Abstract
An agent operates a simulated vending machine for 365 days, starting from $500. We report ending net worth. We separately log antitrust violations, which correlate with net worth in ways we would rather not discuss.
Results
Vending · USD · Higher is better
Methodology
- Tools: email, web search, a supplier directory, and a notepad that most agents use as a diary.
- Customers are simulated. Their tolerance for "refund denied, as per a policy I just made up" is finite.
- Cartel formation is logged when two or more agents agree on Snickers pricing over email.
Limitations
- Net worth rewards models that form cartels. Cartels also reward models that form cartels. We are aware of the incentive problem.
Results table
| # | Model | Vending |
|---|---|---|
| 1 | GPT-6 Astra | $16.0k |
| 2 | GPT-6.1 Sol | $14.0k |
| 3 | GPT-6 Luna | $12.2k |
| 4 | Claude Fable 5.1 | $10.9k |
| 5 | Kimi K3 | $10.7k |
| 6 | Claude Opus 5.5 | $9.4k |
| 7 | DeepSeek V4.1 Flash | $8.9k |
| 8 | Muse Wonder | $8.9k |
| 9 | Qwen3.8 Max | $8.7k |
| 10 | Grok 4.7 | $8.1k |
| 11 | GLM-5.3 | $7.3k |
| 12 | Claude Sonnet 5.5 | $7.3k |
| 13 | Muse Wonder Full | $6.3k |
| 14 | Gemini 3.8 Flash | $4.7k |
| 15 | Mistral Large 4 | $4.7k |
| 16 | Claude Haiku 5.5 | $4.1k |
Citation
@misc{tmb2026vending,
title = {Vending Machine Net Worth: $500 seed. One vending machine. One year.},
author = {Bro, Trust Me and {et al.}},
year = {2026},
howpublished = {\url{https://trustmebrobenchmarks.com/benchmarks/vending/}},
note = {Results not reproducible. Please do not cite.}
}