the average race Average victory points at each fraction of game progress, all 50 games. The gold band spans the strongest to weakest seat at every step.
0 5 10 avg VPs start half game over gemini-3.7-flash gemini-3.7-flash 5.7 deepseek-v4-flash deepseek-v4-flash 6.94 gpt-5.6-luna gpt-5.6-luna 7.42 hy3 hy3 5.04 won from behind Winners who trailed badly mid-game, worst dip first.
0 .5 1 29% of leader pace at 37% through the game start game over share of leader deepseek-v4-flash m020_g0 fell to 29% of leader pace - won with 10 VPs
0 .5 1 50% of leader pace at 2% through the game start game over share of leader deepseek-v4-flash m006_g0 fell to 50% of leader pace - won with 10 VPs
0 .5 1 50% of leader pace at 37% through the game start game over share of leader deepseek-v4-flash m029_g0 fell to 50% of leader pace - won with 10 VPs
how close games were Gap between winner and runner-up across all 50 games - most titles were decided late.
4 8 12 16 margin of 1 VPs: 7 games 7 1 margin of 2 VPs: 16 games 16 2 margin of 3 VPs: 9 games 9 3 margin of 4 VPs: 5 games 5 4 margin of 5 VPs: 6 games 6 5 margin of 6 VPs: 6 games 6 6 margin of 7 VPs: 1 games 1 7 median margin 3 VPs median 3 winner's margin in VPs game length by winner How long each model's wins took - gold tick marks the row median.
100 200 300 400 turns deepseek-v4-flash deepseek-v4-flash won in 102 turns deepseek-v4-flash won in 105 turns deepseek-v4-flash won in 109 turns deepseek-v4-flash won in 110 turns deepseek-v4-flash won in 139 turns deepseek-v4-flash won in 146 turns deepseek-v4-flash won in 151 turns deepseek-v4-flash won in 160 turns deepseek-v4-flash won in 180 turns deepseek-v4-flash won in 183 turns deepseek-v4-flash won in 184 turns deepseek-v4-flash won in 187 turns deepseek-v4-flash won in 198 turns deepseek-v4-flash won in 207 turns deepseek-v4-flash won in 214 turns deepseek-v4-flash won in 218 turns deepseek-v4-flash won in 241 turns deepseek-v4-flash won in 248 turns deepseek-v4-flash won in 355 turns deepseek-v4-flash median 183 turns med 183 gpt-5.6-luna gpt-5.6-luna won in 70 turns gpt-5.6-luna won in 73 turns gpt-5.6-luna won in 91 turns gpt-5.6-luna won in 100 turns gpt-5.6-luna won in 100 turns gpt-5.6-luna won in 106 turns gpt-5.6-luna won in 113 turns gpt-5.6-luna won in 127 turns gpt-5.6-luna won in 128 turns gpt-5.6-luna won in 129 turns gpt-5.6-luna won in 132 turns gpt-5.6-luna won in 137 turns gpt-5.6-luna won in 147 turns gpt-5.6-luna won in 150 turns gpt-5.6-luna won in 154 turns gpt-5.6-luna won in 157 turns gpt-5.6-luna won in 165 turns gpt-5.6-luna won in 204 turns gpt-5.6-luna won in 214 turns gpt-5.6-luna median 129 turns med 129 gemini-3.7-flash gemini-3.7-flash won in 58 turns gemini-3.7-flash won in 131 turns gemini-3.7-flash won in 156 turns gemini-3.7-flash won in 194 turns gemini-3.7-flash won in 194 turns gemini-3.7-flash won in 212 turns gemini-3.7-flash won in 234 turns gemini-3.7-flash median 194 turns med 194 hy3 hy3 won in 114 turns hy3 won in 122 turns hy3 won in 134 turns hy3 won in 143 turns hy3 won in 189 turns hy3 median 134 turns med 134 robber diplomacy — the leader gets robbed deepseek gpt-luna gemini hy3 deepseek gpt-luna gemini hy3 a model never robs itself – deepseek-v4-flash robbed gpt-5.6-luna 190 times 190 deepseek-v4-flash robbed gemini-3.7-flash 152 times 152 deepseek-v4-flash robbed hy3 112 times 112 gpt-5.6-luna robbed deepseek-v4-flash 240 times 240 a model never robs itself – gpt-5.6-luna robbed gemini-3.7-flash 159 times 159 gpt-5.6-luna robbed hy3 117 times 117 gemini-3.7-flash robbed deepseek-v4-flash 180 times 180 gemini-3.7-flash robbed gpt-5.6-luna 174 times 174 a model never robs itself – gemini-3.7-flash robbed hy3 83 times 83 hy3 robbed deepseek-v4-flash 168 times 168 hy3 robbed gpt-5.6-luna 137 times 137 hy3 robbed gemini-3.7-flash 101 times 101 a model never robs itself – everyone's favourite target is a co-leader — 60% of all robberies land on deepseek or gpt-luna. Gold outline marks each robber's top victim.
play styles — four different brains ROBBER DEV KNIGHTS TRADES ACTIVE deepseek-v4-flash 19% 5.32 2.64 13.1 95% 2416 moves deepseek-v4-flash ROBBER DEV KNIGHTS TRADES ACTIVE gpt-5.6-luna 20% 7.04 3.92 12.98 97% 2535 moves gpt-5.6-luna ROBBER DEV KNIGHTS TRADES ACTIVE gemini-3.7-flash 18% 4.56 2.72 10 84% 2366 moves gemini-3.7-flash ROBBER DEV KNIGHTS TRADES ACTIVE hy3 19% 3.36 1.74 10.78 81% 2167 moves hy3 spokes normalized to the best model per axis, raw values on the vertices. luna buys dev cards and knights, deepseek deals the most, hy3 sits on its hands — passing a fifth of its turns.
what they trade — wheat out, ore in GIVE GET deepseek-v4-flash gave wheat away 161 times deepseek-v4-flash took wheat 137 times deepseek-v4-flash gave sheep away 114 times deepseek-v4-flash took sheep 28 times deepseek-v4-flash gave wood away 126 times deepseek-v4-flash took wood 51 times deepseek-v4-flash gave ore away 82 times deepseek-v4-flash took ore 188 times deepseek-v4-flash gave brick away 54 times deepseek-v4-flash took brick 133 times gpt-5.6-luna gave wheat away 115 times gpt-5.6-luna took wheat 153 times gpt-5.6-luna gave sheep away 146 times gpt-5.6-luna took sheep 16 times gpt-5.6-luna gave wood away 77 times gpt-5.6-luna took wood 31 times gpt-5.6-luna gave ore away 54 times gpt-5.6-luna took ore 217 times gpt-5.6-luna gave brick away 70 times gpt-5.6-luna took brick 45 times gemini-3.7-flash gave wheat away 89 times gemini-3.7-flash took wheat 83 times gemini-3.7-flash gave sheep away 43 times gemini-3.7-flash took sheep 34 times gemini-3.7-flash gave wood away 80 times gemini-3.7-flash took wood 49 times gemini-3.7-flash gave ore away 73 times gemini-3.7-flash took ore 83 times gemini-3.7-flash gave brick away 44 times gemini-3.7-flash took brick 80 times hy3 gave wheat away 47 times hy3 took wheat 104 times hy3 gave sheep away 71 times hy3 took sheep 6 times hy3 gave wood away 62 times hy3 took wood 24 times hy3 gave ore away 50 times hy3 took ore 125 times hy3 gave brick away 63 times hy3 took brick 34 times all models gave wheat away 412 times wheat 412 all models took wheat 477 times wheat 477 all models gave sheep away 374 times sheep 374 all models took sheep 84 times sheep 84 all models gave wood away 345 times wood 345 all models took wood 155 times wood 155 all models gave ore away 259 times ore 259 all models took ore 613 times ore 613 all models gave brick away 231 times brick 231 all models took brick 292 times brick 292 deepseek-v4-flash: 537 trades deepseek gpt-5.6-luna: 462 trades gpt-luna gemini-3.7-flash: 329 trades gemini hy3: 293 trades hy3 wheat leaves (412 deals) and ore arrives (613) — everyone runs the same engine: sell wheat and sheep, hoard ore for dev cards.
where they settled — the same golden corners chosen 31 times chosen 6 times chosen 9 times chosen 2 times chosen 8 times chosen 2 times chosen 24 times chosen 2 times chosen 25 times chosen 7 times chosen 16 times chosen 6 times chosen 8 times chosen 9 times chosen 20 times chosen 4 times chosen 19 times chosen 5 times chosen 6 times chosen 8 times chosen 18 times chosen 3 times chosen 16 times chosen 32 times chosen 2 times chosen 4 times chosen 15 times chosen 7 times chosen 1 times chosen 11 times chosen 8 times chosen 4 times chosen 22 times chosen 1 times chosen 10 times chosen 2 times chosen 1 times chosen 7 times chosen 1 times chosen 3 times chosen 7 times chosen 3 times chosen 1 times chosen 1 times chosen 1 times chosen 2 times every opening settlement from all 50 games overlaid on the canonical board - the models never saw it
value for money more wins / less $ per win 0.00 0.10 0.20 0.30 0 10 20 30 40 50 25% = coin flip mean $0.16/win win rate % deepseek-v4-flash: 38% wins at $0.067/win - $1.27 total spend deepseek-v4-flash gpt-5.6-luna: 38% wins at $0.064/win - $1.23 total spend gpt-5.6-luna gemini-3.7-flash: 14% wins at $0.204/win - $1.43 total spend gemini-3.7-flash hy3: 10% wins at $0.324/win - $1.62 total spend hy3 Same four opponents, same 50 games — yet a win costs $0.064 for gpt-5.6-luna and $0.324 for hy3: a 5.1x spread. The two winners sit inside the gold corner; everyone else paid more to lose more.
openings lined up with the standings 0.05 0.10 0.15 0.20 deepseek-v4-flash: mean 0.2053, median 0.194 cards/turn over 100 settlements deepseek-v4-flash 0.2053 deepseek · 19 wins gpt-5.6-luna: mean 0.1936, median 0.194 cards/turn over 100 settlements gpt-5.6-luna 0.1936 gpt · 19 wins gemini-3.7-flash: mean 0.1586, median 0.139 cards/turn over 100 settlements gemini-3.7-flash 0.1586 gemini · 7 wins hy3: mean 0.1586, median 0.139 cards/turn over 100 settlements hy3 0.1586 hy3 · 5 wins Opening quality is pip count: each settlement's tiles weighted by their dice-roll probability, averaged over both initial settlements. The order happens to match the standings — but that is four data points, and it does not hold under a different slice: counting only each model's first settlement, gemini (72% on the high-yield nodes) outranks luna (70%) while finishing twelve wins behind. Read this as a description of how they drafted, not as a cause of who won.
fair dice 0% 4% 8% 12% 16% 2: observed 2.50% vs expected 2.78% 2: expected 2.78% 2 3: observed 5.54% vs expected 5.56% 3: expected 5.56% 3 4: observed 8.95% vs expected 8.33% 4: expected 8.33% 4 5: observed 11.39% vs expected 11.11% 5: expected 11.11% 5 6: observed 13.78% vs expected 13.89% 6: expected 13.89% 6 8: observed 13.46% vs expected 13.89% 8: expected 13.89% 8 9: observed 10.81% vs expected 11.11% 9: expected 11.11% 9 10: observed 8.13% vs expected 8.33% 10: expected 8.33% 10 11: observed 5.89% vs expected 5.56% 11: expected 5.56% 11 12: observed 2.72% vs expected 2.78% 12: expected 2.78% 12 observed expected two-dice sum (7 excluded: it moves the robber) chi² 9.7 on df 9 (95% crit 16.9) over 7,561 rolls — fair dice. No model can blame the board.
who kept stalling deepseek-v4-flash minute 0: 52 retries deepseek-v4-flash minute 1: 29 retries deepseek-v4-flash minute 2: 39 retries deepseek-v4-flash minute 3: 54 retries deepseek-v4-flash minute 4: 51 retries deepseek-v4-flash minute 5: 40 retries deepseek-v4-flash minute 6: 46 retries deepseek-v4-flash minute 7: 41 retries deepseek-v4-flash minute 8: 39 retries deepseek-v4-flash minute 9: 39 retries deepseek-v4-flash minute 10: 44 retries deepseek-v4-flash minute 11: 40 retries deepseek-v4-flash minute 12: 47 retries deepseek-v4-flash minute 13: 41 retries deepseek-v4-flash minute 14: 38 retries deepseek-v4-flash minute 15: 40 retries deepseek-v4-flash minute 16: 42 retries deepseek-v4-flash minute 17: 44 retries deepseek-v4-flash minute 18: 42 retries deepseek-v4-flash minute 19: 38 retries deepseek-v4-flash minute 20: 40 retries deepseek-v4-flash minute 21: 40 retries deepseek-v4-flash minute 22: 36 retries deepseek-v4-flash minute 23: 41 retries deepseek-v4-flash minute 24: 41 retries deepseek-v4-flash minute 25: 40 retries deepseek-v4-flash minute 26: 25 retries deepseek-v4-flash minute 27: 44 retries deepseek-v4-flash minute 28: 38 retries deepseek-v4-flash minute 29: 44 retries deepseek-v4-flash minute 30: 33 retries deepseek-v4-flash minute 31: 30 retries deepseek-v4-flash minute 32: 43 retries deepseek-v4-flash minute 33: 41 retries deepseek-v4-flash minute 34: 36 retries deepseek-v4-flash minute 35: 44 retries deepseek-v4-flash minute 36: 34 retries deepseek-v4-flash minute 37: 31 retries deepseek-v4-flash minute 38: 40 retries deepseek-v4-flash minute 39: 38 retries deepseek-v4-flash minute 40: 42 retries deepseek-v4-flash minute 41: 38 retries deepseek-v4-flash minute 42: 35 retries deepseek-v4-flash minute 43: 34 retries deepseek-v4-flash minute 44: 41 retries deepseek-v4-flash minute 45: 37 retries deepseek-v4-flash minute 46: 38 retries deepseek-v4-flash minute 47: 33 retries deepseek-v4-flash minute 48: 33 retries deepseek-v4-flash minute 49: 38 retries deepseek-v4-flash minute 50: 39 retries deepseek-v4-flash minute 51: 42 retries deepseek-v4-flash minute 52: 35 retries deepseek-v4-flash minute 53: 28 retries deepseek-v4-flash minute 54: 32 retries deepseek-v4-flash minute 55: 31 retries deepseek-v4-flash minute 56: 30 retries deepseek-v4-flash minute 57: 31 retries deepseek-v4-flash minute 58: 30 retries deepseek-v4-flash minute 59: 33 retries deepseek-v4-flash minute 60: 32 retries deepseek-v4-flash minute 61: 32 retries deepseek-v4-flash minute 62: 34 retries deepseek-v4-flash minute 63: 28 retries deepseek-v4-flash minute 64: 28 retries deepseek-v4-flash minute 65: 32 retries deepseek-v4-flash minute 66: 36 retries deepseek-v4-flash minute 67: 29 retries deepseek-v4-flash minute 68: 24 retries deepseek-v4-flash minute 69: 39 retries deepseek-v4-flash minute 70: 21 retries deepseek-v4-flash minute 71: 24 retries deepseek-v4-flash minute 72: 22 retries deepseek-v4-flash minute 73: 25 retries deepseek-v4-flash minute 74: 21 retries deepseek-v4-flash minute 75: 30 retries deepseek-v4-flash minute 76: 22 retries deepseek-v4-flash minute 77: 23 retries deepseek-v4-flash minute 78: 27 retries deepseek-v4-flash minute 79: 26 retries deepseek-v4-flash minute 80: 24 retries deepseek-v4-flash minute 81: 24 retries deepseek-v4-flash minute 82: 20 retries deepseek-v4-flash minute 83: 21 retries deepseek-v4-flash minute 84: 28 retries deepseek-v4-flash minute 85: 23 retries deepseek-v4-flash minute 86: 25 retries deepseek-v4-flash minute 87: 21 retries deepseek-v4-flash minute 88: 17 retries deepseek-v4-flash minute 89: 22 retries deepseek-v4-flash minute 90: 20 retries deepseek-v4-flash minute 91: 14 retries deepseek-v4-flash minute 92: 16 retries deepseek-v4-flash minute 93: 8 retries deepseek-v4-flash minute 94: 11 retries deepseek-v4-flash minute 95: 10 retries deepseek-v4-flash minute 96: 20 retries deepseek-v4-flash minute 97: 15 retries deepseek-v4-flash minute 98: 13 retries deepseek-v4-flash minute 99: 19 retries deepseek-v4-flash minute 100: 15 retries deepseek-v4-flash minute 101: 12 retries deepseek-v4-flash minute 102: 8 retries deepseek-v4-flash minute 103: 11 retries deepseek-v4-flash minute 104: 10 retries deepseek-v4-flash minute 105: 14 retries deepseek-v4-flash minute 106: 9 retries deepseek-v4-flash minute 107: 8 retries deepseek-v4-flash minute 108: 9 retries deepseek-v4-flash minute 109: 10 retries deepseek-v4-flash minute 110: 8 retries deepseek-v4-flash minute 111: 7 retries deepseek-v4-flash minute 112: 9 retries deepseek-v4-flash minute 113: 6 retries deepseek-v4-flash minute 114: 10 retries deepseek-v4-flash minute 115: 6 retries deepseek-v4-flash minute 116: 9 retries deepseek-v4-flash minute 117: 8 retries deepseek-v4-flash minute 118: 8 retries deepseek-v4-flash minute 119: 7 retries deepseek-v4-flash minute 120: 5 retries deepseek-v4-flash minute 121: 4 retries deepseek-v4-flash minute 122: 7 retries deepseek-v4-flash minute 123: 8 retries deepseek-v4-flash minute 124: 7 retries deepseek-v4-flash minute 125: 5 retries deepseek-v4-flash minute 126: 8 retries deepseek-v4-flash minute 127: 6 retries deepseek-v4-flash minute 128: 4 retries deepseek-v4-flash minute 129: 5 retries deepseek-v4-flash minute 130: 5 retries deepseek-v4-flash minute 131: 3 retries deepseek-v4-flash minute 132: 5 retries deepseek-v4-flash minute 133: 3 retries deepseek-v4-flash minute 134: 4 retries deepseek-v4-flash minute 135: 4 retries deepseek-v4-flash minute 136: 3 retries deepseek-v4-flash minute 137: 6 retries deepseek-v4-flash minute 138: 1 retries deepseek-v4-flash minute 139: 7 retries deepseek-v4-flash minute 140: 5 retries deepseek-v4-flash minute 141: 3 retries deepseek-v4-flash minute 142: 4 retries deepseek-v4-flash minute 143: 6 retries deepseek-v4-flash minute 144: 3 retries deepseek-v4-flash minute 145: 3 retries deepseek-v4-flash minute 146: 6 retries deepseek-v4-flash minute 147: 6 retries deepseek-v4-flash minute 148: 5 retries deepseek-v4-flash minute 149: 4 retries deepseek-v4-flash minute 150: 5 retries deepseek-v4-flash minute 151: 4 retries deepseek-v4-flash minute 152: 4 retries deepseek-v4-flash minute 153: 3 retries deepseek-v4-flash minute 154: 5 retries deepseek-v4-flash minute 155: 4 retries deepseek-v4-flash minute 156: 4 retries deepseek-v4-flash minute 157: 1 retries deepseek-v4-flash minute 158: 3 retries deepseek-v4-flash minute 159: 2 retries deepseek-v4-flash minute 160: 2 retries deepseek-v4-flash minute 161: 3 retries deepseek-v4-flash minute 162: 4 retries deepseek-v4-flash minute 163: 2 retries deepseek-v4-flash minute 164: 1 retries deepseek-v4-flash minute 166: 2 retries deepseek-v4-flash minute 167: 1 retries deepseek-v4-flash minute 168: 1 retries deepseek-v4-flash minute 169: 2 retries deepseek-v4-flash minute 170: 3 retries deepseek-v4-flash minute 171: 2 retries deepseek-v4-flash minute 172: 2 retries deepseek-v4-flash minute 173: 2 retries deepseek-v4-flash minute 174: 1 retries deepseek-v4-flash minute 176: 1 retries deepseek-v4-flash minute 177: 1 retries deepseek-v4-flash minute 178: 1 retries deepseek-v4-flash minute 179: 1 retries deepseek-v4-flash minute 181: 1 retries deepseek-v4-flash minute 183: 1 retries deepseek-v4-flash minute 185: 1 retries deepseek-v4-flash minute 187: 1 retries deepseek-v4-flash minute 190: 2 retries deepseek-v4-flash minute 191: 1 retries deepseek-v4-flash minute 193: 1 retries deepseek-v4-flash 3,635 retries gpt-5.6-luna minute 24: 1 retries gpt-5.6-luna minute 71: 1 retries gpt-5.6-luna 2 gemini-3.7-flash minute 71: 3 retries gemini-3.7-flash 3 hy3 minute 0: 12 retries hy3 minute 1: 20 retries hy3 minute 2: 22 retries hy3 minute 3: 9 retries hy3 minute 4: 4 retries hy3 minute 5: 6 retries hy3 minute 6: 5 retries hy3 minute 7: 4 retries hy3 minute 8: 8 retries hy3 minute 9: 4 retries hy3 minute 10: 8 retries hy3 minute 11: 4 retries hy3 minute 12: 3 retries hy3 minute 13: 5 retries hy3 minute 14: 4 retries hy3 minute 15: 5 retries hy3 minute 16: 8 retries hy3 minute 17: 5 retries hy3 minute 18: 9 retries hy3 minute 19: 12 retries hy3 minute 20: 13 retries hy3 minute 21: 9 retries hy3 minute 22: 5 retries hy3 minute 23: 5 retries hy3 minute 24: 4 retries hy3 minute 25: 14 retries hy3 minute 26: 15 retries hy3 minute 27: 9 retries hy3 minute 28: 6 retries hy3 minute 29: 7 retries hy3 minute 30: 16 retries hy3 minute 31: 16 retries hy3 minute 32: 6 retries hy3 minute 33: 6 retries hy3 minute 34: 5 retries hy3 minute 35: 5 retries hy3 minute 36: 11 retries hy3 minute 37: 6 retries hy3 minute 38: 7 retries hy3 minute 39: 7 retries hy3 minute 40: 7 retries hy3 minute 41: 7 retries hy3 minute 42: 10 retries hy3 minute 43: 8 retries hy3 minute 44: 6 retries hy3 minute 45: 3 retries hy3 minute 46: 6 retries hy3 minute 47: 11 retries hy3 minute 48: 4 retries hy3 minute 49: 2 retries hy3 minute 50: 5 retries hy3 minute 51: 2 retries hy3 minute 52: 1 retries hy3 minute 53: 8 retries hy3 minute 54: 6 retries hy3 minute 55: 2 retries hy3 minute 56: 5 retries hy3 minute 57: 3 retries hy3 minute 58: 7 retries hy3 minute 59: 5 retries hy3 minute 60: 7 retries hy3 minute 61: 1 retries hy3 minute 62: 2 retries hy3 minute 63: 2 retries hy3 minute 64: 1 retries hy3 minute 65: 2 retries hy3 minute 66: 4 retries hy3 minute 67: 7 retries hy3 minute 68: 1 retries hy3 minute 69: 1 retries hy3 minute 70: 5 retries hy3 minute 71: 8 retries hy3 minute 72: 2 retries hy3 minute 73: 1 retries hy3 minute 74: 4 retries hy3 minute 75: 3 retries hy3 minute 76: 6 retries hy3 minute 77: 3 retries hy3 minute 78: 3 retries hy3 minute 79: 5 retries hy3 minute 82: 1 retries hy3 minute 83: 3 retries hy3 minute 84: 3 retries hy3 minute 85: 2 retries hy3 minute 86: 1 retries hy3 minute 87: 5 retries hy3 minute 88: 2 retries hy3 minute 89: 1 retries hy3 minute 90: 2 retries hy3 minute 91: 1 retries hy3 minute 92: 5 retries hy3 minute 93: 4 retries hy3 minute 94: 7 retries hy3 minute 98: 2 retries hy3 minute 99: 1 retries hy3 minute 100: 1 retries hy3 minute 101: 1 retries hy3 minute 102: 1 retries hy3 minute 103: 1 retries hy3 minute 104: 1 retries hy3 minute 105: 1 retries hy3 minute 106: 1 retries hy3 minute 109: 2 retries hy3 minute 110: 3 retries hy3 minute 111: 1 retries hy3 minute 116: 1 retries hy3 minute 119: 1 retries hy3 minute 120: 2 retries hy3 minute 121: 3 retries hy3 minute 122: 1 retries hy3 minute 124: 1 retries hy3 minute 132: 2 retries hy3 minute 133: 3 retries hy3 minute 134: 1 retries hy3 minute 135: 2 retries hy3 minute 145: 2 retries hy3 minute 151: 1 retries hy3 minute 154: 1 retries hy3 minute 157: 1 retries hy3 minute 158: 1 retries hy3 minute 159: 1 retries hy3 minute 160: 1 retries hy3 minute 164: 1 retries hy3 minute 165: 1 retries hy3 minute 166: 1 retries hy3 minute 167: 1 retries hy3 minute 168: 2 retries hy3 minute 175: 1 retries hy3 minute 194: 1 retries hy3 591 t=0 194 min model errors median decision deepseek-v4-flash remoteprotocol 1 timeout 3,593 unparseable 41 49.1s incl. retries gpt-5.6-luna timeout 2 5.4s incl. retries gemini-3.7-flash timeout 3 2.9s incl. retries hy3 connect 1 timeout 590 13.9s incl. retries
retry depth — how many times a single decision was re-asked
2,298 0 500 1,000 1,500 2,000 attempt 1: 2,298 decisions (deepseek-v4-flash 1,760 gpt-5.6-luna 2 gemini-3.7-flash 3 hy3 533) attempt 1 2,298 attempt 2: 1,144 decisions (deepseek-v4-flash 1,087 hy3 57) attempt 2 1,144 attempt 3: 490 decisions (deepseek-v4-flash 489 hy3 1) attempt 3 490 attempt 4: 201 decisions (deepseek-v4-flash 201) attempt 4 201 attempt 5: 56 decisions (deepseek-v4-flash 56) attempt 5 56 attempt 6: 16 decisions (deepseek-v4-flash 16) attempt 6 16 attempt 7: 9 decisions (deepseek-v4-flash 9) attempt 7 9 attempt 8: 7 decisions (deepseek-v4-flash 7) attempt 8 7 deepest chain: attempt 8 x7 Every timeout was retried until the model answered — nobody was substituted. deepseek stalled almost every minute of the 194-minute run (3,635 retries, median decision 49s incl. retries) yet still tied for first; luna won just as many while barely dropping a packet.