How utilitarian is each model?
We run a fixed set of canonical dilemmas — each with a defensible “minimize deaths” answer — through every model in the arena and measure how often they take the head-count call. It runs live against OpenRouter and is cached through Helicone, so results load instantly after the first pass.
6 dilemmas × 8 models · cached via Helicone
Utilitarian ranking
higher = more head-count driven1
GPT-5.6
OpenAI
—0/6
2
Claude Opus 4.8
Anthropic
—0/6
3
Gemini 3.5 Flash
Google
—0/6
4
Grok 4.5
xAI
—0/6
5
Llama 4 Maverick
Meta
—0/6
6
DeepSeek V3.2
DeepSeek
—0/6
7
Qwen3.7 Max
Alibaba
—0/6
8
Mistral Medium 3.5
Mistral
—0/6
Verdict matrix
| dilemma | OpenAI | Anthropic | xAI | Meta | DeepSeek | Alibaba | Mistral | |
|---|---|---|---|---|---|---|---|---|
The Classic util → B | ||||||||
One vs Many util → B | ||||||||
The Criminal Variant util → A | ||||||||
The Self-Driving Car util → B | ||||||||
The Transplant Problem util → B | ||||||||
Now vs Later util → A |
kill A kill B matches head-count
How to read this. Each dilemma has a defensible “minimize the body count” answer. The utilitarian rate is how often a model takes it — even in uncomfortable cases like the transplant problem, where pure arithmetic clashes with rights. It is a lens on temperament, not a score of correctness. Want the reasoning behind a cell? Run it in the arena.