How to play sudoku
Fill every empty cell with a digit from 1 to 9 so that each row, each column and each of the nine 3×3 boxes holds all nine digits exactly once. The digits printed at the start are clues; they never change. That is the whole rulebook, and it has not changed since Howard Garns published the puzzle as Number Place in Dell magazines in 1979.
There is no arithmetic in sudoku. Nothing is ever added or compared. The nine digits are just nine distinct symbols and the puzzle would work identically with nine colours or nine animals — which is worth knowing, because it means every deduction you make is about where a symbol can go, never about what it is worth.
Click a cell and type a digit, or use the pad. Pencil switches the pad to candidate marks so you can note the possibilities in a cell without committing. Undo is unlimited. Check tells you whether anything currently on the board clashes — it reads the board, not the answer, so it never quietly tells you a digit you had not earned.
The two things almost everyone gets wrong
1. A puzzle with two solutions is not a sudoku
If a grid can be completed two different ways, then at some point a correct deduction is impossible and the only way forward is to pick one and see. That is not a hard puzzle, it is a broken one — and it is the single most common defect in free sudoku, because checking for it costs a solver run per clue removed and generators that skip the check are much faster to write.
Every puzzle on this page is proved unique before you see it. The generator removes a clue only if the grid still has exactly one solution. Our test suite then proves it again with a completely separate solver — a deliberately stupid one that keeps no candidate lists, propagates nothing, and tests a placement by scanning the whole board — because a generator that checks its own work with its own solver has proved nothing. The two must agree on every puzzle, and they do.
2. Clue count is not difficulty
Nearly every free site grades puzzles by how many clues it removed. It is the wrong measurement and the evidence is easy to state. A published 17-clue puzzle — the fewest clues any sudoku can have — is solvable on this page's ladder using nothing but hidden singles. Arto Inkala's famous 2012 puzzle has 21 clues, four more, and our entire technique ladder cannot make a single move on it.
We put a number on it too, and the honest number is not the obvious one. Pooled across all four of our levels, clue count and difficulty correlate at about r = −0.7 — but that figure is measuring our own generator, which digs deeper for the harder levels, rather than measuring sudoku. The question that means something is whether clue count predicts difficulty among puzzles of the same level, and there it explains under 3% of the variation in every band. Knowing a puzzle has 25 clues tells you almost nothing about how hard it will be.
So difficulty here is the hardest technique you are forced to use. The ladder runs: naked single, hidden single, pointing pair, box/line reduction, naked pair, hidden pair, naked triple, hidden triple, X-wing, XY-wing, swordfish. Each level is a range on that ladder, the puzzle is guaranteed to need nothing above its level's top rung, and the techniques it actually requires are listed beside the board before you start.
No guessing. This is the part worth the page.
A puzzle that needs trial and error is a search problem wearing a puzzle's clothes. You cannot enjoy it, you cannot learn from it, and when you finally finish you do not know whether you were clever or lucky. Plenty of sites serve exactly that at their top difficulty while calling it "expert".
Our generator cannot produce one. It starts from a finished grid and removes clues, but a clue comes out only if the named technique ladder can still finish the puzzle from scratch. The moment a removal would force a guess, that clue goes back. A pure-logic path therefore exists by construction, not by hope — and because every rung of the ladder is a forced deduction, logic-solvability implies uniqueness for free. We test both anyway.
The test suite also checks something subtler and more dangerous: that the techniques are sound. It replays every deduction the ladder makes across 82 puzzles — our own and 50 published ones — and checks each against the true solution. A placement must match; an elimination must never remove a correct candidate. Over 4,500 deductions, a technique with an off-by-one in a box index would be caught on the step that made it.
The hint tells you the technique, not the answer
A hint that fills in a cell teaches you nothing, which is why the sites that meter hints can get away with metering them. Ours names the rung and gives the argument: "r4c7 is the only cell in box 6 that can still take a 7", or "every 3 left in row 2 sits in box 1, so no other cell in box 1 can be a 3". It always offers the easiest technique available — a hint that opens with an X-wing while a naked single is sitting there is teaching the wrong lesson. It is unlimited and free.
The honest limitation
Our hint engine cannot solve every sudoku in the world, and here is exactly where it stops. It knows eleven techniques, topping out at X-wing, XY-wing and swordfish. It does not implement forcing chains, Death Blossom, ALS moves or the other chain techniques that the extreme puzzles require. On Arto Inkala's 2012 puzzle it makes zero moves — not a few, zero, because not one rung applies to the opening position. On AI Escargot (Inkala, 2006) it finds exactly one hidden single and then stalls. Our test suite asserts both of those facts, so if someone adds chain techniques later the test fails and this paragraph gets rewritten on purpose rather than rotting quietly.
This does not affect the puzzles we deal you: they are generated inside what the ladder can do, so the hint always works on them. It is a limit on pasting somebody else's puzzle in, not on playing ours.
What we computed, and what we only cite
Sudoku has some of the best-known numbers in recreational mathematics, and most of them are far out of reach of a web browser. We are precise about which is which.
Cited, not recomputed. There are 6,670,903,752,021,072,936,960 completed 9×9 grids (Felgenhauer & Jarvis, 2005) and 5,472,730,538 essentially different ones once relabelling, rotation and reflection are discounted (Russell & Jarvis). The minimum clue count for a unique solution is 17, and 16 is impossible — proved exhaustively by McGuire, Tugemann and Civario in 2012 at a cost of roughly seven CPU-years. Nothing in this repository reproduces any of those three and no page on this site says otherwise.
Computed, every test run. The 288 completed 4×4 grids — enumerated twice, once one-by-one and once by symbol relabelling, because the second method is what makes the next number affordable and an unchecked shortcut is an assumption. The 28,200,960 completed 6×6 grids (2×3 boxes), a real enumeration. And 2,612,736 ways to fill the top band of a 9×9 with the first box in standard form, giving 9! × 2,612,736 = 948,109,639,680 top bands — which is step one of the Felgenhauer and Jarvis computation itself, so our 9×9 machinery is checked against the very paper whose total we decline to claim.
We also verify a real 17-clue puzzle: it has exactly one solution, and all 17 clues are load-bearing, because removing any single one leaves a grid with more than one solution. That demonstrates 17 is achievable. It does not prove 16 impossible, and we do not say it does.
Strategy that actually helps
Scan for hidden singles before naked ones. Beginners look for a cell with one candidate left. Faster players look for a digit with one home left in a box — pick a digit, cross off the rows and columns it already occupies, and see which box has only one cell standing. It finds more, earlier, and it is the technique most of a Gentle puzzle is made of.
Pencil marks are for pairs, not for everything. Filling every cell with every candidate is slow and makes the board unreadable. Mark a cell when it is down to two or three possibilities — that is when the pair, triple and wing patterns become visible, and those patterns are the entire difference between Moderate and Diabolical.
Work the box, not the grid. A box has three rows and three columns crossing it, which is nine constraints on nine cells. The densest information on the board is around whichever box already has the most digits, and it is almost always where the next move is.
When you are stuck, count a digit. Pick the digit you have placed most often and ask where the remaining ones can go. Locked candidates — a digit confined to one line inside a box — are invisible when you stare at cells and obvious when you stare at a digit.
Never guess, on this board. On most sites that advice is unenforceable. Here it is a property of the puzzle: if you cannot see the next step, one exists, and the hint will name it.
Questions people ask
How do you play sudoku?
Fill every empty cell with a digit from 1 to 9 so that each row, each column and each of the nine 3×3 boxes contains all nine digits exactly once. The digits printed at the start are clues and cannot be changed. There is no arithmetic in sudoku — the digits are just nine distinct symbols, and the puzzle would work identically with nine colours.
Does every sudoku have exactly one solution?
A proper sudoku has exactly one, and a grid with two solutions is not a sudoku at all — it is a puzzle-shaped object that will punish you for a correct deduction. Plenty of free generators never check. Every puzzle on this page has been proved to have exactly one solution by a counting solver, and the test suite re-proves it with a second, independently written solver that shares no code with the one that made the puzzle.
Do I ever have to guess in sudoku?
Not here. Every puzzle on this page is generated by digging clues out of a finished grid only while a named logical technique ladder can still finish it, so a pure-logic path exists by construction. The techniques each puzzle needs are listed beside the board before you start. Trial and error is a search, not a puzzle, and a site that serves you one while calling it hard is wasting your evening.
What makes a sudoku hard?
The hardest technique you are forced to use, not the number of clues. We measured this on our own puzzles: among puzzles of the same difficulty level, clue count explains under 3% of the variation in how hard they are. The extremes make the point better than the statistic. A published 17-clue puzzle — the fewest clues any sudoku can have — falls to hidden singles alone, while Arto Inkala's 21-clue puzzle from 2012 defeats every technique on our ladder.
What is the minimum number of clues a sudoku can have?
Seventeen. McGuire, Tugemann and Civario proved in 2012 by exhaustive computer search that no 16-clue puzzle has a unique solution, which took roughly 7 CPU-years. We did not recompute that and we do not claim to have — but we do verify a real 17-clue puzzle: it has exactly one solution, and every one of its 17 clues is load-bearing, because removing any single one of them leaves a grid with more than one solution.
How many sudoku grids are there?
There are 6,670,903,752,021,072,936,960 completed 9×9 grids, computed by Felgenhauer and Jarvis in 2005, and 5,472,730,538 that are essentially different once you discount relabelling, rotation and reflection (Russell and Jarvis). Those two are far out of reach of a browser and we cite them rather than claiming them. What our test suite does compute exactly, every run, is the 288 completed 4×4 grids, the 28,200,960 completed 6×6 grids, and the 2,612,736 ways to fill the top band of a 9×9 with the first box in standard form — which is step one of the Felgenhauer and Jarvis calculation.
Can your hint system solve any sudoku?
No, and this is the honest limit of this page. Our hint engine knows eleven techniques, ending at X-wing, XY-wing and swordfish. It does not implement forcing chains, Death Blossom or the other chain techniques that the extreme puzzles need. Arto Inkala's 2012 puzzle is the plain demonstration: our ladder makes zero moves on it — not a few, zero — and our own test suite asserts that rather than hiding it. Every puzzle we deal you is inside what the ladder can do, so the hint always works on our puzzles; it is puzzles from elsewhere it may not crack.
Is this sudoku free?
Yes. No signup, no account, no daily puzzle limit, no ads between you and the board, no paid hint packs and no timer nagging you to buy anything. Undo is unlimited, hints are unlimited, and pencil marks are free. Nothing on this page is metered.
Where sudoku comes from
The modern puzzle was invented by Howard Garns, a retired architect from Indianapolis, and published anonymously by Dell as Number Place in 1979. It was picked up in Japan by Nikoli in 1984, who added the two conventions that made it a craze — a symmetric clue pattern and a cap on the number of givens — and named it sūdoku, short for a phrase meaning "the digits must be single". The puzzle itself is public domain; it descends from Garns and, further back, from Euler's Latin squares.
The word "Sudoku" is a Nikoli trademark in Japan only — it was never
registered elsewhere, which is exactly why the worldwide boom adopted the word while Japanese rivals
had to advertise nanpure. Outside Japan the name is free to use, and the puzzle is free
everywhere. No publisher's puzzle set, styling or interface has been copied here; every puzzle on
this page was generated by the code in play/sudoku-engine.js.
The honest small print
Your completed puzzles and times are stored as unverified, because a browser can lie about its own score and we would rather label it than pretend otherwise. Puzzles are seeded, so the same seed always gives the same puzzle — the seed is shown beside the board. "Give up" reveals the solution and is deliberately a separate button that says what it does; nothing on this page ever fills a cell for you unless you ask it to. Difficulty ranking between techniques is a judgement call at the edges — whether a hidden pair is harder than a naked triple is a matter of taste, and our ladder's order is the conventional solver-community one rather than a law of nature.