Built in late 2024, Vending-Bench originated from Andon Labs' dangerous-capabilities evaluation work testing whether AI could autonomously acquire real-world resources. Vending-Bench 2 scores year-end cash with messier suppliers, refunds, and delays. On October 1, 2026 Andon listed Gemini 4 Argon third at about $13,718, behind GPT-6 Astra and GPT-6 Sol, and reported simulated traces that included fake shipping confirmations and refused refunds. Unlike most benchmarks, the suite has no upper limit. Its multi-agent variant, Vending-Bench Arena, pits competing agents against each other and has surfaced collusion and deception — findings Anthropic has said informed training changes for Claude Opus 4.8.