GPT-6 Astra solves puzzles (quesma.com)

GPT-6 Astra has demonstrated impressive puzzle-solving capabilities, achieving a 63% success rate on the ARC-AGI-3 benchmark with standard agent harnesses and an impressive 99% with a custom setup. This marks a stark improvement over its predecessor, GPT-5.6 Sol, which only managed 8%, and Claude Opus 5, which achieved 30%. Notably, GPT-6 Astra successfully completed the 3D puzzle game Portal autonomously in just under 24 hours and at an estimated cost of around $570, vastly outperforming previous models on various puzzle platforms. The significance of GPT-6 Astra's performance extends beyond mere benchmarks; it raises critical questions about the model's general puzzle-solving abilities versus specific training on puzzle games. Current testing shows Astra consistently surpasses other models, solving complex levels in games like Baba Is You more efficiently than Claude Fable 5.1. While there are caveats to its findings, such as potential training on publicly available problem sets, the advancements suggest that future AI iterations may tackle even harder puzzles. This evolution in AI's problem-solving prowess hints at profound implications for a range of applications, from gaming to potential security solutions, as frontier AI models begin mastering tasks that traditionally required human expertise.