title: Statistical model for Genshin Impact's droprates
date: June 2021
author: Cgg
For a lighter version & discussion, see our infographic on Reddit.
Over the past few months, genshin-wishes.com has offered Genshin Impact players the chance to preserve, analyze and (sometimes) brag about their gacha rolls. A side goal of building such a database is to figure out the inner workings of Mihoyo's gacha system. We thus used our database (25 million pulls as of June 2021) to build a statistical model of what we assume is representative of Mihoyo's system.
In accordance with what's been observed by the community, we find a soft pity system that starts way before the 90-pull limit guaranteed by Mihoyo. Namely, we find that the pity starts at pull 74 (63 for the Weapon banner) and linearly boosts the odds until they reach a 100% chance at pull 90 (77 in the Weapon banner). A majority of 5* drops occur right after the pity system kicks in, so the most 'usual' pulls happen around the ~75th attempt.
Interestingly, we find a similar system for the 4* items. There is a 4* pity that starts on the 9th pull and reaches a 100% chance at pull 10. Both pities are independent - 5* drops don't reset the 4* pity.
Banners have their own independent pities, and pities on event banners don't reset when the pool of items is cycled (with new events).
Prior work we are aware of comes from a user in CN, who similarly based their work on contributed histories and came to generally compatible conclusions (but with some differing details).
5* droprate on the Permanent and Character Event banners
Our data is consistent with the 0.6% 'base rate' announced by Mihoyo. In other words, you have a 0.6% chance (1 in 167) with every pull. This, of course, only holds true until the pity system kicks in. After 73 unsuccessful attempts, the gacha will offer an additional 6% chance with every new pull. Thus, on the 74th attempt, your chances become a more favorable 6.6% (from just 0.6%). At pull 75, it is up 6% again, and becomes 12.6%. By turn 80, the probability has reached 42.6%. These increasing odds eventually reach their maximum of a 100% chance at turn 90 - where you are therefore guaranteed a 5*.
The probability you ever reach 90 is vanishingly small, however. In fact, that would be epic - please get in touch if it ever happens to you, you'll make the statisticians happy and we'll have a small party with root beer and ping pong. The increasingly better odds in pity make having to pull 90 times extremely unlikely. We calculated the probability at a staggeringly low 0.000007% per 5* item (that's 1 in 14 millions). Instead, the bulk of 5* drops occur within a few rounds of the start of the pity.
You have an overall 1 in 3 chance (35.6%) of pulling before pity. If you then consider the pity alone, separate from everything else, the 5 first rounds of pity alone have a 65.2% chance (~2 in 3). Altogether, 9 out of 10 (91.8%) 5* drops occur by the 80th pull. Pulling 84 or more times has less than a 1% chance, and anything beyond that increasingly becomes more improbable. The highest number of pulls we've observed is 89, and that happened exactly once in our data. Considering the rarity of these pulls, our data is admittedly limited.
The 'consolidated' droprate (i.e. your chances of getting a 5* with any pull, regardless of pity) is 1.60521%. Mihoyo guarantees 1.600%.

For a quick rundown of these graphs:
* cumulative probability: where a 5* might be pulled when you start from 0 (right after a 5* or fresh on your very first pull). There is a 35.6% chance of getting a 5* before the pity starts (at pull 74).
* per-pull probability: your chances with every individual pull. It is constant until pity territory, where the odds increase linearly up to 100%.
* distribution of successful pulls: where drops occur 'the most'. Because of the generous odds during pity, drops surge following the start of the pity.

NOTE: Notice (in the distribution) that before pity, more pulls happen early: more drops happen at pull 1, and pull 2, in turn, has more drops than pull 3. This happens even though the droprate is constant. The rationale for this behavior is that 0.6% of first pulls result in a drop, and these 0.6% will then be out. As a result, the second pull has less to work with. This is a property of what we call geometric distributions.

5* droprate on the Weapon Event banner
While built on the same model, the Weapon banner has more generous odds. First, its soft pity is triggered on the 63th pull (instead of the 74th). Then, the 'base rate' is 0.7% (up from 0.6%). Once under pity, every new pull adds an extra 7% chance - ten times the 'base rate'. Because of these better odds, it only takes 15 turns of pity to top 100%. So, at pull 77, your 5* is guaranteed. Again, though, players don't usually make it through the whole pity. The maximum number of pulls we've observed for this banner is 75.
We calculated the 'consolidated' droprate for this banner at 1.87792%. Mihoyo guarantees 1.85%.
The following graph compares the odds of the Weapon banner against the odds of the Permanent and Character banners (grey line). On the Weapon banner, 97.7% of 5* drops occur by the 70th pull.

4* droprates
The 4* items have easier parameters but some extra peculiarities. The pity for these items starts on the 9th pull. Like with 5* items, every round of pity adds 10 times the 'base rate' (of 5.1%). Consequently, the first round of pity has probability 5.1% + 51% = 56.1%. The second round then reaches 100% and guarantees the drop. There is a catch, though. In all situations, 5* drops have priority over 4*. If, say, you're on the 10th pull for a 4*, but a 5* comes out, you're up for an 11th pull on that 4* item - despite the guarantee.
Things get even more convoluted as there is no stopping the 5* items from turning up many times in a row - and delaying the 4* drop all the while. The number of pulls for a 4* has no theoretical limit. But because getting 5* items 10 times in a row is implausible enough, you can safely say that it won't happen.
We do observe this phenomenon in the database. So far, we've had 4* items at 11 and 12 pulls. The larger the dataset, the more apparent these rare events tend to be.
Not accounting for the interference from 5* items, the probability of pulling a 4* before pity is 34.2% (roughly 1 in 3). The 9th pull sees the majority of drops, but for simplification we can say the 9th and 10th pulls are roughly on par, and together take the lion's share of 4* drops.
We found through numerical analysis and simulation that the 'consolidated' droprate is about 13.0567% (Mihoyo guarantees 13%).
The weapon banner again has more favorable odds - the 'base rate' is 6% instead of 5.1%, and the pity starts 1 pull earlier (at pull 8). Every turn of pity then adds 60% instead of 51%. We found the 'consolidated' droprate to be 14.8568% (Mihoyo guarantees 14.5%).
Novice banner
The novice banner has odds similar to that of the Permanent banner. There is no 5* pity to speak of, as only a limited number of 20 pulls is offered. The banner has one notable quirk however: Noelle will always drop at pull 8. This suggests a hardcoded behavior by Mihoyo as nothing seems to avert it - your 8th pull on the Novice banner will always be Noelle. That forced drop resets the 4* pity like any other drop.
Discussion
Pulling the gacha in Genshin Impact could be construed as spinning a roulette, with items of different qualities occupying different areas on the wheel - 5* items would have a very small area, whereas 3* items would hold most of the space. As the pity system kicks in, these smaller - rarer - areas grow in size to the detriment of lower-priority item qualities. For instance, you could say that the 5* pity 'pushes out' the 3* items. This is corroborated in our data by the absence of 3* drops when both the 4* and 5* pities are engaged to certain levels. For instance, on the permanent banner, a 5* pity at 81 (48.6%) coupled with a 4* pity at 9 (56.1%) occupy the entire outcome space, and thus prevent the occurrence of any 3* drops.
Mihoyo's gacha system is rather simple but efficient at balancing the odds towards earlier drops. In all cases the system is more generous than actually announced. However the mechanics described here only cover the star rating of the drops. Which items are picked within the pool of available items is subject to yet another set of rules.
Pseudocode
_rate5 = .006 # 0.6%
_rate4 = .051 # 5.1%
_pity5 = 73 # 5* pity starts after 73 attempts
_pity4 = 8 # 4* pity starts after 8 attempts
# counter5 and counter4 are the (persistent) user state for this banner (initialize to 1)
x = random() # between 0.0 and 1.0
prob5 = _rate5 + max(0, (counter5-_pity5) * 10 * _rate5)
prob4 = _rate4 + max(0, (counter4-_pity4) * 10 * _rate4)
if (x < prob5) {
drop(5star)
counter5 = 1;
counter4 += 1
} else if (x < prob4 + prob5) {
drop(4star)
counter5 += 1;
counter4 = 1;
} else {
drop(3star)
counter5 += 1;
counter4 += 1;
}
History
2021-07-03: added missing consolidated droprates for 4* and fixed some issues with the pseudocode.

