Draw From the Discard Pile or the Stock? The Gin Rummy Rule of Thumb

Every turn in gin rummy starts with the same choice. Take the top card of the discard pile, the one everyone can see, or draw blind from the stock? You make it several times a hand, usually on feel. Most tips pages give the same answer: take it only if it completes a meld. Then they stop. No numbers, no weighing of the reasons, no tested exceptions.

So we measured it. In our bot simulation of 100,000 hands, only about 1 turn in 7 offered a face-up card that fit a meld. Below you'll find the one-question rule and what each option is worth in deadwood. After that, we cover what a pickup gives away, when to break the rule, and how it all applies against an app's computer opponent.

What's the rule for drawing from the discard pile or the stock?

The rule of thumb. Take the face-up card only when it goes straight into a meld: it completes a set or run, or extends one you already hold. Otherwise draw from the stock. In 100,000 simulated hands, a fitting discard cut deadwood by 16.3 points on average, against 4.9 for a blind draw (Atay Games bot simulation, September 2026).

Turn it into one question you can answer in a second: will this card be in a meld at the end of my turn? If yes, take it. If no, draw from the stock. There's one carve-out. Take a card that doesn't fit if it lets you knock this turn. The hand ends before anyone can use what they saw. It's rare. In our simulation, a non-fitting pickup allowed a knock in fewer than 1 draw in 1,000.

The draw decision in one flow

Shares from the Atay Games simulation, September 2026: 100,000 hands and 1,335,313 draw decisions by bots under EAAI research rules. A fitting face-up card appeared in 15.4% of decisions; when it didn't fit, taking it would have allowed a knock only 0.08% of the time.

The very first upcard of a hand works differently, because a pass gives your opponent the option. Our guide covers whether to take the very first upcard. New to the terms? The glossary defines the stock and discard pile.

If you count deadwood as you play, there's a sharper version of the rule, and it lands in the same place. A blind draw is worth about 5.5 points of deadwood in your first three draws. That drops to 4.1 in draws 4 to 7, and 2.9 from your 8th draw on. A discard is only worth taking if it beats that. Fitting discards cleared the bar 86% of the time. Non-fitting ones almost never did once the first upcard was gone.

Do gin rummy research bots follow the same rule?

Yes. Researchers who build gin rummy bots arrived at the same rule. Todd Neller's baseline player for the EAAI Gin Rummy challenge follows one line: "Draw face up card only if it becomes part of a meld. Draw face down card otherwise." Neller calls the whole player's policy "trivial, poor": it's the simple baseline supplied for the competition. Yet the draw rule is the part stronger agents kept.

Tonic is one of those agents, presented at AAAI 2021. Its authors wrote that "Domain experts often advocate only drawing the face up card if it completes a meld." Then they added narrow exceptions, covered below.

Why does a discard that fits a meld beat a blind draw?

A fitting discard is a sure thing. A face-up card that fits a meld is certain. A blind stock card lands in a meld only about 1 time in 5. Across 205,507 simulated decisions where the discard fit, taking it cut deadwood by 16.3 points on average, against 4.9 for drawing blind (Atay Games bot simulation, September 2026).

The gap holds at every stage of the hand. In your first three draws, a fitting pickup cut deadwood by 20.3 points, against 5.9 for a blind draw. In draws 4 to 7 it was 13.5 against 4.2. From the 8th draw on, it was 10.1 against 3.1. Late in the hand, the difference turns into knocks. A fitting pickup let the player knock that same turn 49% of the time; a blind draw, 13%.

The gap is wide for a simple reason. A card that fits costs nothing to hold. When it completes a meld, it wipes out the deadwood of the cards it joins, and you still get to throw your worst remaining card. A blind card rarely does that. When it helps, it usually just swaps a high card for a lower one.

Average deadwood cut per draw: take the face-up card or draw blind

Atay Games simulation, September 2026: 100,000 hands, 1,335,313 draw decisions, bots under EAAI research rules (not Gin Rummy Cash's settings). Values: fits, take 16.3 vs blind 4.9; doesn't fit, take −0.4 vs blind 4.8.

There's a caveat. In about 14% of fitting cases, taking the card still didn't beat the average blind draw. A card can fit a meld and leave you no better off. That happens, for example, when using it breaks up another meld you hold. That's why the test is "will it be in a meld at the end of my turn," not "does it match anything in my hand."

Over whole games, the habit adds up. In our identical-deck bot test, a bot that never takes the face-up card lost 91.4% of 100,000 games. Its opponent was an otherwise identical bot that takes the card when it fits. Our rerun for this post, on new seeds, matched it at 91.5%.

Why draw blind when the face-up card doesn't fit?

You can't throw it back. Pagat's rules say that if you take the top discard, you "must discard a different card." So a card that doesn't fit is stuck in your hand for a turn. Taking a non-fitting discard left deadwood 0.4 points worse on average, while a blind draw improved it by 4.8 (Atay Games bot simulation, September 2026).

That rule does most of the work. Pick up a card that doesn't fit, and it sits in your hand as fresh deadwood while you throw something else. Our guide to the draw-then-discard rules covers the mechanics, including the rare tournament exception.

The stock is also better than it feels. A blind card improved the hand 62% of the time once the player made their best discard. Mostly, it replaced a high card with a lower one. It landed in a meld 19% of the time. Those odds shrink as the hand goes on, from 69% early to 57% mid-hand and 46% late. By then there's less high deadwood left to replace.

Chance a blind stock card improves your hand, by stage of the hand

Atay Games simulation, September 2026: 1,335,313 draw decisions from 100,000 hands, bots under EAAI research rules. "Improves" means the stock card lowers deadwood after the best legal discard. Draw numbers count each player's own draws.

So does a non-fitting card ever beat a blind draw? Almost never. After the first upcard, it beat the average blind draw in 0.36% of decisions, 3,537 out of 978,989. It beat it by more than 2 points in just 51, and never by more than 5. That's why the rule needs no "take it if it cuts your deadwood" clause.

Strategy guides add an argument we didn't measure. GameColony's gin guide makes the usual information point, "It is generally better to draw a deck card than an upcard because less of your hand will be known to your opponent." Then it adds that "drawing from deck you may be preventing him from going for a gin hand." The upcard, after all, is a card your opponent already threw away. Treat that second point as reasoning, not data.

What does picking up a discard tell your opponent?

One card for certain, and a direction. A pickup tells your opponent one card you hold, and that it sits in or next to a meld. In 2021 AAAI research, guessing an opponent's hand from the public record of discards and pickups got about 33% of its guesses right. A blind guess got about 20% (Hein et al., 2021).

Under the fit-only rule, our bots' pickups joined a set 49% of the time and a run 47% of the time. The other 4% could go either way. So your opponent learns one card for sure and a coin flip on which kind of meld it's in. A careful reader has to guard both: the same rank, and the suited cards on either side.

What your pickup reveals

Atay Games simulation, September 2026: fit-only bots over 100,000 hands under EAAI research rules. Pickup joined a set only 49.2%, a run only 46.8%, either 4.1%. Our reading bot treated the same rank, or the same suit within 2 ranks, as off limits whenever that cost it 3 deadwood points or less: the pattern shown in pink.

Penn State researchers who built a gin agent spelled out what an opponent can do with that. Drawing face up "can indicate to the opponent what cards the agent is collecting. This information can help the opponent avoid future discards that might be added to a meld, as well as allow them to collect cards that can be laid off on the agent's melds" (Goldstein et al., AAAI 2021). That second part matters if you knock without gin, because it shows them where they can lay off.

How big is the leak?

Small but measurable. Call a discard "connected" if it shares a rank with a card the player had picked up, or sits within 2 ranks of it in the same suit. Against a bot that ignored pickups, 17.9% of the opponent's later discards were connected. About 0.15 of those per pickup, roughly 1 in every 6.5 pickups, went straight into the player's hand. A bot that read pickups cut connected discards to 8.2%, and cut those fed-and-taken cards by 64%.

A blind draw, by contrast, tells your opponent almost nothing. The one clean signal is the pass itself: the top discard didn't fit. Reading their takes and passes is the mirror image of this post. Our advanced guide covers reading their pickups and counting the pile, and our cash-game tactics cover tracking what they throw.

Is hiding your hand worth more than the cards?

The cards decide most of it. In 100,000 duplicate-dealt games per matchup, a bot that also took non-fitting cards whenever they beat the average blind draw won 50.1% against a fit-only bot. Against a pickup-reading bot it won 51.9%, versus 52.1% for the fit-only bot itself. Information is the tiebreaker (Atay Games bot simulation, September 2026).

The tournament worked like this. Every deck was dealt twice with the seats swapped, so each bot played the other's cards. Never taking the face-up card lost 91.5% of games. Adding the knock-now carve-out changed nothing measurable, at 50.1%, because it came up so rarely. Adding loose swaps, about one extra pickup every 7 hands, also landed at 50.1%.

That loose-swap bot also made fewer gins than the fit-only bot (10,021 against 10,499) and was undercut more often (22,212 against 21,567). That fits the idea that a useless card clogs the hand.

The surprise came from the reader. Our pickup-reading bot refused to throw any card near a card its opponent had picked up. It held back whenever avoiding the throw cost 3 deadwood points or less. It won just 47.9% against an identical bot that ignored pickups. In our model, holding a live high card "because they might want it" cost more deadwood than the information saved.

What can't this simulation tell you?

These are bots, not people. The reader is deliberately crude, and a strong human may get more out of a pickup than it did. The bots judge cards by immediate deadwood, a short-sighted yardstick that ignores future outs, and they don't speculate. The face-up cards they saw depend on what the other bot threw, usually its highest deadwood card, so mid-hand discards skew high. And the rules are the research version, not Gin Rummy Cash's settings, which we haven't verified.

When should you break the rule?

Rarely, and only for a reason you can name. Take a non-fitting discard if it lets you knock this turn; the hand ends before the read matters. In its first two turns, the Tonic agent takes non-face cards that form a pair. Stretching that window past three turns cut its performance by more than 10% (Larkin et al., AAAI 2021).

Here are the exceptions worth knowing, each labeled with how much evidence stands behind it.

• Knock now (tested). Our knock-now bot won 50.1% against the fit-only bot (95% CI 49.9% to 50.4%). It's harmless, and it's rare.
• Early pair speculation (research-backed, not tested by us). Tonic takes the face-up card in its first two turns if it forms a pair. It does so only while the hand has many unmelded cards and few pairs. It never does this for face cards, because "exploring the deck is more productive than forming high-deadwood pairs."
• Late defensive swap (strategy-guide advice, not tested). GameColony: "When you suspect an imminent knocking from your opponent (far into the deck > 2/3) it may be worth taking a low upcard and throw any high-value deadwood." RummyTalk's Mark Kotsckowski is stricter. Take a late discard when "the discard puts you under a count," meaning low enough to knock, and never just to get closer. Our bots throw high cards mid-hand, so our data can't judge this one fairly.
• Deception pickups (strategy-guide advice, not tested). Gin Rummy Palace says drawing to confuse is "useless against beginners and offensive players who don't keep track of what you pick up or discard." Try it only against someone who watches closely, and only when your hand can afford a dead card.
• The first upcard (its own decision). Card Games For Dummies says to be prepared to take it "even if it has no relevance to your hand." Most guides say take it only if it melds. Our opening-hand guide works through it.

Notice what isn't on the list: "late in the hand, take anything that cuts deadwood." Our loose-swap bot took any non-fitting card that beat the average blind draw, at every stage of the hand, and gained nothing.

Does the rule change in a same-deck app like Gin Rummy Cash?

No, and it gets simpler. Gin Rummy Cash's App Store listing says both players "are dealt the same deck and play against the same computer opponent," and the higher score wins. The deadwood math behind the rule doesn't depend on who's watching, so it holds whatever that computer opponent does with your pickups.

Here's what we don't know. The game hasn't published how its computer opponent uses your pickups, so we won't guess whether it reads them. It doesn't change the advice. A card that doesn't fit is a bad trade even when nobody is watching.

Your competitor is a different story. As the listing describes the format, they play their own copy of the same deal against the same computer opponent. Then your points are compared. (Here's how competitors are matched.) So the classic "don't tip off the player across the table" argument applies mostly to home games and head-to-head apps. In a same-deck match, the deadwood half of the rule does the work.

Where does the draw decision bite in the app? Skillz, operated by Firy Inc (NYSE: FIRY, formerly SKLZ), runs the matchmaking and prizes. Its developer docs say random numbers "are in the same sequence for each player in a match." They add: "Over time, the player's decisions may diverge and the same random number will not be used in the same place" (Skillz Random and Fairness). So once you and your competitor make a different draw choice, your games stop matching.

To practice, drill the one-question test in free games until it's automatic. Then try a $1 match in Gin Rummy Cash. It's on iPhone and iPad through the App Store, and on Android as an APK from the official Atay site. Here's why Android installs from our site. Real-money play is 18+ and depends on your location, so check whether prize play is available where you live. Keep expectations realistic, too: a skill edge is worth a few dollars, not an income. For formats and prizes, see how real-money gin rummy works.

Frequently Asked Questions

What should you do on your next turn?

• Ask one question: will this card be in a meld at the end of my turn? Yes, take it. No, draw from the stock.
• Know the benchmark: a blind draw is worth about 5, then 4, then 3 points of deadwood as the hand goes on. A fitting discard is worth about 16.
• Keep the leak in proportion: a pickup gives away one card and a meld type. That's real, but deadwood decides most turns.
• Break the rule only for a reason you can name. Knocking this turn is the carve-out we tested: it cost nothing, and the hand ends before anyone can use the read.
• Against an app's computer opponent, the rule doesn't change.

Next, see what to keep and throw in your first three turns. Then learn to read the other side of the pile in the advanced guide. When your opponent knocks first, the undercut playbook takes over. Ready to test the rule? Practice free, then play a $1 match in Gin Rummy Cash.

A note on results. Skill-based cash games involve real money. Better decisions raise your odds over many hands, but they never guarantee a single outcome, and you can lose entry fees. The simulation figures in this post come from bots under research rules and are not a prediction of your results. Set a daily limit, never wager what you cannot afford to lose, and use the responsible play tools available in-app. Real-money play is 18+ and is not available in every jurisdiction.