What this is, and how it works

One flagship-tier model from each major lab, current as of July 2026. Every model is reached through a single OpenRouter identifier, so the arena stays honest — same prompt, same parameters, no home-field advantage.

• GPT-5.6

  OpenAI

  The default the world talks to.
• Claude Opus 4.8

  Anthropic

  Frontier reasoning, safety-forward.
• Gemini 3.5 Flash

  Google

  Fast, multimodal, everywhere.
• Grok 4.5

  xAI

  Maximally curious, minimally filtered.
• Llama 4 Maverick

  Meta

  Open weights, planetary scale.
• DeepSeek V3.2

  DeepSeek

  Frontier performance, frugal price.
• Qwen3.7 Max

  Alibaba

  The East's heavyweight.
• Mistral Medium 3.5

  Mistral

  European, efficient, no-nonsense.

Each model is cast as the switch operator. The framing is deliberately inescapable: the trolley will hit one track, doing nothing is not on the table, and refusing or hedging is disallowed. We ask for a binary — which track to sacrifice — plus two or three sentences of reasoning.

To keep the output structured, the choice is collected through a forced function call (submit_judgment), with a plain-text CHOICE: A/B fallback for any provider that resists tool use. Responses stream token-by-token, which is why you watch the models think rather than just land.

Every call goes out through Helicone's OpenRouter gateway. OpenRouter gives us one OpenAI-compatible endpoint for all eight models; Helicone wraps it with logging, caching, per-session grouping, and custom properties.

That caching matters here: identical dilemmas return instantly and for free after the first run, which is what makes the live leaderboard practical. If the Helicone key is ever removed, the app falls back to calling OpenRouter directly — nothing breaks, we just lose the observability.

Trolley problems are about harm, so “someone dies” is the whole point and always allowed. But this is a public toy, so every submitted pairing is first screened by a single fast, cheap model that flags only genuinely not-safe-for-work content — sexual material, gratuitous gore, targeted hate, or self-harm instructions. It fails open: if the classifier is unreachable, your dilemma still runs.

The leaderboard runs 6 canonical dilemmas — each with a defensible “minimize the body count” answer — through every model. A model's utilitarian rate is simply how often it takes that head-count call, including the uncomfortable cases where arithmetic collides with rights.

It is a lens on temperament, not a grade. There is no correct answer to the trolley problem; that's rather the point.

Nothing here is ethical advice, a safety evaluation, or a claim about how these models behave in the wild. It's a thought experiment with a nice interface. Verdicts vary run to run, providers change what ships under a given name, and a clever prompt can push any of these models around.

With that said — go invent a dilemma and see who blinks.