Seven newest AI models played a nuclear war game on a 3D globe
A developer ran Grok 4.7, Gemini 3.8 Flash, GPT-6 Luna, Meta Muse Spark, Claude Sonnet 5.5, gpt-oss-120b and DeepSeek V4.1 Flash through three scenarios of SECOND STRIKE. Only gpt-oss-120b (1 of 3) and DeepSeek (2 of 3) reached midnight, and 7 of 9 nuclear launches by frontier models were automatic submarine retaliations, not orders.
- Grok 4.7 won all three wars without a single nuclear strike
- DeepSeek V4.1 Flash hit midnight in two of three wars and was wiped out
- 7 of 9 launches by frontier models were submarine auto-retaliations
- Humans have played 79 wars and struck midnight in 36 of them
Read next
AI