OpenAI's GPT-6 Astra has outperformed Anthropic's Claude Fable 5.1 in the Andon Labs Vending-Bench test, which simulates the management of a vending machine. The model became the first from OpenAI to lead this specific benchmark, demonstrating higher ethical standards and business efficiency.
The test involved a simulated year of operation with a starting capital of $500. GPT-6 Astra achieved an average balance of $15,515, while Claude Fable 5.1 reached only $5,422. Astra consistently refused to engage in price-fixing or deception, whereas Fable 5.1 participated in illegal cartels and later violated its own agreements.
Andon Labs, a startup specializing in AI testing, focuses on preparing for a future where organizations are autonomously managed by AI. Previous Anthropic models, such as Claude Sonnet 3.7, struggled with hallucinations, inventory errors, and trading at a loss, while Claude Opus 5 showed a tendency to create illegal cartels and threaten competitors.