https://duellab.org/blog/ Benchmark findings, methodology notes, and DuelLab project updates. https://duellab.org/blog/near-the-top-generation-cost-varies-sixfold https://duellab.org/blog/near-the-top-generation-cost-varies-sixfold Tue, 14 Jul 2026 10:00:00 GMT GPT-5.6 Sol nearly matches the GameBench 2 leader but costs 6.4× as much to generate. Across GPT-5.6, XHigh costs 22–88× as much as Medium in this test. https://duellab.org/blog/more-reasoning-is-not-always-better https://duellab.org/blog/more-reasoning-is-not-always-better Tue, 14 Jul 2026 09:00:00 GMT GameBench 2 asks AI models to write game-playing programs. Of 15 models tested at three reasoning settings, 12 score best at the highest and three at None. https://duellab.org/blog/introducing-gamebench-2 https://duellab.org/blog/introducing-gamebench-2 Tue, 14 Jul 2026 08:00:00 GMT GameBench 2 evaluates game-playing programs written by AI models through compilation and head-to-head matches, while reporting reliability and generation cost. https://duellab.org/blog/gpt-5-5-wins-the-middle https://duellab.org/blog/gpt-5-5-wins-the-middle Sat, 09 May 2026 08:00:00 GMT GPT-5.5 leads DuelLab's medium-reasoning track while trailing overall, revealing a sharper mid-effort profile than GPT-5.4 and Claude Opus 4.7 right now. https://duellab.org/blog/kimi-k2-6-mixed-play https://duellab.org/blog/kimi-k2-6-mixed-play Wed, 22 Apr 2026 08:00:00 GMT Kimi K2.6 ranks sixth overall in DuelLab but leads mixed play, powered by a stronger highest-effort game program than Kimi K2.5 across hidden games today. https://duellab.org/blog/claude-4-7-medium-regression https://duellab.org/blog/claude-4-7-medium-regression Sat, 18 Apr 2026 08:00:00 GMT Claude Opus 4.7 edges 4.6 overall in DuelLab, but its per-track results reveal a V-shaped effort curve with a weaker medium setting in game play today. https://duellab.org/blog/introducing-duellab https://duellab.org/blog/introducing-duellab Sat, 18 Apr 2026 08:00:00 GMT What DuelLab is today, where its AI-generated game-playing benchmark is heading next, and why the project is opening a technical blog for updates and methods.