The same figures as a 1080×1080 card, for posting. Click it to copy the image.Full size ↗
Matches counted813,994
Characters19
Clear of 50%5 of 19
Pairs told apart49 of 171
Rank bandRank 21–46
Period2026/03/26 – 2026/10/18
Tiers
S
Taka53.3%
El Blaze51.6%
A
Eileen51.4%
Lau51.3%
Kage51.1%
Jean51.0%
Jeffry51.0%
B
Brad50.5%
Lion50.5%
Lei Fei50.4%
Jacky49.9%
Pai49.8%
C
Sarah49.7%
Vanessa49.7%
Shun49.6%
Wolf49.4%
Goh49.1%
D
Akira48.4%
Aoi47.9%
SABCD
The letters sort characters by win rate; they are not a test. 49 of 171 character pairs can be told apart.
How sure is each rate?
Clear of 50%, aboveClear of 50%, belowNot clear of 50% after correction
Ranges count players, not matches: one person's games rise and fall together, so 813,994 matches from 5,583 players carry less evidence than 813,994 coin flips. They do not also allow for repeat opponents, so if anything they are too narrow. Coloured characters clear 50% after Benjamini–Hochberg correction across all 19. Each range is for one character alone. 10 of 19 miss 50% that way; with 19 tested at once the bar is higher, so only the coloured ones count.
Win rate before and after v1.10
Ranks 21–46, both players in one band · before: 2 Nov 2025 – 25 Mar 2026, 141,461 matches · after: 26 Mar 2026 – 18 Oct 2026, 869,835 matches
None of the 19 characters' win rates changed at v1.10 by more than chance. Closest: Lion, 45.3% → 50.5% (p = 0.033, q = 0.234).
Monthly rate95% range across players50% lineAverage before / afterv1.10 launch, 26 Mar 2026Part of a month
Where a month's range runs past 40–56%, the band stops at the edge; that point's tooltip gives the range in full.
The change is the average after v1.10 minus the average before, with its error counted per player, since one player's matches are not independent. p is the share of 1,000 random deals of the calendar days to the two periods (each keeping its number of days) that produced a change as large; the textbook p would have called 5.6% of those chance changes significant at 0.05. q corrects p across the 19 characters (Benjamini–Hochberg), and a character is marked where q < 0.05.
A change here happened at v1.10; it does not show the patch caused it. Players arrived and left around the launch, and a character's rate can move because different people play it.
In both periods a player's character is read from the opponent's record of the same match, since there is no openings pull before v1.10 or for the low band's players. On 1,545,822 v1.10 player-matches that the openings pull also covers, the two methods agree 100.00% of the time.
The same figures as a 1080×1080 card, for posting. Click it to copy the image.Full size ↗
Matches counted374,113
Characters19
Clear of 50%4 of 19
Pairs told apart26 of 171
Rank bandRank 21–29
Period2026/03/26 – 2026/10/18
Tiers
S
Jean53.2%
Taka52.9%
A
Jeffry51.5%
El Blaze51.4%
Brad51.0%
Lau50.8%
Jacky50.7%
B
Vanessa50.6%
Lei Fei50.5%
Wolf50.5%
Eileen50.4%
Kage50.3%
C
Shun50.0%
Sarah50.0%
Pai49.9%
Lion49.8%
Goh49.6%
D
Akira47.6%
Aoi46.6%
SABCD
The letters sort characters by win rate; they are not a test. 26 of 171 character pairs can be told apart.
How sure is each rate?
Clear of 50%, aboveClear of 50%, belowNot clear of 50% after correction
Ranges count players, not matches: one person's games rise and fall together, so 374,113 matches from 4,774 players carry less evidence than 374,113 coin flips. They do not also allow for repeat opponents, so if anything they are too narrow. Coloured characters clear 50% after Benjamini–Hochberg correction across all 19. Each range is for one character alone. 4 of 19 miss 50% that way; with 19 tested at once the bar is higher, so only the coloured ones count.
Win rate before and after v1.10
Ranks 21–29, both players in one band · before: 4 Nov 2025 – 23 Mar 2026, 66,536 matches · after: 26 Mar 2026 – 18 Oct 2026, 399,566 matches
None of the 19 characters' win rates changed at v1.10 by more than chance. Closest: Aoi, 53.1% → 46.6% (p = 0.007, q = 0.133).
Monthly rate95% range across players50% lineAverage before / afterv1.10 launch, 26 Mar 2026Part of a month
Where a month's range runs past 32–80%, the band stops at the edge; that point's tooltip gives the range in full.
The change is the average after v1.10 minus the average before, with its error counted per player, since one player's matches are not independent. p is the share of 1,000 random deals of the calendar days to the two periods (each keeping its number of days) that produced a change as large; the textbook p would have called 5.1% of those chance changes significant at 0.05. q corrects p across the 19 characters (Benjamini–Hochberg), and a character is marked where q < 0.05.
A change here happened at v1.10; it does not show the patch caused it. Players arrived and left around the launch, and a character's rate can move because different people play it.
In both periods a player's character is read from the opponent's record of the same match, since there is no openings pull before v1.10 or for the low band's players. On 1,545,822 v1.10 player-matches that the openings pull also covers, the two methods agree 100.00% of the time.
The same figures as a 1080×1080 card, for posting. Click it to copy the image.Full size ↗
Matches counted376,139
Characters19
Clear of 50%5 of 19
Pairs told apart42 of 171
Rank bandRank 30–39
Period2026/03/26 – 2026/10/18
Tiers
S
Taka53.6%
Eileen53.0%
A
Kage51.9%
El Blaze51.4%
Lion51.4%
Jeffry50.9%
Lau50.8%
B
Brad50.4%
Vanessa50.1%
Aoi50.0%
Sarah49.9%
Jean49.8%
C
Pai49.8%
Lei Fei49.4%
Shun49.4%
Jacky49.0%
Akira48.8%
D
Goh48.8%
Wolf48.1%
SABCD
The letters sort characters by win rate; they are not a test. 42 of 171 character pairs can be told apart.
How sure is each rate?
Clear of 50%, aboveClear of 50%, belowNot clear of 50% after correction
Ranges count players, not matches: one person's games rise and fall together, so 376,139 matches from 2,321 players carry less evidence than 376,139 coin flips. They do not also allow for repeat opponents, so if anything they are too narrow. Coloured characters clear 50% after Benjamini–Hochberg correction across all 19. Each range is for one character alone. 8 of 19 miss 50% that way; with 19 tested at once the bar is higher, so only the coloured ones count.
Win rate before and after v1.10
Ranks 30–39, both players in one band · before: 2 Nov 2025 – 23 Mar 2026, 58,754 matches · after: 26 Mar 2026 – 18 Oct 2026, 399,837 matches
None of the 19 characters' win rates changed at v1.10 by more than chance. Closest: Goh, 53.7% → 48.8% (p = 0.027, q = 0.513).
Monthly rate95% range across players50% lineAverage before / afterv1.10 launch, 26 Mar 2026Part of a month
Where a month's range runs past 40–60%, the band stops at the edge; that point's tooltip gives the range in full.
The change is the average after v1.10 minus the average before, with its error counted per player, since one player's matches are not independent. p is the share of 1,000 random deals of the calendar days to the two periods (each keeping its number of days) that produced a change as large; the textbook p would have called 6.2% of those chance changes significant at 0.05. q corrects p across the 19 characters (Benjamini–Hochberg), and a character is marked where q < 0.05.
A change here happened at v1.10; it does not show the patch caused it. Players arrived and left around the launch, and a character's rate can move because different people play it.
In both periods a player's character is read from the opponent's record of the same match, since there is no openings pull before v1.10 or for the low band's players. On 1,545,822 v1.10 player-matches that the openings pull also covers, the two methods agree 100.00% of the time.
The same figures as a 1080×1080 card, for posting. Click it to copy the image.Full size ↗
Matches counted63,742
Characters19
Clear of 50%1 of 19
Pairs told apart1 of 171
Rank bandRank 40–46
Period2026/03/26 – 2026/10/18
Tiers
S
Lei Fei54.6%
Lau54.3%
A
Taka53.3%
El Blaze53.1%
Akira52.0%
Eileen51.2%
Jacky51.0%
B
Kage49.6%
Lion49.1%
Brad49.0%
Aoi48.7%
Pai48.3%
C
Shun48.0%
Wolf47.9%
Goh47.7%
Jeffry46.6%
Jean46.2%
D
Sarah46.2%
Vanessa45.1%
SABCD
The letters sort characters by win rate; they are not a test. 1 of 171 character pairs can be told apart.
How sure is each rate?
Clear of 50%, aboveClear of 50%, belowNot clear of 50% after correction
Ranges count players, not matches: one person's games rise and fall together, so 63,742 matches from 462 players carry less evidence than 63,742 coin flips. They do not also allow for repeat opponents, so if anything they are too narrow. Coloured characters clear 50% after Benjamini–Hochberg correction across all 19. Each range is for one character alone. 3 of 19 miss 50% that way; with 19 tested at once the bar is higher, so only the coloured ones count.
Win rate before and after v1.10
Ranks 40–46, both players in one band · before: 2 Nov 2025 – 25 Mar 2026, 16,171 matches · after: 26 Mar 2026 – 18 Oct 2026, 70,432 matches
None of the 19 characters' win rates changed at v1.10 by more than chance. Closest: Wolf, 53.8% → 48.1% (p = 0.032, q = 0.272).
Monthly rate95% range across players50% lineAverage before / afterv1.10 launch, 26 Mar 2026Part of a month
Where a month's range runs past 28–72%, the band stops at the edge; that point's tooltip gives the range in full.
The change is the average after v1.10 minus the average before, with its error counted per player, since one player's matches are not independent. p is the share of 1,000 random deals of the calendar days to the two periods (each keeping its number of days) that produced a change as large; the textbook p would have called 7.9% of those chance changes significant at 0.05. q corrects p across the 19 characters (Benjamini–Hochberg), and a character is marked where q < 0.05.
A change here happened at v1.10; it does not show the patch caused it. Players arrived and left around the launch, and a character's rate can move because different people play it.
In both periods a player's character is read from the opponent's record of the same match, since there is no openings pull before v1.10 or for the low band's players. On 1,545,822 v1.10 player-matches that the openings pull also covers, the two methods agree 100.00% of the time.