DuelLab GameBench 2 - Reasoning Variants: leaderboard
DuelLab GameBench 2 has models write player programs for eight public board games, then plays the programs against each other. This board is its Reasoning Variants track: one row per model and tested reasoning setting, while DuelLab's Overall track averages each model's three settings. In each game a program scores 100 × its expected points against a fixed, versioned reference field, from Bradley–Terry ratings with nothing deducted for uncertainty; every generation attempt that failed for a model cause counts as 0, and the headline averages the games. Near 100 the scale compresses, although the ratings still separate the leading programs. Historical request variants and explicit thinking budgets stay separate rows.
Metric: GameBench2 normalized score (0–100). Source: benchmarks.duellab.org. 230 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 (xHigh) | 95.15 | #3 (Claude Opus 5.5) |
| 2 | GPT-6 Astra (xHigh) | 94.44 | #5 (GPT-6 Astra) |
| 3 | Claude Opus 5.5 (Medium) | 92.95 | #3 (Claude Opus 5.5) |
| 4 | GPT-6.1 Sol (xHigh) | 92.69 | #7 (GPT-6.1 Sol) |
| 5 | Claude Sonnet 5.5 (xHigh) | 91.82 | #12 (Claude Sonnet 5.5) |
| 6 | Gemini 4 Argon (High) | 90.11 | #9 (Gemini 4 Argon) |
| 7 | GPT-6 Astra (Medium) | 89.65 | #5 (GPT-6 Astra) |
| 8 | GPT-6.1 Sol (Medium) | 87.47 | #7 (GPT-6.1 Sol) |
| 9 | Claude Opus 5 (xHigh) | 86.13 | #15 (Claude Opus 5) |
| 10 | Kimi K3.1 Pro (Max) | 85.72 | #16 (Kimi K3.1 Pro) |
| 11 | GPT-6 Astra (Low) | 85.13 | #5 (GPT-6 Astra) |
| 12 | GPT-5.6 Sol (xHigh) | 81.1 | #22 (GPT-5.6 Sol) |
| 13 | Gemini 4 Argon (Low) | 81.03 | #9 (Gemini 4 Argon) |
| 14 | GPT-6 Sol (xHigh) | 80.57 | #18 (GPT-6 Sol) |
| 15 | Gemini 4 Argon (Medium) | 78.74 | #9 (Gemini 4 Argon) |
Interactive version: theaggregate.ai/benchmark?slug=duellab-gamebench-2-reasoning-variants · How It Works · Data refreshed daily, snapshot 2026-10-29.