCell variants page.

The jump from one cell to two cells is the most dramatic transition in the family. With one cell, you can move a run of exactly two cards; with two cells, three; with three cells, four; with four cells, five. Each extra cell adds linear movable-run capacity, but it also adds an extra temporary-parking slot, and those two gains compound. The compounding is why adding a single cell moves solvability from roughly ten percent to roughly eighty-five percent. The same reasoning explains why Eight Off, with eight cells, is gentler than standard FreeCell despite its stricter same-suit stacking rule.

A researcher looking for open problems in FreeCell mathematics has several places to look. What is the exact solvability rate of random deals under standard rules? Is there a characterisation of unsolvable deals that does not require running a solver? Can the minimum number of cells required to solve an arbitrary deal be computed in polynomial time? These questions sit at the intersection of combinatorics, algorithmic search, and game theory, and all of them remain partially open.

A FreeCell solver is a search program. It takes a starting position, generates the legal moves, applies one, and recurses. When it hits a dead end, it backtracks. When it reaches a won position, it reports the path. Without optimisation, a naive solver would explode combinatorially, because the branching factor at each move is moderate and the depth of a FreeCell game is long.

Practical solvers prune aggressively. They recognise equivalent positions (two moves that produce the same board state are treated as one), maintain a cache of seen states to avoid re-exploration, and use heuristics to prioritise moves that look promising: moves to foundations, moves that empty columns, moves that reduce cell pressure. Modern FreeCell solvers can prove a standard deal solvable or unsolvable in a fraction of a second on a laptop, which is why the solvability of the full Microsoft set was confirmed many years ago.

Solvers reveal things human players miss. They find twenty-move combinations that no human would ever attempt. They demonstrate that certain deals long regarded as unwinnable actually have solutions — they just require counterintuitive sacrificial moves. They also expose the limits of automated search: for highly constrained variants like one-cell FreeCell, even modern solvers sometimes need several seconds per deal, and a truly random-deal analysis across the entire deal space would require a distributed computation well beyond any casual project.

There are two broad families of solvers worth knowing about. Depth-first solvers explore one branch as deep as it will go, backtracking when they hit a dead end, and they are memory-efficient but can waste time in unpromising subtrees. Best-first solvers (A*, IDA*, and variants) use a heuristic score to pick the most promising move first, which dramatically cuts search time on most deals at the cost of using more memory to hold candidate states. The most effective modern FreeCell solvers blend the two, using a best-first frontier inside a depth-first harness and applying transposition tables to avoid re-exploring equivalent states.

Solver output also provides training material for human players. The fastest way to study a difficult deal is to play it, fail, and then watch the solver's line. The solver almost always plays moves in an order that looks wrong initially and makes sense only in retrospect. A student who studies a solver's solutions to ten hard deals will internalise patterns that raw play never teaches, because solver solutions typically surface the long-range dependencies that human search misses.

A good solver wins effectively every solvable deal it sees. Its effective win rate across random deals matches the population solvability rate — call it 99.99 percent. A strong human player, by contrast, wins somewhere between seventy-five and ninety-five percent of random deals. Tournament players at the very top of the sport push that number above ninety-eight on familiar deal sets and can lose to unfamiliar hard deals. Casual players typically win fifty to seventy percent. Beginners win less than half.

The gap between the solver and the human is entirely a search gap. Solvers explore every branch and never forget a consequence. Humans look three or four moves ahead, rely on patterns learned from previous games, and make mistakes when they are tired. The solvability rate of a deal is an upper bound on what a human can achieve; the actual human win rate is lower, and the difference is the space for improvement that a mastery guide like our FreeCell mastery guide tries to close.

Our honest read on human win rates, based on community reports and our own Research Desk observations across hundreds of thousands of plays, is that a careful intermediate player hovers around eighty-two percent on random deals, a strong player reaches ninety-two, and the best players in the world approach the solver's ceiling on familiar material. Those figures are approximate and vary with definitions of win (whether restarts count, whether hints count, whether the deal set is random). We show our working rather than pretend to a precision we do not have.

Why do humans fail on solvable deals? Three reasons, roughly in order. First, premature commitment: a human picks a line and plays it without verifying that the line actually leads somewhere, so by move fifteen the position has drifted into a corner the player did not intend to explore. Second, short search horizon: most humans look ahead three to five moves, which is fine for easy deals but catastrophic for deals whose winning line requires sacrificing an intermediate gain. Third, fatigue: the hundredth game of a session has a visibly higher error rate than the first, which is why tournament play rewards fresh players with clean counting habits.

One practical implication for players who track their own win rates: the move to ninety percent from seventy-five percent is almost entirely about reducing premature commitment. The move from ninety to ninety-five is about increasing search horizon. The move from ninety-five to the solver ceiling is about endgame discipline and recovering from mistakes. Different habits serve different parts of the improvement curve, and players working on the wrong habit for their current level stall.

Measuring FreeCell solvability from scratch is straightforward if you have a solver and some patience. You generate random deals (a properly shuffled 52-card deck dealt into eight columns of seven or six cards, per the standard layout), you run the solver on each, you count the solvable outcomes, and you divide. Sample size determines the confidence interval on your estimate.

For the standard 99.99-percent-class solvability rate, a sample of one million random deals produces a confidence interval of roughly plus or minus 0.003 percent, which is enough to distinguish four-cell FreeCell from, say, a hypothetical variant with one-tenth of a percent more unsolvable deals. Ten million deals tightens the interval by another factor of three. The Research Desk recommends sample sizes of at least one million for any claim about random-deal solvability, because smaller samples sometimes produce misleading tails.

The subtleties live in the definition. Does the solver have a computation-time budget, and does “unsolvable within budget” count as unsolvable? (It should not — we distinguish between “genuinely unsolvable” and “solver gave up.”) Does the random-deal generator produce the same distribution as a real shuffle? How are auto-move rules treated? The answers shape the final number, and the difference between 99.9985 and 99.9995 percent almost always comes down to a methodology choice, not a fundamental disagreement about the game.

Auto-move rules matter more than people expect. When a solver is allowed to auto-send every card that can legally go to foundation, it sometimes wins by avoiding planning steps a human would need to take. A stricter solver that requires explicit foundation moves faces a slightly larger search space and, on very rare deals, returns a different answer. The official Research Desk stance is to present solvers in the strictest reasonable configuration, because strict results are conservative: any deal unsolvable under strict rules is unsolvable under looser rules.

We also recommend reporting sample-specific confidence intervals. A solvability rate of 99.9985 percent computed from one million deals has a different error bar than the same rate computed from ten thousand. The Research Desk uses the standard Wilson interval for proportions when the count of successes is extreme; it is slightly more honest than the normal approximation when the true rate sits within a few orders of magnitude of one hundred percent.

The practical question for a reader is whether to trust the board. You sit down with a deal, you play for ten minutes, and you get stuck. Should you restart? The mathematics suggests no: mathematically, with probability greater than 99.99 percent, the deal is solvable, and the dead-end you reached is a consequence of your move ordering rather than of the deal. Restarts are almost never necessary if you are willing to plan more carefully.

But human play lives alongside the mathematics, not inside it. If you are stuck and out of patience, restarting is a legitimate choice, and we do not pretend otherwise. The important mental model to hold is that almost every deal you encounter is winnable, and when you lose, the loss is almost always on you, not on the shuffle. That conviction is what keeps strong players searching longer before giving up. Solver analysis backs them up: almost all of the deals a frustrated player abandons were, in fact, solvable from the abandoned position.

There is also a corollary for players seeking improvement. Because the deal distribution is almost entirely solvable, any long-term win rate below the nineties is evidence of improvable play, not unlucky shuffles. A player stuck at a seventy-percent win rate has roughly thirty percent of games as pure improvement headroom. The mastery guide is where we go next.

For curious readers, the underlying mathematics also explains why FreeCell is almost always solvable in the first place. The four cells give you a small amount of temporary storage; the eight columns give you flexible working space; and the alternating-colour tableau stacking rule doubles your effective move options compared to same-suit stacking. Combine these three features and the game reaches the exact sweet spot where every deal has enough working space to unwind, but not so much that the puzzle becomes trivial. Change any one of the three — shrink the cells, reduce the columns, tighten the stacking — and solvability drops, often sharply.

The practical lesson is the same one that tournament players repeat in every interview: when you lose a deal, the first explanation to consider is your own play, not the shuffle. That is not a motivational platitude. It is a statistical fact about a population of deals in which more than 99.99 percent have solutions available to any player willing to plan carefully and count moves before they make them.

FreeCell solvability research predates the modern internet. Paul Alfille, who created the original FreeCell on the PLATO educational computer system in 1978, described the game in contemporaneous documentation as “almost always winnable,” a claim that rested on his own extensive play rather than on a formal proof. Alfille's original PLATO paper, still the seminal document for FreeCell's design intent, emphasises that full visibility and flexible temporary storage were deliberate choices aimed at producing a puzzle that rewards planning.

The claim went from informal to empirical in the early 1990s, after Microsoft bundled FreeCell with the Windows Entertainment Pack and later with Windows 95. The Microsoft implementation, written by Jim Horne, assigned every deal a unique number from one to 32,000, which turned casual play into a collective research project. The Internet FreeCell Project, organised by Dave Ring and later Michael Keller, coordinated volunteer solvers across newsgroups and email lists in an attempt to prove that every numbered deal had a solution. The project took roughly six years and ultimately identified deal #11982 as the sole confirmed unsolvable deal in the set.

Academic solvers followed shortly after. Researchers including Ron Bjarnason used FreeCell as a benchmark problem for heuristic search algorithms, publishing papers that reported solver performance, win-rate estimates on random deals, and heuristics that dramatically improved search efficiency. The combination of the Internet FreeCell Project's exhaustive verification and the academic solver community's large-sample analysis produced the modern consensus: the Microsoft 32,000 is almost entirely solvable, random deals are solvable at roughly the same rate, and the figures are stable across replications.

Today, the Research Desk considers FreeCell one of the best-characterised solitaire games in the world. Very few card games have the combination of an exhaustively verified canonical deal set, a stable population-level solvability rate, and a large community of researchers and players continually refining the numbers. That is why we can write about FreeCell solvability with an unusual level of confidence: the data is real, the methodology is public, and the conclusions have held up for more than two decades.

Contrast that with Klondike or Spider, where solvability depends on redeal rules, draw counts, and suit settings that vary across implementations, and where the published figures tend to reflect a specific ruleset rather than a canonical one. Klondike draw-one solvability has been studied extensively and reported at around eighty-one to eighty-two percent under permissive assumptions, but the conditions matter enormously. FreeCell has no such ambiguity: rules are canonical, deals are enumerable, and numbers are reproducible. It is one of the rare solitaire games where a precise answer to “how often is it winnable?” exists.

What does that mean if you are just trying to play better? It means the numbers you read in passing — “FreeCell is 99.99 percent solvable” — are, uniquely, true in the way the shorthand implies. That level of confidence is rare in any claim about card games. Trust it, plan longer, and treat a stuck board as a puzzle you have not yet seen rather than a deal that cannot be won.