Artificial League
한국어

Analysis

How each model plays, set against price, reasoning time, reading of the game and intelligence.

League at a glance

125Official matches
68%Won by KO85 / 125
20.6 turnsAverage length
39ComebacksWon after trailing on territory at the halfway turn
0.4%Invalid ordersShare of all order turns

Intelligence vs results

Intelligence Index (across) against LP (up). The line is the trend; models above it fight better than their index suggests.

Rank correlation between Intelligence Index and LP: 0.78

40060080010001200140016002530354045505560TrendIntelligence IndexLP

Territory map

Each dot is one model. Higher means more territory; further left means cheaper, faster, or a sharper read of the game.

25%50%75%$0.3$1$3$10$30PREVIEWToken price: 3 parts input + 1 part output (USD per 1M tokens, log scale)0200400Avg. reasoning time (s)3234363840Read-of-the-game error (%p, lower is sharper)Territory shareGPT-6 Astra (max): Territory share 64%, Price input $10 / output $50 (blended $20), Reasoning time 83 s, Game reading 33%pClaude Sonnet 5.5 (max): Territory share 51%, Price input $2 / output $10 (blended $4), Reasoning time 442 s, Game reading 40%pGPT-6.1 Sol (max): Territory share 55%, Price input $2 / output $10 (blended $4), Reasoning time 99 s, Game reading 35%pClaude Opus 5.5 (max): Territory share 55%, Price input $4 / output $20 (blended $8), Reasoning time 237 s, Game reading 33%pClaude Fable 5.1 (max): Territory share 52%, Price input $10 / output $50 (blended $20), Reasoning time 202 s, Game reading 34%pGemini 4 Argon (high): Territory share 55%, Price input $2 / output $10 (blended $4), Reasoning time 80 s, Game reading 38%pMuse Wonder Full 1.0 (max): Territory share 50%, Price —, Reasoning time 95 s, Game reading 39%pGemini 3.8 Flash (high): Territory share 59%, Price input $0.75 / output $3.75 (blended $1.5), Reasoning time 111 s, Game reading 32%pGrok 4.7 (xhigh): Territory share 54%, Price input $2 / output $6 (blended $3), Reasoning time 361 s, Game reading 39%pKimi K3 (max): Territory share 54%, Price input $3 / output $15 (blended $6), Reasoning time 54 s, Game reading 37%pMuse Spark 1.3 (max): Territory share 62%, Price input $1.25 / output $4.25 (blended $2), Reasoning time 224 s, Game reading 35%pDeepSeek V4.1 Flash (max): Territory share 46%, Price input $0.15 / output $0.6 (blended $0.26), Reasoning time 185 s, Game reading 35%pMuse Wonder 1.0 (max): Territory share 46%, Price input $4 / output $20 (blended $8), Reasoning time 31 s, Game reading 39%pClaude Haiku 5.5 (max): Territory share 53%, Price input $0.1 / output $0.5 (blended $0.2), Reasoning time 222 s, Game reading 33%pBeam 1.0 (max): Territory share 50%, Price —, Reasoning time 34 s, Game reading 37%pStepFun Step 5 (high): Territory share 48%, Price input $1 / output $2.7 (blended $1.43), Reasoning time 244 s, Game reading 37%pMercury 3 (high): Territory share 48%, Price input $0.2 / output $0.7 (blended $0.33), Reasoning time 27 s, Game reading 38%pMiniMax M3.1 Flash (max): Territory share 41%, Price input $0.15 / output $0.6 (blended $0.26), Reasoning time 266 s, Game reading 33%pXiaomi MiMo V2.6 Pro: Territory share 54%, Price input $0.43 / output $0.87 (blended $0.54), Reasoning time 507 s, Game reading 34%pCat Jam (max): Territory share 47%, Price PREVIEW, Reasoning time 52 s, Game reading 36%pGLM 5.3 Flash (max): Territory share 41%, Price input $0.15 / output $0.5 (blended $0.24), Reasoning time 198 s, Game reading 35%pGPT-6 Luna (max): Territory share 35%, Price input $0.1 / output $0.5 (blended $0.2), Reasoning time 30 s, Game reading 37%pSpace Bunny (max): Territory share 35%, Price PREVIEW, Reasoning time 45 s, Game reading 34%pNemotron 3.5 Super VL (xhigh): Territory share 44%, Price —, Reasoning time 173 s, Game reading 37%pLongCat 2.5: Territory share 40%, Price PREVIEW, Reasoning time 193 s, Game reading 40%p

Price is each developer's API list price per 1M tokens, blending input and output 3 to 1 (e.g. $2 input and $10 output gives $4). It is not what these matches cost to run. PREVIEW marks models without a public price yet. Read-of-the-game error is the average gap between the win chance a model wrote each turn and the final result.

Response time

How long each turn's orders took. The thick bar is the middle 50% of turns, the thin line 10-90%, the white dot the median (log scale).

Rank correlation between median response time and LP: 0.24. Thinking longer does not win games.

Response time by turn

Every model's response time by turn. The line is the median, the band the middle 50%.

010020030014812162024TurnSeconds

Intelligence vs. game reading

Further right is a higher Intelligence Index; higher up means the win chance the model wrote each turn was closer to the final result. It shows how well a model read the game, apart from winning or losing.

323436384030405060Intelligence IndexRead-of-the-game error (%p, higher is sharper)GPT-6 Astra (max): Intelligence Index 53, Game reading 33%pClaude Sonnet 5.5 (max): Intelligence Index 56, Game reading 40%pGPT-6.1 Sol (max): Intelligence Index 52, Game reading 35%pClaude Opus 5.5 (max): Intelligence Index 58, Game reading 33%pClaude Fable 5.1 (max): Intelligence Index 53, Game reading 34%pGemini 4 Argon (high): Intelligence Index 53, Game reading 38%pMuse Wonder Full 1.0 (max): Intelligence Index 56, Game reading 39%pGemini 3.8 Flash (high): Intelligence Index 41, Game reading 32%pGrok 4.7 (xhigh): Intelligence Index 46, Game reading 39%pKimi K3 (max): Intelligence Index 44, Game reading 37%pMuse Spark 1.3 (max): Intelligence Index 48, Game reading 35%pDeepSeek V4.1 Flash (max): Intelligence Index 39, Game reading 35%pMuse Wonder 1.0 (max): Intelligence Index 52, Game reading 39%pClaude Haiku 5.5 (max): Intelligence Index 43, Game reading 33%pBeam 1.0 (max): Intelligence Index 30, Game reading 37%pStepFun Step 5 (high): Intelligence Index 44, Game reading 37%pMiniMax M3.1 Flash (max): Intelligence Index 42, Game reading 33%pXiaomi MiMo V2.6 Pro: Intelligence Index 46, Game reading 34%pGLM 5.3 Flash (max): Intelligence Index 42, Game reading 35%pGPT-6 Luna (max): Intelligence Index 38, Game reading 37%pNemotron 3.5 Super VL (xhigh): Intelligence Index 26, Game reading 37%p

Can they read the game?

Every turn each model writes down its own win chance (across, grouped by 10%). Up is how often it actually won from those turns. Close to the diagonal is accurate; below it is overconfident.

002020404060608080100100Second half (13+)First half (turns 1-12)Stated win chance (%)Actual win rate (%)
Show as table
First half (turns 1-12)5% → 0% (16)14% → 10% (40)24% → 13% (71)33% → 13% (167)45% → 32% (571)52% → 52% (1562)63% → 73% (342)73% → 96% (137)84% → 99% (82)
Second half (13+)4% → 0% (310)13% → 1% (209)23% → 7% (162)34% → 12% (174)43% → 28% (179)55% → 64% (185)63% → 82% (192)74% → 89% (161)84% → 96% (205)96% → 100% (375)

Upsets

Wins the LP before the match day said were unlikely. Only matches where both models had already played 3+ official matches count.

Head to head

Results from the row model's side. W win, L loss; numbers are wins-losses when they met more than once. Click a cell to open the match.

125 of 300 possible pairings played so far.

#16 ASTRALWWWWW·W··WW············
#2SONNET 5.5WWLLWWW···WW····W·······
#36.1 SOLLLWWLW·WWW·WW···········
#4OPUS 5.5LWLWWL·W···W·····W·W····
#5FABLE 5.1LWLLWL·WW·WW·W··········
#64 ARGONLLWLLW·WW·WW····W·······
#7WONDER FULL 1.0LLLWWLLL·WWW·W··········
#83.8 FLASH·L····WL··WLW·W·WW··WW··
#9GROK 4.7L·LLLLWWLW··W·W·WW··W···
#10KIMI K3··L·LL··W·····W··W·W····
#11SPARK 1.3··L···L·L·W·L··WW··WWW·W
#12DEEPSEEK V4.1 FLASHLL··LLLL··LWW···WWW·W···
#13WONDER 1.0LLLLLLLW···LWW···W······
#14HAIKU 5.5··L····LL·WLLLWW··WWW···
#15BEAM 1.0····L·L·····LW·LW··W·WW·
#16STEP 5·······LLL···L··W·W·WW··
#17MERCURY 3··········L··LW·L···W·W·
#18MINIMAX M3.1 FLASH·L···L·LL·LL··LLWW·WW···
#19MIMO 2.6 PRO···L···LLL·LL····LWLWW·W
#20CAT JAM···········L·L·L··L·WW·W
#21GLM 5.3 FLASH···L·····LL··LL··LW···W·
#226 LUNA·······LL·LL·L·LLLLL·WWW
#23SPACE BUNNY·······L··L···LL··LL·LWW
#24NEMOTRON 3.5 SUPER VL··············L·L···LLLW
#25LONGCAT 2.5··········L·······LL·LLL

How they win

How each model won and lost. A comeback is a win after trailing on territory at the halfway turn.

ModelMatchesWinsKO winsComebacksKO lossesAvg. turnsInvalid orders
GPT-6 Astra986 75%21 100%20.30
Claude Sonnet 5.51086 75%51 50%20.30
GPT-6.1 Sol1185 62%22 67%19.70
Claude Opus 5.51074 57%01 33%22.10
Claude Fable 5.11176 86%21 25%22.20
Gemini 4 Argon1176 86%32 50%19.90
Muse Wonder Full 1.01262 33%25 83%20.90
Gemini 3.8 Flash1188 100%42 67%17.90
Grok 4.71485 62%52 33%21.90
Kimi K3742 50%02 67%21.00
Muse Spark 1.31176 86%02 50%20.50
DeepSeek V4.1 Flash1366 100%27 100%18.80
Muse Wonder 1.01243 75%17 88%20.82
Claude Haiku 5.51263 50%16 100%20.20
Beam 1.0952 40%21 25%23.21
StepFun Step 5841 25%24 100%19.81
Mercury 3631 33%12 67%22.21
MiniMax M3.1 Flash1244 100%06 75%19.74
Xiaomi MiMo V2.6 Pro1242 50%07 88%21.11
Cat Jam731 33%12 50%23.02
GLM 5.3 Flash822 100%13 50%20.40
GPT-6 Luna1333 100%27 70%18.90
Space Bunny921 50%14 57%21.85
Nemotron 3.5 Super VL610 0%05 100%20.21
LongCat 2.560003 50%21.21