Why a 1-in-128 drop does not arrive in 128 kills
A drop rate is a per-attempt probability, not a schedule. If an item drops at 1 in 128, each kill is an independent roll at 0.78125%, and the dice keep no record of the 127 kills that came before. The chance of surviving 128 rolls with nothing is (127/128)128 = 0.3664, so 36.6% of players are still empty-handed at the point the rate seemed to promise a drop, and only 63.4% have one.
The number 63.4% is not a coincidence of that particular rate. For any small p, running exactly 1/p attempts gives a success chance approaching 1 - 1/e = 63.21%. At 1 in 50 it is 63.58%; at 1 in 1,000 it is 63.23%. The rate barely matters, which means the intuition "I have done the expected number of attempts, so I am due" is wrong in a completely predictable way at every rate.
Two distributions do all the work here. The binomial distribution answers questions about a fixed number of attempts: how likely am I to see exactly two copies in 300 runs. The geometric distribution answers questions about waiting: how many attempts until the first drop. They are two views of the same independent-trials model, and both are memoryless in the sense that matters - after 500 dry attempts, the chance of a drop on attempt 501 is still exactly p.
The four formulas and what each answers
At least one in n attempts. Compute the failure path and subtract: the chance of missing every attempt is (1 - p)n, so the chance of at least one is 1 - (1 - p)n. Working through failure is not a trick, it is the only tractable route, because "at least one" covers one, two, three and more drops, while "none" is a single case.
Exactly k in n attempts. This is the binomial term C(n,k) pk (1-p)n-k. The binomial coefficient C(n,k) counts the orderings - which particular attempts dropped - while pk and (1-p)n-k price one specific ordering. Sum the terms from k upward for "at least k copies", which is the question that matters when you need four of the same component.
Average attempts per drop. The expected number of trials to the first success in a geometric distribution is 1/p, so a 1-in-128 item takes 128 attempts on average. That average is dragged upward by a long tail: a minority of players who go 400 or 600 attempts pull the mean well above the typical experience.
Attempts for a target confidence. Set 1 - (1-p)n = C and solve for n: n = ln(1 - C) / ln(1 - p). Both logarithms are negative, so the result is positive. This is the figure to plan around, because it converts a vague "probably" into a countable number of runs.
The gap between the mean and the median is the single most useful thing on this page. At 1 in 128 the mean is 128 attempts but the median is 89 - half of all players get the drop by attempt 89. If you feel unlucky at 130 attempts, you are, but only mildly: about 36% of players are in the same position.
Worked example: hunting a 1-in-128 item for 150 kills
Take a drop rate of 1 in 128 and 150 kills completed, and work every figure by hand.
- Per-attempt probability. p = 1 / 128 = 0.0078125, or 0.78125%.
- Failure probability. 1 - p = 0.9921875.
- Failure across all 150. ln(0.9921875) = -0.00784318. Multiply by 150: -1.176477. Exponentiate: e-1.176477 = 0.308363. So there is a 30.84% chance of 150 kills producing nothing.
- At least one. 1 - 0.308363 = 0.691637, or 69.16%.
- Expected drops. 150 x 0.0078125 = 1.172 items on average.
- Exactly one. C(150,1) p (1-p)149 = 150 x 0.0078125 x 0.310789 = 0.364209, so 36.4% end with exactly one - more likely than any other single count.
- 90% confidence. n = ln(0.10) / ln(0.9921875) = -2.302585 / -0.00784318 = 293.6, so 294 kills.
- 99% confidence. n = ln(0.01) / ln(0.9921875) = -4.605170 / -0.00784318 = 587.2, so 588 kills.
Read the shape of that: getting from 69% to 90% costs another 144 kills, and getting from 90% to 99% costs 294 more. Confidence is bought at an exponentially rising price, which is why "one more run" is a poor stopping rule and a pre-committed attempt budget is a good one.
How to tell bad luck from a bad assumption
Use the chance-of-nothing figure as your luck percentile. If your attempt count leaves a 30% chance of nothing, then roughly 30% of players doing the same grind are where you are, and you are unremarkable. Once that figure falls below 5% you are in the unluckiest twentieth; below 1% and you are in the unluckiest hundredth. Those are the only three thresholds worth remembering, and note that at any large drop-hunting community, thousands of players are grinding, so somebody is always in the 1% tail - that person is not evidence of a broken drop table.
If your figure falls below about 0.1%, question the model rather than your luck. The usual explanations are mundane: the wiki rate is for a different difficulty tier, the item requires a condition you are not meeting, the rate is per-chest rather than per-kill, or the drop is gated behind a quest or a threshold contribution. A genuinely wrong random number generator is far rarer than a misread drop table.
Watch for mechanics that break independence outright. Pity timers, bad-luck protection and guaranteed-after-N systems all mean your effective chance rises with each failure, so these formulas understate your position - for gacha-style banners with a hard pity, use the gacha pull probability calculator instead. Bundled rolls, where a single kill rolls the table several times, mean the effective per-kill rate is higher than the listed one. And a rate quoted as "1 in 128 per hour" is not a per-kill rate at all.
Attempts needed by drop rate and confidence
| Drop rate | Median (50%) | 1/p attempts, giving | 90% | 95% | 99% |
|---|---|---|---|---|---|
| 1 in 10 | 7 | 10 → 65.1% | 22 | 29 | 44 |
| 1 in 25 | 17 | 25 → 64.0% | 57 | 74 | 113 |
| 1 in 50 | 35 | 50 → 63.6% | 114 | 149 | 228 |
| 1 in 100 | 69 | 100 → 63.4% | 230 | 299 | 459 |
| 1 in 128 | 89 | 128 → 63.4% | 294 | 382 | 588 |
| 1 in 256 | 178 | 256 → 63.3% | 589 | 766 | 1,177 |
| 1 in 512 | 355 | 512 → 63.2% | 1,178 | 1,533 | 2,356 |
| 1 in 1,000 | 693 | 1,000 → 63.2% | 2,302 | 2,995 | 4,603 |
| 1 in 5,000 | 3,466 | 5,000 → 63.2% | 11,512 | 14,978 | 23,024 |
Notice that the 99% column is consistently about 4.6 times the drop rate denominator, and the median is about 0.69 times it. Those ratios are ln(100) and ln(2), and they hold for any small drop rate.
Assumptions and where they break
- Independence. Every attempt is a fresh roll. Pity counters, bad-luck protection and duplicate protection all violate this, and always in the player's favour, so these figures become a pessimistic floor.
- A constant rate. Rate-up banners, difficulty scaling, luck stats, drop-rate buffs and level-dependent tables all move p during the grind. Split the grind into segments at a fixed rate and multiply the failure probabilities.
- One roll per attempt. If a kill rolls the table three times, the effective per-kill chance is 1 - (1-p)^3, not 3p - close for small p, but not the same number.
- The listed rate is right. Community-datamined rates are usually accurate; community-estimated rates from small samples are often not. A rate inferred from 200 observations has a wide confidence interval.
- Rates are per-attempt, not per-hour. Convert time-based rates to attempts before using them, or the answer will be out by whatever your kill rate happens to be.
- No cap on copies. Some tables stop dropping an item once you own it, or convert extras to currency. That changes what "exactly k" means but not the chance of at least one.
The gambler's fallacy in loot form
You are never "due". After 500 failed attempts at a 1-in-128 item, the chance on attempt 501 is still 0.78125%, unchanged. What does change is the distribution of the total: 500 dry attempts is a genuinely rare position to be in, but rarity in the past buys nothing in the future. The correct use of these numbers is forward-looking - deciding how many more attempts to budget - not backward-looking justification for continuing. If the expected time to reach 90% confidence from here is longer than you want to spend, that is the real answer, and it does not depend on how long you have already spent.
Related probability tools
The binomial model here covers sampling with replacement, which is what a loot table is: the table does not shrink as you roll it. When the pool does shrink - drawing cards from a deck, pulling from a fixed prize pool without replacement - the correct model is hypergeometric instead, which is what a lottery odds calculator uses and what governs card-draw questions in deck-building games. The two agree closely when the pool is large relative to the draws and diverge sharply when it is not.
For dice-based systems the same independent-trials logic applies with much larger per-attempt probabilities, where the full distribution is more interesting than the tail: the dice probability calculator gives exact sum distributions and advantage odds. For card equity in poker, where the unseen cards form a shrinking pool, the outs and equity calculator applies the hypergeometric version. And once you have the item, the DPS calculator answers whether the upgrade was worth the grind.
One last framing that helps at very low rates. When p is small and n is large, the number of drops is close to Poisson with mean np, so the chance of nothing is approximately e-np. At 1 in 128 over 150 attempts that approximation gives e-1.171875 = 0.30979 against the exact 0.30836 - accurate to about a tenth of a percentage point, and easy to do in your head with a calculator that has an exponential key. It is the fastest sanity check available when a wiki quotes a figure that looks wrong.