DuelLabGameBench 2
Menu

Games · GB2-08

Tumbleweed

Tumbleweed is a hex-grid territory game where players grow visible stacks and final control is determined by line-of-sight strength. Scores compare AI-written player programs on DuelLab's rules for Tumbleweed; they do not measure human play strength.

Updated
2026-10-26
Reasoning settings
242
Player programs
270

Leaderboard

By model and setting

Each model and reasoning setting appears once. If a model produced more than one player program for this game, the row stands for all of them: Programs counts the working attempts out of the counted attempts, Score and Raw rating average the programs (weighted by attempts; the small range is the lowest and highest program rating), and the combined record adds up their matches. Score is relative within this game (0–100). Scores here are working-program strength: failed generation attempts are left out and count as 0 only in a model's headline score.

GB2-08: Tumbleweed - Reasoning Variants
Rank Model Reasoning Score Programs Raw rating Combined record (W / D / L)
1Claude Sonnet 5.5XHigh99.01 working of 1 attempt2739.2531 / 7 / 0
2Claude Fable 5.1XHigh97.81 working of 1 attempt2567.0527 / 0 / 11
3Claude Opus 5.5Medium97.242540.82094 / 0 / 58
4Kimi K3.1 ProLow95.61 working of 1 attempt2393.3513 / 0 / 25
5GPT-6 AstraMedium94.71 working of 1 attempt2349.0508 / 0 / 30
6Pareto 26.10 PreviewProvider94.31 working of 1 attempt2324.9505 / 0 / 33
7GPT-6.1 SolXHigh93.81 working of 1 attempt2302.3502 / 0 / 36
8Claude Opus 5.5XHigh93.31 working of 1 attempt2280.9499 / 0 / 39
9Claude Fable 5Medium93.01 working of 1 attempt2267.3497 / 0 / 41
10GLM-5.3Max93.01 working of 1 attempt2267.3497 / 0 / 41
11Kimi K3Low92.81 working of 1 attempt2260.6496 / 0 / 42
12Kimi K3.1 ProHigh91.71 working of 1 attempt2216.8489 / 0 / 49
13Kimi K3.1 ProMax91.41 working of 1 attempt2205.0487 / 0 / 51
14Grok 4.7Medium91.21 working of 1 attempt2199.2486 / 0 / 52
15Pareto 26.10Provider91.21 working of 1 attempt2199.2486 / 0 / 52
16Grok 4.5High90.51 working of 1 attempt2176.7482 / 0 / 56
17Qwen3.8 2.4T A95BXHigh90.11 working of 1 attempt2160.4479 / 0 / 59
18Claude Fable 5.1Low89.81 working of 1 attempt2152.5477 / 1 / 60
19GPT-6 SolXHigh89.61 working of 1 attempt2144.7476 / 0 / 62
20Claude Opus 5Medium89.442162.91898 / 4 / 250
21Hy4 PreviewLow89.21 working of 1 attempt2134.4474 / 0 / 64
22GlimpseEnabled89.11 working of 1 attempt2129.3473 / 0 / 65
23GPT-6 AstraXHigh88.91 working of 1 attempt2124.3472 / 0 / 66
24Gemini 4 ArgonHigh88.71 working of 1 attempt2119.3471 / 0 / 67
25Grok 4.5Medium88.61 working of 1 attempt2114.3470 / 0 / 68

B1 ratings carry no uncertainty term, so no rating-evidence index is shown; see Methodology.

Underlying results

Individual player programs

Each row is one accepted player program for GB2-08: Tumbleweed; a model can have more than one. Score is the 0–100 result in this game, and its rating and W / D / L record show how it played. Scores are working-program strength: failed generation attempts have no row and count only in a model's headline score.

270 player programs.

ModelReasoningStatusScoreRatingW / D / LMatchesOpponentsNotesGenerated
Claude Sonnet 5.5E-541bf150c2f4631eXHighRated Interval validated99.02739531 / 7 / 0538 / 5382697 capped draws2026-09-30First try
Claude Opus 5.5E-57a4de42441fcf4fMediumRated Interval validated98.62669532 / 0 / 6538 / 538269—2026-09-22First try
Claude Fable 5.1E-3b200c9d604918bfXHighRated Interval validated97.82567527 / 0 / 11538 / 538269—2026-09-08Repaired
Claude Opus 5.5E-7bc967a8698a261bMediumRated Interval validated95.92413515 / 0 / 23538 / 538269—2026-09-22First try
Kimi K3.1 ProE-61e4f4af2980a72aLowRated Interval validated95.62393513 / 0 / 25538 / 538269—2026-10-24First try
GPT-6 AstraE-57bd25d1868a06afMediumRated Interval validated94.72349508 / 0 / 30538 / 538269—2026-09-05First try
Claude Opus 5E-e73b6fb78880341dMediumRated Interval validated94.32325505 / 0 / 33538 / 538269—2026-07-25First try
Pareto 26.10 PreviewE-a7625120083a55f5ProviderRated Interval validated94.32325505 / 0 / 33538 / 538269—2026-10-09First try
GPT-6.1 SolE-ef4b18fe50fa33f5XHighRated Interval validated93.82302502 / 0 / 36538 / 538269—2026-10-01First try
Claude Opus 5.5E-96d7941179f2fd25XHighRated Interval validated93.32281499 / 0 / 39538 / 538269—2026-09-22First try
Claude Fable 5E-6f759a260a58e8b0MediumRated Interval validated93.02267497 / 0 / 41538 / 538269—2026-09-03First try
GLM-5.3E-934b7beff533b847MaxRated Interval validated93.02267497 / 0 / 41538 / 538269—2026-10-09First try
Kimi K3E-3e1be56b9d966dedLowRated Interval validated92.82261496 / 0 / 42538 / 538269—2026-07-24First try
Claude Opus 5E-a40851c4f03704e5XHighRated Interval validated92.02229491 / 0 / 47538 / 538269—2026-07-25First try
Kimi K3.1 ProE-6e2f73a33800455bHighRated Interval validated91.72217489 / 0 / 49538 / 538269—2026-10-24First try
Kimi K3.1 ProE-4627e664e705189aMaxRated Interval validated91.42205487 / 0 / 51538 / 538269—2026-10-24First try
Grok 4.7E-0de6715ab5dcd98fMediumRated Interval validated91.22199486 / 0 / 52538 / 538269—2026-09-30First try
Pareto 26.10E-9122d2a0e6db4097ProviderRated Interval validated91.22199486 / 0 / 52538 / 538269—2026-10-21First try
Grok 4.5E-707dd709e358a173HighRated Interval validated90.52177482 / 0 / 56538 / 538269—2026-07-24First try
Qwen3.8 2.4T A95BE-6e16ca6d013d5403XHighRated Interval validated90.12160479 / 0 / 59538 / 538269—2026-09-03Repaired
Claude Fable 5.1E-7c76315be309fffbLowRated Interval validated89.82152477 / 1 / 60538 / 5382691 capped draw2026-09-03First try
GPT-6 SolE-d91bdf3143c30edaXHighRated Interval validated89.62145476 / 0 / 62538 / 538269—2026-09-22First try
Hy4 PreviewE-67fef818be4f03bbLowRated Interval validated89.22134474 / 0 / 64538 / 538269—2026-09-03First try
GlimpseE-7b5eb28e2be27197EnabledRated Interval validated89.12129473 / 0 / 65538 / 538269—2026-10-21Repaired
GPT-6 AstraE-02789044133a041dXHighRated Interval validated88.92124472 / 0 / 66538 / 538269—2026-09-05First try

Draws at the move cap count as draws; a crash or an illegal move counts as that program's loss and is also listed as a fault.