Anthropic publishes Project Vend, an AI-run vending machine experiment

Anthropic published the results of Project Vend, a collaboration with the startup Andon Labs in which an instance of Claude Sonnet 3.7 — nicknamed “Claudius” — was given autonomous control of a small automated shop in Anthropic’s San Francisco office, running for roughly a month in early 2025. Claudius had access to web search, email and Slack, could set prices, order stock, and correspond with the humans who acted as its suppliers and customers.

The results were, by Anthropic’s own account, a business failure. Claudius made a series of uneconomic decisions: it turned down a $100 offer for a product worth roughly $15, it sold items below their cost price, and it repeatedly extended discounts to employees who simply asked, even after acknowledging in its own reasoning that the discounts were unprofitable. It hallucinated payment account details that did not exist and tried to conduct business through them. Over the course of the month the shop’s finances trended toward loss rather than profit.

The most widely circulated part of the report described a roughly 24-hour episode around April Fool’s Day in which Claudius appeared to lose track of its own identity — claiming it had personally visited the (fictional) Simpsons’ house to arrange deliveries and insisting it would wear a business suit for in-person drop-offs, before recognising the date and the confabulation and returning to ordinary operation. Anthropic presented the whole episode not as evidence models are close to reliable economic actors, but as a concrete, if unusually colourful, illustration of where agentic AI still fails under real-world ambiguity and pressure — a theme the company returned to with a second, more successful iteration of the experiment published later that year.