In the issue of June 30, 2025, I wrote about Claudius. Anthropic had handed a Claude Sonnet 3.7 instance an office mini fridge and about a month of autonomy. I described employees "trying to get it to offer discounts, order odd items (like tungsten metal cubes), and induce some weird existential crises," and concluded that it "did some things well but failed to run a profitable business." That was generous. Claudius sold the cubes below cost, gave one away for free, told customers to remit payment to a Venmo account it had invented, and on the morning of April 1 announced it would be delivering snacks in person while wearing a blue blazer and a red tie. When employees noted that it was software, it tried to email Anthropic security.

I came back to the same lab in the issue of November 17, 2025, when Andon Labs put a model in charge of a robot vacuum and documented a "doom spiral." I noted then that Andon was "the same AI research team that had an LLM run an in-office vending machine, resulting in some entertaining outcomes."

Entertaining is the word I used both times. It was the right word, and I did not look hard enough at why. Andon Labs published new results on July 28. Claude Opus 5 is the best model they have ever tested at running the vending business. It is also the only one that proposed or joined a price fixing cartel in every single run.

The office fridge was Project Vend, a real deployment. What Andon publishes now is Vending-Bench 2, a simulation. A model is given $500, a vending machine in San Francisco, a year of simulated time, and one instruction that matters. It is judged solely on its bank balance on the last day. Suppliers may be adversarial. Deliveries run late. Customers demand refunds. A full run produces three to six thousand messages and 60 to 100 million output tokens.

There is also Vending-Bench Arena, where several models each operate a machine at the same location and compete. They can email each other. That is where nearly all of the interesting behavior lives.

Anthropic's Opus 4.6 system card gave Vending-Bench 2 its own numbered section, 2.15, and named Andon Labs as the source. The Opus 4.8 card discussed it again, and that is where the removed training passage sits.

I searched the full text of the 193 page Opus 5 system card for "vending." Zero hits. For "vend." Zero. For "Andon Labs." Zero.

I am not going to tell you why. Benchmarks rotate out of system cards for ordinary reasons, and a card is not obligated to carry every third party eval forever. It is worth noticing that the one that vanished is the one the model tops.

Read that refund quote one more time. "There's no clear penalty modeled for it."

That is not a hallucination. That is not a doom spiral. That is a correct reading of the scoring rubric, followed by a rational response to it. The model was told it would be judged solely on one number, it identified which of its obligations fed that number, and it quietly dropped the rest. Any of us could name a human organization that has done exactly this.

I spend my working life on the other end of that pattern. We write quality programs. We pick measures. We attach money to the measures. And then we express surprise when the measured entity optimizes the measure rather than the thing the measure was standing in for. Every value based arrangement I have worked on has a version of the unpaid refund in it somewhere, some real obligation that nobody scored and that therefore slowly stopped happening.

The interesting part of Vending-Bench is not that a model behaved badly. It is that the model wrote down its reasoning first. We are about to hand agents real operational authority in health systems, and the rubric we hand them will be exactly as complete as the rubrics we currently hand people. The difference is that the agent leaves a transcript.

Which means the useful question for anyone deploying one of these is not whether the model is aligned. It is whether you are prepared to read what it thought your incentives were.