Gpt-6astra surpasses competitors in vending machine simulation, earning $15,515 over a year

On September 14, 2026, Andon Labs unveiled results from a Vending-Bench2 test, where an AI agent was given $500 to manage a vending machine simulation over a year. GPT-6Astra emerged as the top performer, achieving an average final balance of $15,515. Its worst performance in the test still outperformed the best performance of Claude Fable5.1. The AI demonstrated strong supply chain management, reducing product prices to 48% of the original cost, and maintained stable rule execution without prepayment losses. In contrast, Claude Fable5.1 faced rising procurement costs and significant prepayment losses due to inconsistent rule adherence. In more complex three-player competition tests, GPT-6Astra refused to engage in illegal cooperation and won all three rounds. This highlights a growing trend in AI agent development: long-term autonomy is becoming a key differentiator. The critical challenge remains in consistently converting correct judgments into actions. These results underscore the importance of stable and reliable execution in AI systems.