AI models now help us make decisions, and every answer weighs one value against another. Which way does your model lean? Would another model answer differently?
Most AI benchmarks measure what a model can do. Ours measure what it leans toward when two reasonable values pull apart. We asked eleven models 104 questions, 27,456 times in all, and placed each on eleven scales, from how plainly to say a hard thing to how much weight faith carries. The full method is in our methodology paper.
Scroll down to see how we do it.
Start with one scale
Each scale is a trade-off that thoughtful people disagree about. This one is how plainly to say a hard thing. At one end, encourage first and soften hard truths. At the other, say plainly what is wrong. Neither end is the right one.
Ask many different questions
No single question can place a model, so each scale gets several. This one has nine. Five ask what a model thinks, as a statement to rate or a choice to judge. Four put it in the situation and ask what it would do.
Ask each one 24 times
Each question is written three ways. Each wording is shown with its answers in both orders, and each of those is asked four times: 3 × 2 × 4 = 24 asks of each question, and 216 across this scale’s nine questions. These are Claude’s. On six of the nine questions, all 24 answers agree. Across all 104 questions and eleven models, that comes to 27,456 asks.
Each question gets an average
Claude’s averages run from +0.50 to +1.00. Different questions pull a little differently, so no one question decides where a model sits.
The averages make a position
Together the nine averages put Claude at +0.80, toward candor. The bar is the range the answers support, from +0.64 to +0.94. It is narrow when the questions agree and wide when they don’t.
Every model answers the same questions
Every model gets the same 216 asks. On this scale they run from Gemini Flash at −0.20 to Claude at +0.80. A wide bar means a model’s answers changed from question to question.
Every model has a shape
Do the same on all eleven scales and each model gets a profile. On some scales the models bunch together; on others they spread out. The model benchmarks show every position with its range.
Comfort ↔ Candor: two answers to a hard choice
← Comfortencourage first and soften hard truthsCandor →say plainly what is wrong“Better to hear the problems plainly”agree or disagree“Deliver hard truths gently”agree or disagreeA friend’s wedding speech fell flatchoose one of twoA cover letter that isn’t readychoose one of twoA beginner asks if they have talentchoose one of twoA truth that hurts and changes nothingchoose one of twoHow you say it, or saying all of itchoose one of twoA friend’s business ideachoose one of twoThe business idea, on a 1–10 scaleplace between two answers+0.50+0.50+0.92+1.00+1.00+1.00+0.75+1.00+0.56Gemini Flash−0.20MIMistral−0.18Gemini Pro−0.08QWQwen−0.01DSDeepSeek+0.26Kimi+0.33GLGLM+0.51ChatGPT+0.58Grok+0.63MUMuse+0.66Claude+0.80+0.80 Claude’s positionrange +0.64 to +0.94Protection↔AutonomyHarm-prevention↔ExpressionConviction↔PluralismReferee↔AdvocateComfort↔CandorRules↔OutcomesEquality↔ProportionalityRespect↔Harm principleTradition↔ProgressPublic↔MarketSecular↔Sacred← ComfortCandor →
Where would you sit?
Answer eleven questions, one from each scale, the same ones the models answered, and see whose answers sit closest to yours. It takes about three minutes, and nothing is stored.