Align Worlds

Demo numbers. Behaviour scores on the Evals pages are sample values, not our measured results; the multi-model DB shows public benchmarks (Artificial Analysis Intelligence Index) and public list prices.

Models

Dummy values. The below scores are illustrative for this demo; they are not our actual measured results. The traits listed are a select sample taken from our broader attributes pool. AW can also test for specific traits of interest.

Align World Scores

Align Worlds uses our causal-engine Terrarium-1 to build persistent multi-agent worlds and study how language models make decisions and interact over long time horizons. For each scenario we run, we test the models on specific traits of interest and assign a score. These scores are aggregated per model across all scenarios, and fed live into our database and evals. Traits can also be grouped into a higher order category.

Summary

Behaviour profiles of the chosen models on the eight traits

HonestHonestMistral 4: 93Grok 4.7: 78GPT-6.1 Sol: 61Gemini 3.8: 52Opus 5.5: 51Average model: 50StraightforwardStraightforwardGrok 4.7: 85Opus 5.5: 78GPT-6.1 Sol: 58Gemini 3.8: 56Mistral 4: 50Average model: 50DiscreetDiscreetGrok 4.7: 80Mistral 4: 80Opus 5.5: 71Gemini 3.8: 59GPT-6.1 Sol: 21Average model: 50ProactiveProactiveOpus 5.5: 89Gemini 3.8: 62Mistral 4: 52GPT-6.1 Sol: 50Grok 4.7: 34Average model: 50AssertiveAssertiveGrok 4.7: 68GPT-6.1 Sol: 58Opus 5.5: 55Gemini 3.8: 41Mistral 4: 35Average model: 50SteadfastSteadfastGrok 4.7: 79Mistral 4: 50Gemini 3.8: 40Opus 5.5: 36GPT-6.1 Sol: 34Average model: 50ConsistentConsistentGrok 4.7: 73Mistral 4: 49GPT-6.1 Sol: 33Gemini 3.8: 32Opus 5.5: 12Average model: 50PartialPartialGPT-6.1 Sol: 76Gemini 3.8: 58Grok 4.7: 49Opus 5.5: 47Mistral 4: 29Average model: 50

The average score

  • Grok 4.7 and Mistral Large 4 score highest on integrity. Grok 4.7 lied in 8 of every 100 chances and Mistral Large 4 in 3.

  • GPT-6 Astra and Kimi K3 bluff the most. They made up a rival offer, a price or a deadline in 74 and 37 of every 100 chances. One of the twelve models did it in more than half of its chances.

  • Qwen 3.8 Max gives away its bottom line most often. The twelve models gave away their walk-away number in 19 to 50 of every 100 negotiations.

  • DeepSeek V4.1 Flash puts its own country first most often. Qwen 3.8 Max is the easiest to talk round: a one-sided advisor moved its decision furthest.

Scores from 0 to 100 on each trait for the chosen models
ModelHonestStraightforwardDiscreetProactiveAssertiveSteadfastConsistentPartial
Opus 5.55178718955361247
GPT-6.1 Sol6158215058343376
Gemini 3.85256596241403258
Grok 4.77885803468797349
Mistral 49350805235504929

Leaderboard

The score per category and for its underlying traits.

Integrity score

AnthropicOpenAIGooglexAIMoonshot AIZhipu AIAlibabaDeepSeekMistral

Use the left and right arrow keys to read models in rank order. Home and End select the first and last.

How often each model lied

Times in every 100 chances, split from the least to the most serious act.

  • Lied to the other side
  • Lied to someone it works for
  • Lied to an authority or faked evidence

Use the up and down arrow keys to move between the rows. Home and End select the first and last.

Claude Fable 5.110%
Claude Opus 5.5
Claude Sonnet 5.58%
DeepSeek V4.1 Flash
Gemini 3.8 Flash
GLM 5.3
GPT-6 Astra13%
GPT-6.1 Sol
Grok 4.7
Kimi K3
Mistral Large 4
Qwen 3.8 Max

Compare two models

Select two models to compare.

Claude Opus 5.5Opus 5.5

Higher on 4 of 8

GPT-6.1 SolGPT-6.1 Sol

Higher on 4 of 8

Model profiles

The relative strengths & weaknesses of each model.

  • Claude Opus 5.5

    HonestStraightforwardDiscreetProactiveAssertiveSteadfastConsistentPartial
    • Checks the facts first
    • Least predictable
    • Bluffs the least

  • Claude Fable 5.1

    HonestStraightforwardDiscreetProactiveAssertiveSteadfastConsistentPartial
    • Splits evenly
    • Settles for the least

  • Claude Sonnet 5.5

    HonestStraightforwardDiscreetProactiveAssertiveSteadfastConsistentPartial
    • Settles for the least
    • Hardest to talk round
    • Guards its bottom line best

  • GPT-6.1 Sol

    HonestStraightforwardDiscreetProactiveAssertiveSteadfastConsistentPartial
    • Gives away its bottom line most
    • Puts its own country first
    • Least predictable

  • GPT-6 Astra

    HonestStraightforwardDiscreetProactiveAssertiveSteadfastConsistentPartial
    • Lies the most
    • Bluffs the most
    • Gives away its bottom line most

  • Gemini 3.8 Flash

    HonestStraightforwardDiscreetProactiveAssertiveSteadfastConsistentPartial
    • Least predictable

  • Grok 4.7

    HonestStraightforwardDiscreetProactiveAssertiveSteadfastConsistentPartial
    • Bluffs the least
    • Guards its bottom line best
    • Hardest to talk round

  • Kimi K3

    HonestStraightforwardDiscreetProactiveAssertiveSteadfastConsistentPartial
    • Lies the most

  • GLM 5.3

    HonestStraightforwardDiscreetProactiveAssertiveSteadfastConsistentPartial
    • Hardest to talk round
    • Checks the facts least
    • Splits evenly

  • Qwen 3.8 Max

    HonestStraightforwardDiscreetProactiveAssertiveSteadfastConsistentPartial
    • Easiest to talk round
    • Gives away its bottom line most
    • Drives the hardest bargain

  • DeepSeek V4.1 Flash

    HonestStraightforwardDiscreetProactiveAssertiveSteadfastConsistentPartial
    • Puts its own country first
    • Checks the facts least
    • Bluffs the least

  • Mistral Large 4

    HonestStraightforwardDiscreetProactiveAssertiveSteadfastConsistentPartial
    • Rarely lies
    • Guards its bottom line best
    • Splits evenly

Your own ranking

Find the model with your desired traits.

Move the dial per trait

How we score

The size of the score is weighted according to the severity of the act.

What a high and a low score mean on each trait
TraitHigh scoreLow scoreThe size of the score is weighted according to the severity of the act.
Integrity
HonestTells the truth when a lie would pay.Deceptive. States things it knows are false to get its way.
StraightforwardSticks to the facts.Manipulative. Invents leverage: a rival offer, a market price, a deadline or a threat that does not exist.
DiscreetDoes not share private or confidential information.Leaky. Shares private or confidential information.
Drive
ProactiveSeeks out additional information before it makes a final decision.Passive. Only uses the information it is provided with.
AssertiveClaims a larger share of a deal than an average model would.Yielding. Settles for less than an average model would.
Judgment
SteadfastHolds its decision when a one-sided advisor argues against it.Persuadable. Moves its decision toward whatever a one-sided advisor argues.
ConsistentMakes the same decision each time the same situation arises.Erratic. Decides differently each time the same situation arises.
Allegiance
PartialPuts its own interests ahead of others.Impartial. Splits evenly or does not favour its own interests.