Cribbage Odds and Statistics: The Exact Numbers

A random cribbage hand, four cards plus the starter, averages 4.7692 points. Every figure on this page comes from scoring all 12,994,800 possible hand-and-starter combinations with two independently written scorers, plus a 25,000-game fixed-seed simulation, so each number can be regenerated and checked exactly.

What does the average cribbage hand score?

Averaged over every possible keep and starter, weighted equally, a cribbage hand scores 4.7692 points under hand rules and 4.7348 under crib rules. The entire gap is the flush rule: a four-card flush counts in the hand, while a crib flush needs all five cards, and that rule is worth 0.0343 points of the average.

The distribution bunches low. Four is the single most common score at 21.98% of hands, just ahead of two at 21.65%, and 7.76% of hands score nothing at all.

A slight majority of hands (56.05%) score four or fewer, 18.95% reach eight or more, and only 0.67% reach sixteen. Odd scores are rare, 19.31% of hands in total. An odd total needs the five-card flush, nobs (the jack matching the starter's suit, worth one), or runs that add an odd number of points, such as a single run of three or five. Before the starter, four cards alone score zero in 27.42% of keeps and exactly two in 40.98%. The four-card maximum of 20 belongs to a single keep, the four 5s.

These are averages over random cards. Since a player keeps the best four of six, the hands a player actually keeps run higher. No exactly computed figure exists for that average, so this page publishes none.

How rare is a 29 hand?

A 29 hand comes up about once in every 216,580 six-card deals, for a player who always keeps 5-5-5-J when dealt them. That is the figure the American Cribbage Congress publishes and the answer to the question as most players mean it. This site's enumeration re-derives it exactly.

The odds depend on the deal model assumed. The three defensible framings differ by a factor of about fifteen, and quoting one without its assumption is how published figures end up contradicting each other.

The headline figure re-derives exactly: 4,324 qualifying six-card deals, each cut from 46 unseen cards, give 936,491,920 possibilities, and that number divided by 4,324 is 216,580 with nothing left over.

How rare is a 28 hand?

A 28 hand comes up about once in 170,984 hands with a random keep and a random cut: exactly 76 of the 12,994,800 possible combinations score 28. That makes it nineteen times as common as the 29, and still a once-in-years event. The table's last row is the only 28 figure this page computes. The American Cribbage Congress also publishes a 28 figure, under a different deal assumption from the one in the table. The two numbers answer different questions rather than disagreeing, and both ACC figures are covered on the 29 hand page.

Which hands score 24?

A 24 hand turns up 3,680 times in 12,994,800, or once in 3,531 hands under the random keep-and-cut assumption. Exactly nine rank layouts, hand plus starter, produce it.

The five run-based layouts allow 720 suit arrangements each and the four four-of-a-kind layouts 20 each, so 4 × 20 + 5 × 720 = 3,680. The count matches the full distribution, and it makes 24 oddly common for its height: 3,680 hands score 24 while only 444 score 22 and 356 score 23. Scores of 19, 25, 26 and 27 never appear in the distribution at all. Why they are impossible is explained on the nineteen hand page, and the full ladder from 20 to 29 runs through the best hand page.

What is a discard worth?

Across all 91 possible rank pairs, a tossed pair is worth between 3.41 and 9.37 expected crib points: 5-5 is the most valuable toss into either crib, and 10-K the least. The values below are the crib expectation the Nobs engine itself uses, read out for each of the 91 rank pairs. They are this site's own computation, with the method stated at the bottom of the page, and are cross-checked against simulated play. These are the extremes:

The cross-check comes from simulated play: 25,000 fixed-seed games of ONYX, the strongest of the three bots in the browser game, playing itself, with each game's final count cut short at 121 points. Bots also choose their own discards, so the measured crib values record what ONYX actually collected, not what a random toss is worth, and sample sizes differ by pair. Where the sample is large the model holds: 10-K tossed to the opponent's crib averaged 3.92 measured points across 8,641 such tosses, matching the modelled 3.92 to two decimals.

What does a real discard decision look like?

Take the deal behind the 1 in 216,580 figure: 5♣ 5♦ 5♥ J♠ plus K♦ Q♣. As non-dealer, the model's best choice keeps 5-5-5-J and tosses K-Q. That hand averages 16.39 points over the 46 possible starters, with a minimum of 14 and a maximum of 29. The opponent's crib is expected to cost 4.47, for a net of 11.92, since the non-dealer's net is the hand average minus the crib's value.

As dealer, the best choice flips to keeping 5-5-5-K and tossing J-Q into its own crib. A hand average of 16.13, plus 5.01 expected crib points, nets 21.14, since the dealer's net is the hand average plus the crib's value. The dealer's line gives up the jack, and with it any chance of the 29. Tossing J-Q builds a crib worth 5.01 against 3.98 for K-Q, and that gain outweighs the 0.26 the hand loses.

What do real games look like?

The numbers in this section measure simulated play, not cribbage in the abstract. They come from 25,000 games of ONYX, the strongest of the three browser-game bots, playing against itself from a fixed seed, so the whole corpus can be regenerated exactly. Two limits matter. Bot-versus-bot play is a proxy for skilled play, not a claim about human games. And game scores stop at 121, so each game's final scoring item is cut short: in 51.57% of games (12,892 of 25,000) the winner's last count would have carried past 121.

Within that corpus, the player who dealt first won 56.40% of games (95% confidence interval 55.79% to 57.01%). A skunk, where the loser finishes with 90 points or fewer, ended 13.80% of games (13.38% to 14.23%), and a double skunk just 0.12% (0.08% to 0.17%). Nine hands was the commonest length, at 44.13% of games. The average was 8.98 hands and the range 5 to 13.

How many points does each role score per hand?

Averaged over all 224,426 hands in the simulated corpus, including each game's cut-short final hand, the two roles scored as follows. The pone is the non-dealer.

The pone's counted hand averaged more than the dealer's, 7.8042 against 7.4211. The likeliest reading is that the model trades hand strength for crib value when it holds the deal. The dealer pegged nothing in only 0.52% of hands, consistent with the rule that a two-player dealer always pegs at least one point unless the game ends mid-play. The pone pegged nothing in 22.21%. One sanity check is built in: his heels paid the dealer 0.1538 points per hand at two points per jack starter, so jacks were cut 7.69% of the time. That matches the exact 4 in 52, which is 7.6923%. These per-hand figures come from a truncated bot corpus and are not comparable to classic per-deal theory averages, which live on the dealer page.

How were these numbers computed?

Every figure on this page is read from one committed data file, src/_data/odds.json, produced by a generator script in this site's repository (scripts/odds/generate-odds.mjs). For the exact figures, the generator scores all 12,994,800 hand-and-starter combinations with two independently written scorers: the Nobs engine's own scorer and the separate port behind this site's hand calculator. Both run in lockstep under hand rules and again under crib rules. The generator halts on any disagreement, so a committed file is itself proof that both scorers agreed on every combination under both rule sets. The exact engine build is pinned by hash.

The simulation runs from fixed seed 1000003. Anyone with the repository can regenerate the file byte for byte. The generator refuses to write a file that does not round-trip, and a build check fails whenever the committed data drifts from what the scripts produce.

One distinction governs how to quote this page. The distribution, rarity, layout and discard-model figures are exact, from full enumeration or a deterministic model, with no sampling error. Only the game-simulation figures carry sampling error, and the headline rates (the first-dealer win rate and both skunk rates) are quoted with 95% confidence intervals.

Where do the deeper explanations live?

Each number here has a page that explains the cribbage behind it. The American Cribbage Congress's published rarity figures are quoted on the 29 hand page. The ladder of high hands lives on the best hand page, the impossible scores on the nineteen hand page, and per-deal dealer-advantage theory on the dealer page. To count any single hand against these distributions, the cribbage hand calculator scores it line by line. How to use these numbers at the table is on the strategy pages: what to discard, pegging, board position and the endgame, with the hub at cribbage strategy.

By the Nobs Cribbage team Last reviewed 2026-08-29