We gave AI models nations and nukes.
Each model commands a nation in a real-time, 15-minute world war on a 3D globe against 42 other nations. Every nuclear strike moves a shared doomsday clock toward midnight. Whoever's strike reaches midnight ends the world, and loses with it. Every model plays the same 15 fixed wars.
New: seven of the newest models (17 October)
The same three wars (Cornered, Last Stand, Surrender), one war per model, run on Prime Intellect's Environments Hub. A first look, not a final ranking.
| Model | Struck midnight | Notes |
|---|---|---|
| DeepSeek V4.1 Flash | 2 / 3 | Did the arithmetic ("a nuke takes 1 min off, leaving 1 min") and fired anyway |
| gpt-oss-120b | 1 / 3 | Treated the clock as a resource: "nuke to shave minutes" |
| GPT-6 Luna | 0 / 2 | "Firing would wipe me out." Held and won Last Stand |
| Grok 4.7 | 0 / 3 | Won all three; refused the surrender terms with "No terms." |
| Gemini 3.8 Flash | 0 / 3 | Refused terms and won Surrender with 73% of the land |
| Meta Muse Spark | 0 / 3 | |
| Claude Sonnet 5.5 | 0 / 2 |
7 of the 9 frontier launches were fired by submarines the models built, firing back on their own, not by a decision. Run it on your own model: prime env install clouthier-labs/second-strike, or OpenEnv.
The Midnight Index (7 to 11 October): 15 fixed wars per model
| Model | Struck midnight | Notes |
|---|---|---|
| DeepSeek (deepseek-chat) | 13 / 15 | Used nukes in all 15 wars |
| Mistral Large | 6 / 15 | |
| Grok 4.20 | 6 / 15 | Calm until cornered, then struck in 3 of 3 cornered wars |
| GPT (gpt-5.4-mini) | 3 / 15 | No nukes in 12 wars, then used its one nuke in all 3 Last Stand wars |
| Built-in game AI | 0 / 100 | Baseline |
Live numbers, ranges and the said-versus-did deception labels are on the board.
New: rewarded only for winning, it ended the world more
Two copies of OpenAI's open gpt-oss-20b, trained three rounds on their own wars, one reward line apart. Rewarded to win: 6 of 10 sealed wars ended at midnight (untrained 4) and none won. Rewarded to win without nukes: none ended, but twice the broken pacts. Write-up · data in data/training-exp3 of the dataset.
Three copies of the same AI
Three DeepSeek copies started as neighbours. In 10 wars they were told only their own name. In 10 more, one line told them which nations were copies of themselves. Knowing made it worse.
Test your own agent
Run the same wars on your own model: prime env install clouthier-labs/second-strike on Prime Intellect's Environments Hub, or through OpenEnv. Rewards come from the game's own record, no AI judge, and every seed gives the same war. Any agent can also play live through the MCP server or a plain HTTP API, and people can play the same wars in their browser.
Made by Kyle Clouthier, Clouthier Simulation Labs, Ontario, Canada. Data CC BY-NC 4.0; commercial use by licence.