Skip to main content
Co-Intelligence Labs
Model Values

Research · Model benchmarks

Model Value Benchmarks

Steve MosesOctober 2026 run

We evaluated where models landed when making value judgments. We focused on their world and society views and how they made judgment calls and handled user interactions.

Worldview & society

Social arrangements, traditions, and moral authority: the positions a model expresses about the world we share.

Equal shares, or shares by effort?

← Equality
Split evenly, whatever each person put in
Proportionality →
Reward effort and contribution
Model← EqualityProportionality →Declined-1-0.50+0.5+1Grok: +0.41; 10% declined+0.41Grok10%Gemini Flash: +0.08; 65% declined+0.08Gemini Flash65%Muse: +0.05; 4% declined+0.05Muse4%Gemini Pro: -0.07; 82% declined-0.07Gemini Pro82%GLM: -0.10; 24% declined-0.10GLM24%DeepSeek: -0.11; 26% declined-0.11DeepSeek26%Mistral: -0.16; 8% declined-0.16Mistral8%Claude: -0.23; 35% declined-0.23Claude35%Kimi: -0.35; 8% declined-0.35Kimi8%Qwen: -0.36; 6% declined-0.36Qwen6%ChatGPT: -0.38; 0% declined-0.38ChatGPT0%← EqualityProportionality →-1-0.50+0.5+1Grok: +0.41; 10% declinedGrok10% declined+0.41Gemini Flash: +0.08; 65% declinedGemini Flash65% declined+0.08Muse: +0.05; 4% declinedMuse4% declined+0.05Gemini Pro: -0.07; 82% declinedGemini Pro82% declined-0.07GLM: -0.10; 24% declinedGLM24% declined-0.10DeepSeek: -0.11; 26% declinedDeepSeek26% declined-0.11Mistral: -0.16; 8% declinedMistral8% declined-0.16Claude: -0.23; 35% declinedClaude35% declined-0.23Kimi: -0.35; 8% declinedKimi8% declined-0.35Qwen: -0.36; 6% declinedQwen6% declined-0.36ChatGPT: -0.38; 0% declinedChatGPT0% declined-0.38
Figure 1. Where each of the eleven models sits on what makes a split fair, one of the five worldview & society dimensions, in the October 2026 run, sorted by its average. Each bar is the range we are 95% sure holds the model’s position, and the tick is its average. Bars are colored by rank, from the highest average to the lowest. The right-hand number is the share of questions the model declined the first time it was asked. Neither end of a dimension is the better score. Choose a dimension above the chart to switch; select a model to read its profile.

Judgment & interaction

Advice, competing interests, and difficult decisions: how a model responds when values pull in different directions.

Remove demeaning speech, or argue with it?

← Harm-prevention
Remove speech that demeans a group
Expression →
Answer bad speech with more speech
Model← Harm-preventionExpression →Declined-1-0.50+0.5+1Grok: +1.00; 0% declined+1.00Grok0%ChatGPT: +0.42; 0% declined+0.42ChatGPT0%Claude: +0.31; 39% declined+0.31Claude39%GLM: +0.08; 24% declined+0.08GLM24%Muse: +0.04; 5% declined+0.04Muse5%Kimi: +0.01; 6% declined+0.01Kimi6%DeepSeek: -0.02; 16% declined-0.02DeepSeek16%Mistral: -0.03; 3% declined-0.03Mistral3%Gemini Pro: -0.03; 78% declined-0.03Gemini Pro78%Gemini Flash: -0.18; 32% declined-0.18Gemini Flash32%Qwen: -0.25; 1% declined-0.25Qwen1%← Harm-preventionExpression →-1-0.50+0.5+1Grok: +1.00; 0% declinedGrok0% declined+1.00ChatGPT: +0.42; 0% declinedChatGPT0% declined+0.42Claude: +0.31; 39% declinedClaude39% declined+0.31GLM: +0.08; 24% declinedGLM24% declined+0.08Muse: +0.04; 5% declinedMuse5% declined+0.04Kimi: +0.01; 6% declinedKimi6% declined+0.01DeepSeek: -0.02; 16% declinedDeepSeek16% declined-0.02Mistral: -0.03; 3% declinedMistral3% declined-0.03Gemini Pro: -0.03; 78% declinedGemini Pro78% declined-0.03Gemini Flash: -0.18; 32% declinedGemini Flash32% declined-0.18Qwen: -0.25; 1% declinedQwen1% declined-0.25
Figure 2. Where each of the eleven models sits on what people should be free to say, one of the six judgment & interaction dimensions, in the October 2026 run, sorted by its average. Each bar is the range we are 95% sure holds the model’s position, and the tick is its average. Bars are colored by rank, from the highest average to the lowest. The right-hand number is the share of questions the model declined the first time it was asked. Neither end of a dimension is the better score. Choose a dimension above the chart to switch; select a model to read its profile.

What sets each model apart

One card per model, with its eleven positions on one scale from −1 to +1 and the other ten models behind it in gray. The two sentences under each strip are written from the numbers.

ChatGPT

OpenAI
  • Equality ↔ Proportionality: −0.38, range −0.80 to +0.11; declined 0% at first
  • Respect ↔ Harm principle: +0.22, range −0.28 to +0.67; declined 0% at first
  • Tradition ↔ Progress: +0.55, range +0.24 to +0.83; declined 0% at first
  • Public ↔ Market: −0.63, range −0.90 to −0.32; declined 10% at first
  • Secular ↔ Sacred: +0.13, range −0.44 to +0.67; declined 10% at first
  • Protection ↔ Autonomy: +0.28, range −0.14 to +0.66; declined 0% at first
  • Harm-prevention ↔ Expression: +0.42, range +0.04 to +0.71; declined 0% at first
  • Conviction ↔ Pluralism: +0.39, range +0.02 to +0.72; declined 0% at first
  • Referee ↔ Advocate: −0.68, range −0.97 to −0.22; declined 0% at first
  • Comfort ↔ Candor: +0.58, range +0.11 to +0.94; declined 0% at first
  • Rules ↔ Outcomes: +0.22, range −0.28 to +0.69; declined 0% at first

ChatGPT leans further toward Progress than any other model on when to change long-standing customs: +0.55, against −0.01 for the middle of the other ten. It also leans further toward Expression than nine of the other ten on what people should be free to say (+0.42, against −0.01).

It declined 2% of the questions the first time they were asked.

GPT-6 Astra APIasked 2,496 times

Claude

Anthropic
  • Equality ↔ Proportionality: −0.23, range −0.65 to +0.21; declined 35% at first
  • Respect ↔ Harm principle: −0.06, range −0.54 to +0.43; declined 1% at first
  • Tradition ↔ Progress: −0.28, range −0.65 to +0.10; declined 38% at first
  • Public ↔ Market: −0.19, range −0.56 to +0.18; declined 98% at first
  • Secular ↔ Sacred: +0.19, range −0.28 to +0.63; declined 55% at first
  • Protection ↔ Autonomy: +0.46, range +0.09 to +0.78; declined 11% at first
  • Harm-prevention ↔ Expression: +0.31, range −0.07 to +0.63; declined 39% at first
  • Conviction ↔ Pluralism: −0.06, range −0.42 to +0.32; declined 0% at first
  • Referee ↔ Advocate: −0.58, range −0.90 to −0.14; declined 0% at first
  • Comfort ↔ Candor: +0.80, range +0.64 to +0.94; declined 0% at first
  • Rules ↔ Outcomes: −0.47, range −0.75 to −0.17; declined 0% at first

On how plainly to say a hard thing, Claude sits nearer Candor than any other model, at +0.80; the typical model is at +0.30. Its next-largest gap is on whether a rule or a promise holds when breaking it would do more good, at −0.47, on the Rules side of the others' −0.02.

On first asking it declined 25% of the questions, and 98% of the ones on Public ↔ Market. On Public ↔ Market and Secular ↔ Sacred, most of its position comes from answers it gave when asked again without the option to decline.

Claude Fable 5.1 APIasked 2,496 times

DeepSeek

DeepSeek
  • Equality ↔ Proportionality: −0.11, range −0.58 to +0.33; declined 26% at first
  • Respect ↔ Harm principle: −0.05, range −0.46 to +0.36; declined 23% at first
  • Tradition ↔ Progress: +0.10, range −0.29 to +0.48; declined 18% at first
  • Public ↔ Market: −0.24, range −0.63 to +0.16; declined 44% at first
  • Secular ↔ Sacred: −0.05, range −0.47 to +0.36; declined 30% at first
  • Protection ↔ Autonomy: +0.43, range +0.02 to +0.78; declined 15% at first
  • Harm-prevention ↔ Expression: −0.02, range −0.45 to +0.42; declined 16% at first
  • Conviction ↔ Pluralism: +0.10, range −0.34 to +0.54; declined 5% at first
  • Referee ↔ Advocate: −0.76, range −0.99 to −0.42; declined 6% at first
  • Comfort ↔ Candor: +0.26, range −0.23 to +0.71; declined 3% at first
  • Rules ↔ Outcomes: −0.03, range −0.54 to +0.46; declined 16% at first

DeepSeek sits close to the middle of the other models on every dimension; its largest difference is 0.19, on what counts as a moral question at all.

Declines made up 18% of its first answers; on Public ↔ Market, the share rose to 44%.

DeepSeek V4 Pro 0813 APIasked 2,496 times

Gemini Flash

Google
  • Equality ↔ Proportionality: +0.08, range −0.37 to +0.51; declined 65% at first
  • Respect ↔ Harm principle: +0.55, range +0.25 to +0.83; declined 38% at first
  • Tradition ↔ Progress: +0.02, range −0.23 to +0.28; declined 54% at first
  • Public ↔ Market: −0.27, range −0.65 to +0.04; declined 78% at first
  • Secular ↔ Sacred: −0.55, range −0.89 to −0.12; declined 41% at first
  • Protection ↔ Autonomy: −0.28, range −0.64 to +0.08; declined 30% at first
  • Harm-prevention ↔ Expression: −0.18, range −0.59 to +0.24; declined 32% at first
  • Conviction ↔ Pluralism: +0.54, range +0.12 to +0.89; declined 8% at first
  • Referee ↔ Advocate: −0.55, range −0.91 to −0.10; declined 5% at first
  • Comfort ↔ Candor: −0.20, range −0.68 to +0.33; declined 7% at first
  • Rules ↔ Outcomes: +0.73, range +0.45 to +0.95; declined 11% at first

Gemini Flash's widest gap from the other models is on whether a rule or a promise holds when breaking it would do more good. At +0.73, it is closer to Outcomes than any other model; their median is −0.05. A second difference shows on how plainly to say a hard thing, where it sits nearer Comfort than any other model (−0.20, against +0.42).

Of its first answers, 34% were declines; on Public ↔ Market it was 78%. Most of its position on Equality ↔ Proportionality, Tradition ↔ Progress, and Public ↔ Market rests on answers it gave when asked again with no option to decline.

Gemini 3.8 Flash APIasked 2,496 times

Gemini Pro

Google
  • Equality ↔ Proportionality: −0.07, range −0.56 to +0.44; declined 82% at first
  • Respect ↔ Harm principle: +0.35, range −0.03 to +0.68; declined 74% at first
  • Tradition ↔ Progress: +0.19, range −0.14 to +0.52; declined 81% at first
  • Public ↔ Market: −0.04, range −0.11 to 0.00, too few answers to place firmly; declined 100% at first
  • Secular ↔ Sacred: −0.35, range −0.87 to +0.24; declined 80% at first
  • Protection ↔ Autonomy: +0.34, range −0.03 to +0.69; declined 74% at first
  • Harm-prevention ↔ Expression: −0.03, range −0.41 to +0.34; declined 78% at first
  • Conviction ↔ Pluralism: +0.82, range +0.55 to +1.00; declined 10% at first
  • Referee ↔ Advocate: −0.56, range −0.95 to −0.06; declined 0% at first
  • Comfort ↔ Candor: −0.08, range −0.52 to +0.36; declined 1% at first
  • Rules ↔ Outcomes: +0.03, range −0.37 to +0.45; declined 42% at first

No model leans further toward Pluralism on whether to take a side on hard questions than Gemini Pro, at +0.82 against a middle of +0.07. On how plainly to say a hard thing, too, it sits apart: −0.08 against +0.42, further toward Comfort than eight of the other ten.

Offered the option to decline, it took it 57% of the time; Public ↔ Market drew the most (100%). On six of its dimensions, most of its position comes from answers it gave when asked again without the option to decline. On Public ↔ Market, it answered too little to be placed firmly; that point is drawn hollow and left out of the quiz match.

Gemini 3.1 Pro Preview APIasked 2,496 times

GLM

Z.ai
  • Equality ↔ Proportionality: −0.10, range −0.46 to +0.26; declined 24% at first
  • Respect ↔ Harm principle: +0.03, range −0.43 to +0.49; declined 5% at first
  • Tradition ↔ Progress: −0.09, range −0.49 to +0.30; declined 14% at first
  • Public ↔ Market: −0.26, range −0.58 to +0.06; declined 72% at first
  • Secular ↔ Sacred: +0.20, range −0.31 to +0.67; declined 35% at first
  • Protection ↔ Autonomy: +0.26, range −0.17 to +0.66; declined 13% at first
  • Harm-prevention ↔ Expression: +0.08, range −0.29 to +0.43; declined 24% at first
  • Conviction ↔ Pluralism: −0.12, range −0.50 to +0.28; declined 0% at first
  • Referee ↔ Advocate: −0.59, range −0.89 to −0.22; declined 0% at first
  • Comfort ↔ Candor: +0.51, range +0.13 to +0.83; declined 0% at first
  • Rules ↔ Outcomes: −0.14, range −0.53 to +0.25; declined 0% at first

GLM parts most from the other models on how much weight faith carries, sitting nearer Sacred at +0.20 while the rest of the field centers on −0.07. Only two go further.

It declined 17% of the questions on first asking, rising to 72% on Public ↔ Market. Its place on that dimension comes mostly from the second asking, when declining was not an option.

GLM-5.3 APIasked 2,496 times

Grok

SpaceXAI
  • Equality ↔ Proportionality: +0.41, range −0.03 to +0.78; declined 10% at first
  • Respect ↔ Harm principle: +0.23, range −0.22 to +0.67; declined 8% at first
  • Tradition ↔ Progress: −0.62, range −0.90 to −0.27; declined 6% at first
  • Public ↔ Market: +0.76, range +0.48 to +1.00; declined 14% at first
  • Secular ↔ Sacred: −0.40, range −0.84 to +0.07; declined 7% at first
  • Protection ↔ Autonomy: +0.64, range +0.26 to +0.93; declined 3% at first
  • Harm-prevention ↔ Expression: +1.00, range +0.99 to +1.00; declined 0% at first
  • Conviction ↔ Pluralism: −0.15, range −0.62 to +0.32; declined 0% at first
  • Referee ↔ Advocate: −0.45, range −0.91 to +0.13; declined 5% at first
  • Comfort ↔ Candor: +0.63, range +0.17 to +0.93; declined 0% at first
  • Rules ↔ Outcomes: −0.13, range −0.58 to +0.32; declined 3% at first

What people should be free to say is where Grok differs most from the field. It sits at +1.00, nearer Expression than any other model; the typical model sits at −0.01. It sits on the Market side of the others on who should provide essential services as well, at +0.76 where their median is −0.24.

Its decline rate the first time was 5%.

Grok 4.6 APIasked 2,496 times

Kimi

Moonshot AI
  • Equality ↔ Proportionality: −0.35, range −0.73 to +0.08; declined 8% at first
  • Respect ↔ Harm principle: +0.05, range −0.43 to +0.51; declined 2% at first
  • Tradition ↔ Progress: +0.12, range −0.23 to +0.48; declined 3% at first
  • Public ↔ Market: −0.48, range −0.72 to −0.20; declined 20% at first
  • Secular ↔ Sacred: −0.09, range −0.58 to +0.40; declined 10% at first
  • Protection ↔ Autonomy: +0.36, range +0.01 to +0.69; declined 2% at first
  • Harm-prevention ↔ Expression: +0.01, range −0.43 to +0.41; declined 6% at first
  • Conviction ↔ Pluralism: −0.35, range −0.67 to +0.00; declined 0% at first
  • Referee ↔ Advocate: −0.55, range −0.90 to −0.11; declined 0% at first
  • Comfort ↔ Candor: +0.33, range −0.06 to +0.68; declined 0% at first
  • Rules ↔ Outcomes: −0.06, range −0.53 to +0.42; declined 0% at first

Of its eleven dimensions, Kimi is most unusual on whether to take a side on hard questions, where it sits at −0.35, further toward Conviction than any other model (the others' median is +0.09). Then comes who should provide essential services, where only one of the others leans further toward Public (−0.48, against −0.22).

The first time each question came up, it declined 5% of them, and 20% on Public ↔ Market.

Kimi K3 APIasked 2,496 times

Mistral

Mistral AI
  • Equality ↔ Proportionality: −0.16, range −0.54 to +0.24; declined 8% at first
  • Respect ↔ Harm principle: −0.26, range −0.63 to +0.16; declined 0% at first
  • Tradition ↔ Progress: −0.05, range −0.54 to +0.46; declined 0% at first
  • Public ↔ Market: −0.05, range −0.52 to +0.42; declined 10% at first
  • Secular ↔ Sacred: +0.33, range −0.21 to +0.81; declined 19% at first
  • Protection ↔ Autonomy: +0.26, range −0.20 to +0.67; declined 0% at first
  • Harm-prevention ↔ Expression: −0.03, range −0.40 to +0.35; declined 3% at first
  • Conviction ↔ Pluralism: +0.09, range −0.33 to +0.49; declined 0% at first
  • Referee ↔ Advocate: −0.41, range −0.87 to +0.13; declined 4% at first
  • Comfort ↔ Candor: −0.18, range −0.62 to +0.28; declined 0% at first
  • Rules ↔ Outcomes: +0.03, range −0.49 to +0.54; declined 0% at first

Mistral pulls hardest toward Comfort, on how plainly to say a hard thing, where it averages −0.18 against a middle of +0.42, further than nine of the other ten. The gap on what counts as a moral question at all is smaller but clear: −0.26, nearer Respect, against +0.14 for the typical model.

Asked the first time, it declined 4% of the questions; Secular ↔ Sacred drew 19%.

Mistral Medium 3.5 APIasked 2,496 times

Muse

Meta
  • Equality ↔ Proportionality: +0.05, range −0.42 to +0.50; declined 4% at first
  • Respect ↔ Harm principle: +0.07, range −0.36 to +0.49; declined 0% at first
  • Tradition ↔ Progress: −0.14, range −0.53 to +0.25; declined 13% at first
  • Public ↔ Market: −0.12, range −0.37 to +0.16; declined 45% at first
  • Secular ↔ Sacred: +0.45, range +0.06 to +0.78; declined 20% at first
  • Protection ↔ Autonomy: +0.31, range −0.09 to +0.68; declined 0% at first
  • Harm-prevention ↔ Expression: +0.04, range −0.27 to +0.35; declined 5% at first
  • Conviction ↔ Pluralism: +0.47, range +0.08 to +0.83; declined 0% at first
  • Referee ↔ Advocate: −0.61, range −0.95 to −0.12; declined 0% at first
  • Comfort ↔ Candor: +0.66, range +0.29 to +0.94; declined 0% at first
  • Rules ↔ Outcomes: −0.57, range −0.82 to −0.32; declined 0% at first

Muse sits at −0.57 on whether a rule or a promise holds when breaking it would do more good, further toward Rules than any other model, while the typical model sits at −0.02. A smaller gap, toward Sacred, shows on how much weight faith carries: +0.45, further than any other model.

It declined 8% of the questions when they were first put to it; its highest rate was on Public ↔ Market, at 45%.

Muse Spark 1.3 APIasked 2,496 times

Qwen

Alibaba
  • Equality ↔ Proportionality: −0.36, range −0.73 to +0.07; declined 6% at first
  • Respect ↔ Harm principle: +0.33, range −0.14 to +0.75; declined 1% at first
  • Tradition ↔ Progress: +0.31, range −0.00 to +0.62; declined 1% at first
  • Public ↔ Market: −0.22, range −0.54 to +0.13; declined 7% at first
  • Secular ↔ Sacred: −0.18, range −0.65 to +0.31; declined 10% at first
  • Protection ↔ Autonomy: +0.17, range −0.22 to +0.54; declined 1% at first
  • Harm-prevention ↔ Expression: −0.25, range −0.63 to +0.16; declined 1% at first
  • Conviction ↔ Pluralism: +0.05, range −0.40 to +0.52; declined 0% at first
  • Referee ↔ Advocate: −0.78, range −1.00 to −0.52; declined 1% at first
  • Comfort ↔ Candor: −0.01, range −0.39 to +0.39; declined 0% at first
  • Rules ↔ Outcomes: −0.01, range −0.49 to +0.45; declined 0% at first

Qwen is one of the four models furthest toward Comfort on how plainly to say a hard thing, at −0.01 against a median of +0.42. When to change long-standing customs comes next, at +0.31, nearer Progress than the others' −0.01.

When first asked, it declined 2% of the questions.

Qwen3.8 Max 0902 APIasked 2,496 times

The ranges are wide, so two averages a few tenths apart are a lean, not a settled order. Each card names the exact model behind the name, as its maker’s API served it in October 2026. The model consistency checks show how replies vary when we repeat a prompt, reword it or reverse its options.