War Blackjack

1The rules, as data

One card each, aces low, dealer takes ties, and the winner may roll the win into the blackjack bet.

Alongside the blackjack wager a player may post a War bet. The dealer gives the player one card face up and takes the upcard as their own, higher card wins even money, aces rank low, and the dealer wins ties. A player who wins may instead parlay: take back the original wager and add the win to the blackjack bet, knowing their own first card and the dealer's upcard. The published house edge is 1.16% on six decks, dealer hitting soft 17, blackjack paying 3:2, doubling after splits allowed.

One detail in the placard was the reason this bet got looked at. The War wager takes any amount within table limits and is not tied to a multiple of the blackjack wager, which is unusual for a side bet and would matter a great deal if the thing ever went positive.

This bet needed a resolution the rest of the site never does. Blackjack is blind to suits and treats every ten-valued card alike, so the engine collapses cards to values from 1 to 10. A war between a ten and a jack is a win, though, so the tie question lives on the full 13-rank scale, and the shoe already carried raw ranks underneath the collapsed values.

2The published-figure gate

Reproducing 1.16% turned into a reading problem worth a factor of six, and a strategy table settled it with no reference to the edge at all.

The published description says a war winner may collect the original wager and the winnings, or collect the original wager only and parlay the win by adding it to the blackjack bet. Read literally, the second option throws away a certain unit to put one unit on a hand, so a player would parlay only when the hand returns more than the unit it costs. That never happens, since the best first card against the friendliest upcard returns +0.71. Priced under that reading the whole wager sits at −7.3955%, and the published 1.16% becomes unreachable.

Under the other reading the win is yours either way, exercising the option costs nothing, and the criterion collapses to whether the blackjack hand ahead of you has positive expectation. That prices the wager at −1.2172% against the published −1.1600%. The check that decided it never touched the house edge. Deriving the rule "parlay when the hand's expectation is positive" from the expected-value calculator alone reproduces the published parlay strategy across all 78 cells where the player wins the war, with zero mismatches, including its one irregular entry, parlay a player 8 against a dealer 2.

A correction worth keeping: while scoping this I argued that the literal reading meant nobody should ever parlay. The calculation behind that was fine and the reading of the game was wrong, and a published parlay strategy table existing at all should have told me so before I computed anything.

3The oracle

Exact at any composition, in two pieces that behave nothing alike: one is settled by a symmetry argument, and the other needed a calculation nobody had reason to do before.

The comparison leg comes out of a symmetry argument that needs no calculation beyond counting ties. Both war cards come off the same shoe and nothing distinguishes which one is dealt first, so at any composition the chance your card beats the dealer's is exactly the chance the dealer's beats yours. Those two cancel, the dealer collects ties, and the leg's expectation is exactly minus the probability of a tie. On the fresh six-deck shoe that's 13 × 24 × 23 over 312 × 311, or −7.3955%.

The ace-low rule can't move this leg at all, which is easy to get backwards: relabelling ranks changes who wins a given pair and never changes how many pairs tie. The statistic a counter would have to track is the chance two cards drawn from the shoe share a rank, which is a sum of squared rank counts. That's a second moment, and it's smallest when the shoe is flat.

The parlay leg needed something I hadn't built yet: the expected value of a blackjack round given only your first card and the dealer's upcard, with the rest of the deal still unknown. That's the decision state of any bet resolved between the upcard and the rest of the deal, so it's reusable well beyond this game. Averaging it back over the deal distribution reproduces the ordinary game expectation to 2.3e-17. Put together, the bet's exact value at any shoe is the tie probability subtracted from the parlay option's value, where the option is exercised only when it helps.

4The state dataset

Ten million rounds, with the exact 13-rank composition read at the moment the bet must be posted. The comparison leg never once came within reach.

I ran ten million rounds across 232,865 shoes at 0.75 penetration, reading the exact rank composition at every round start, which is when the bet has to be on the felt. The comparison leg averaged −7.3952% with a standard deviation of 0.1349 points. Breakeven, meaning the point where the leg stops costing more than the parlay option brings in, sits at −6.2355%. The single best round out of ten million came in at −6.5823%, still 0.35 points short, and the gap to breakeven is 8.6 standard deviations. Zero rounds cleared it.

Penetration can't help, because the mean is pinned to the fresh-shoe value at every depth by the symmetry above. Only the spread grows, from nothing at the fresh shoe to 0.295 points with a deck and a half left, which is nowhere near enough to matter.

Neither count sees the leg, and the sign runs the wrong way regardless. Against the hi-lo true count the leg traces a shallow U, sitting at −7.385% around ±1 and sagging to −7.64% out at ±8, a slope of −0.00024 points per true count. The mechanism is the one the theory predicted before the run: a high count means a shoe that has drifted away from a flat rank distribution, a lumpier shoe produces more ties, and so the comparison leg is at its worst exactly where the parlay leg is at its best.

I put a prediction on paper before measuring and it scored two out of three. Direction and mechanism came out right. The spread I predicted, 0.34 points, was 2.5 times too wide, because 0.34 describes the deepest rounds of a shoe and shallow rounds dominate the average.

5The fit, and the count

No linear count fits this bet, and the correct statistic is second order. Escalating to it changes nothing, because the leg has no spread to exploit.

This is the page where the higher-order escalation gets considered first and abandoned fastest. A linear count is the wrong shape by construction: the comparison leg depends on the shoe only through the chance two cards share a rank, and that quantity is a sum of squared rank counts. No set of per-card tags can track it, since a tag system adds up cards one at a time and this statistic is about how they cluster.

So I went straight to the second moment, which is the same escalation the 21+3 page makes when squared suit surpluses turn out to carry the value. Here it arrives and buys nothing. The collision statistic is real and correct and would need its own dedicated tracking, and it moves the leg by two hundredths of a point across the entire observed range while breakeven is a full point away. A statistic can be the right one and still be worthless when the quantity it describes barely moves.

The parlay leg has no count of its own to derive, because it isn't a side bet in any meaningful sense. It's a blackjack hand, so it rides the ordinary hi-lo count that every counter already keeps, and the parlay's value climbs from +6.18% at a neutral shoe to +10.02% at a true count of +8. Nothing new needs learning to play it, and nothing new can be learned to make it better.

6Certified live

With no human system to certify, what got certified is the theorem. Ten million dealt pairs agree with it to a third of a standard error.

Every other page on this site certifies a card by replaying it on fresh shoes. There's no card here, so the certification went after the claim the whole verdict rests on. Over ten million actually dealt war pairs the players won 4,629,737 and lost 4,630,597, a difference of −0.000086 against a standard error of 0.000632, and the realized expectation came in at −0.074053 against the predicted −0.073952, a z of −0.32. The symmetry argument holds as a measured result here, on top of holding as algebra.

Three more checks hold up the count sweep. Depth invariance is proved exactly, by exhaustive enumeration over fractions, for flat and lopsided rank profiles alike. The composition model behind the true-count sweep reproduces the independently measured per-count tie curve to about one part in ten thousand, U shape included. And the first-card calculation averages back to the ordinary game expectation to 2.3e-17 on two different rulesets.

7The ledger

Played as well as it can be played, the bet makes about $7.65 per 100 rounds on a $100 wager, and every one of those dollars would have earned more on the hand.

The best available version of this play is to post the War bet only when the hi-lo true count is +3 or better, which is 11.10% of rounds. At $100 a wager and 200 rounds an hour that comes to about $15.30 an hour, or $7.65 per 100 rounds. Treat it as a floor, since the highest count bin in the measurement is clipped and gets priced here at the +8 rate even where the true count runs above it.

The same $100 on the blackjack hand beats it at every count in the sweep, and the margin grows: 0.37 points at a neutral shoe, 0.63 at +3, 1.11 at +8, and 2.72 out at +20. The reason is structural. The War bet drags a −7.4% comparison leg behind its parlay option, and the option is only ever a lever on the same underlying hand. You're paying a fixed 7.4% toll for a smaller share of the same edge.

This page has no bankroll figure, and it's the only one on the site that doesn't. I never measured the swing on this play, because the bet lost on expectation before its variance was worth the compute. Bankroll couldn't rescue it anyway: it already trails blackjack on edge while carrying a coin flip on the comparison leg stacked on top of blackjack-like parlay swings.

8The verdict

Not worth it, and the distinction from unbeatable is worth being precise about.

A player who counts and posts this bet only at a true count of +3 and above will make money on it. That's why the call isn't unbeatable, even though half of this page reads like it should be. The wager does turn positive, on 11% of rounds, using a count that's been in print since 1962.

The call is that no dollar belongs here. Every count at which the bet pays is a count at which the same dollar on the blackjack hand pays more, and the shortfall grows precisely as conditions improve. The one exception anyone can construct, extra capacity at the table maximum, is undone by the parlay adding to the very wager that's already capped.

One piece of this does earn the flat call, and it's the piece with the bet's name on it. The card comparison is immune to counting as a matter of proof, and ten million rounds confirmed it without producing a single exception. That leg is 84% of what the wager costs, and everything worth having in the rest of it belongs to the blackjack hand sitting next to it.

§What this verdict depends on

• Six decks, dealer hits soft 17, blackjack pays 3:2, double after split. Aces rank low in the war and the dealer wins ties.
• The parlay reading matters more than any other input: this page prices the version where you keep the war win whether or not you parlay. A floor that makes parlaying forfeit the win prices the whole wager at a flat −7.4% and the parlay strategy table stops making sense. The published strategy table matching on all 78 cells is the evidence for the reading used here.
• Whether a parlayed total is capped at the table maximum is unverified, and it decides the only argument for the bet's existence in an advantage player's game.
• The expected-value calculator's baseline runs slightly pessimistic against the published figure for the host game, so option values are conservative and the count sweep is reported calibrated onto the published fresh-shoe number. The calculator also doesn't remove the two known cards from the shoe weights.
• Variance and bankroll are not priced anywhere on this page.
• Cut-card shoes only. A continuous shuffling machine removes the count the parlay leg rides on, leaving the frozen −7.4% comparison leg with a smaller option on top.
• Nobody has confirmed whether the game is spread at any particular property.

Last updated 2026-08-10.