How the Doubling Cube Works in Backgammon (2026 Guide)

The doubling cube is a small numbered die — values 1, 2, 4, 8, 16, 32, 64 — that turns backgammon from a single-shot dice game into a stake-management negotiation. Either side can offer to double the stake at the start of their turn; the receiver either accepts (and now owns the cube — only they can redouble) or refuses (and forfeits the game at its current value). This guide explains the actual decision thresholds used by the Nardex backgammon AI, the official Crawford and Holland rules, and the most common cube mistakes that cost amateurs match points.

Cube in backgammon

In backgammon (short narde), the cube has been a standard part of tournament and match play for nearly a century — it is where the offer/take thresholds and the Crawford/Holland rules below come from. The «Cube in long narde» section further down the page covers how the cube differs in the other rule set.

What is the doubling cube?

The cube was invented in New York around 1927. Before it, backgammon was a pure dice game — you played one game, scored 1 / 2 / 3 points depending on win type, moved on. The cube added a layer of negotiation: at any point before your roll, you can offer to double the stake. Refusal means the game ends immediately at the current cube value (so refusing a 1-cube is a 1-point loss; refusing a 4-cube is a 4-point loss). Acceptance means the game continues, but now the receiver owns the cube and is the only side that can redouble next.

Mechanically: the cube starts in the centre of the bar at value 1 (centred = either side can offer). After the first acceptance, the cube physically moves to the accepter's side and now shows 2. The accepter can later offer at 4×; if accepted, the cube moves back, shows 4, and so on up to 64. Cap is 64.

When should you double — why the 66% threshold?

The classical doubling threshold for money games is 66% winning probability (equity ≈ +0.32 on the −1/+1 scale). Below that, doubling is premature and gives away equity: your opponent will gladly take a position you should have kept playing without giving them the cube ownership. Above 80%+, you risk the opponent dropping — you collect the current value instead of the larger one you would have won by playing on.

The Nardex AI uses these exact thresholds in game/bot_doubling.rs:

• Money game / even match score: offer at equity ≥ 0.66 (≈ 66% wins)
• Trailing in a match: offer at equity ≥ 0.60 (more aggressive)
• Leading in a match: offer at equity ≥ 0.70 (more conservative)

The match-score adjustment matters because trailing doubles a chance to catch up; leading doubles risk losing more than necessary. These calibrations come from match-equity tables empirically derived over decades of expert play.

When should you take, and when should you drop?

On the receiving side the question flips: should you accept the doubled stake, or eat the current loss and move to the next game? The classical cubeless take point is approximately 25% — meaning if you have at least a 25% chance of winning, taking is correct expected value, even though you are favoured to lose.

Why is the take point 25%?

The intuition: if you take at 25%, you win 1× the new cube value 25% of the time and lose 1× the new cube value 75% of the time. EV = 0.25×(+1) + 0.75×(−1) = −0.5, expressed in cube units. Compare against dropping: −1.0 (you lose the current cube value for sure). Taking is −0.5, dropping is −1.0. Taking is correct.

Nardex AI take-point thresholds:

• Money / even match: take at equity ≥ −0.50 (≈ 25% wins)
• Trailing: take at equity ≥ −0.60 (≈ 20% wins — more aggressive)
• Leading: take at equity ≥ −0.40 (≈ 30% wins — more conservative)

These are cubeless take points. In reality, owning the cube has additional value (the right to redouble), which lowers the take point further. Expert tables put the practical take point closer to 22% in money games. The Nardex AI uses cubeless thresholds for simplicity and predictability.

How does the cube work in match play: Crawford and post-Crawford?

Match play introduces wrinkles. The biggest is the Crawford rule:

When either side reaches match-point minus 1 (e.g. 4−5 in a 5-point match), the next single game is played without the cube. This is the Crawford game. The cube panel is greyed out — neither side can offer.

Why: without Crawford, the trailing player would simply offer immediately on roll one. The leader would have to accept (refusing = match over) and gamble at 2× from a starting position — which makes the cube degenerate at match point. Crawford forces one cube-free game and preserves the lead.

Post-Crawford (the Crawford game has been played, both sides can use the cube again) the trailing side should always offer at the start of their turn, because they cannot win the match without doubling. The Nardex matchmaker auto-detects Crawford state and disables the cube panel for that one game.

What is the optional Holland rule?

Post-Crawford, the cube can be offered immediately on the trailing side's first roll. To avoid that auto-roll, the Holland rule says: post-Crawford, the cube cannot be turned until both players have moved twice each (so doubling becomes available from move 3 onward). It gives the leader a small window to actually play the position before facing a forced cube decision.

Holland is optional — common in tournaments, less common online. Nardex supports it as a feature flag but it is off by default. If you want it on, mention it in your match settings when joining a long match.

Equity-based decisions: how does the AI decide?

Putting it all together, here is the Nardex AI decision flow each time it is on roll:

1. Estimate your equity. Equity is your math expectation of the game outcome on a [-1, +1] scale: +1.0 = certain normal win, +2.0 = certain gammon, -1.0 = certain loss. Use a position evaluator (or pip-count rules of thumb in race endgame) to read it. Equity is also expressed as winning probability — convert via win_probability ≈ (equity + 1) / 2.
2. Check whether you can double at all. Three blockers: (1) you must own the cube or it must be centered (the side that just declined a previous double cannot redouble); (2) cube value < 64 (cube cap); (3) the game is not the Crawford game in match play (one game without cube after either side reaches match-point − 1).
3. If equity ≥ 66% — offer the double. Money-game offer threshold is approximately 66% winning probability (equity ≈ +0.32). When trailing in a match this drops to 60%; when leading, rises to 70%. The Nardex AI bot uses exactly these thresholds (game/bot_doubling.rs). Below 66% — keep playing without offering; doubling too early leaks equity.
4. On the receiving side — apply the take point (~25%). When the opponent doubles, you decide take or drop. Cubeless take point is ~25% winning probability: above that, take (accept double) and play on at the new stake; below, drop (forfeit current value, game ends). Trailing in a match shifts take point down to 20%; leading shifts it up to 30%.
5. After acceptance — repeat from step 1 next decision. When a double is accepted, the cube goes to the accepter — only they can offer the next double. Re-evaluate equity each turn; once it crosses your threshold again, double again (it is now a redouble). The cube doubles each time: 1 → 2 → 4 → 8 → 16 → 32 → 64 (cap).

For an exact equity number you can use the Nardex Analysis API — submit a \`cube_decision\` request with the current position and \`offer_made\` flag, get back a ranked alternative for «double / no_double» (offer side) or «take / drop» (response side) with equity for each. Same engine as the in-game bot.

What are the most common cube mistakes?

1. Doubling too early. «I am ahead, why not double» — at 55% wins, doubling loses ~0.3 cube units of expected value because you give cube ownership to the opponent. Wait until 66% (equity ≥ 0.32).
2. Doubling too late. «I want to make sure» — at 90% wins, the opponent drops and you collect 1× the current value instead of 2×. Equity loss ≈ 0.4 per game. Double around 70-80%, before the opponent's drop becomes obvious.
3. Dropping reasonable takes. Especially in race endgames: at 28% wins people often drop because they think they are «too far behind». They lose 1×; taking and losing 2× has worse single-game outcome but better expected value. Compute take points, don't trust gut.
4. Taking suicidal positions. Mirror image: at 18% wins many casual players take because «anything can happen». EV is significantly worse than dropping. Drop below 20% (or 25% in even matches).
5. Forgetting Crawford. Trailing player at match point − 1 sometimes tries to double during the Crawford game itself; the cube is locked. After Crawford it returns; use it.
6. Ignoring gammon risk. If your opponent has a 30% gammon chance, the cube value is effectively 1.3× higher when computing take points. A position with 28% wins is actually a drop if gammon risk is ≥ 25% — equity calculation must include the distribution, not just win/lose.

Cube in long narde

In long narde, the cube is not exclusive to short backgammon. The Russian Ministry of Sports order №347 of 30.04.2025, article 23, officially adopts it for tournament long narde with the same 1→2→4→8→16→32→64 progression and the same Crawford rule. Holland is also optional in long narde and supported by Nardex. Beaver/raccoon and Jacoby are not implemented for either game family.

The strategic role of the cube differs slightly. In long narde the absence of hits means equity swings are smoother — there is no single dice roll that flips a 70% position to 40%. That makes cube decisions tighter: thresholds are closer to the theoretical optima because variance is lower. See our article on long narde strategy for the cube-aware sections of opening, mid-game, and endgame decisions.

Frequently asked questions