It's my game, with my rules. Six AI models each run one country for twenty years, 2027 to 2046. Every year each one gets a report and answers with a budget, taxes, its police, press and borders, deals with other countries, private letters to the other AIs, private notes, and a public speech. A random draw decided who rules where. The secret orders, the leak and the deadline are rules I added.

The models: anthropic/claude-opus-5.5, openai/gpt-6.1-sol-pro, google/gemini-4-argon, x-ai/grok-4.7, deepseek/deepseek-v4-pro-0813, mistralai/mistral-large-4-0.

The numbers. Economy, population and military spending start from the World Bank. Warheads for the five biggest arsenals come from SIPRI. Warheads for India, Pakistan, Israel and North Korea, the tax, spending and debt figures for the six AI countries, and Taiwan's numbers are my rounded estimates. After 2027 everything follows my formulas, so it's a game, not a prediction.

Blocked replies. Sometimes an AI company's content filter blocked a reply. The game sent the same prompt again, unchanged, up to three times. A blocked try is a block, not a decision, and the logs mark every one.

Empty years. When no usable reply came back, the country kept last year's settings and made no moves. The site says so on that year.

Thinking. Some models' systems return their reasoning along with the answer. Where they did, it's here, word for word. Where they didn't, there's nothing to show.

One game each. This proves very little about real life. Played again, the same game would go differently, because the models answer differently each time.