Lead story
MiniMax M3.1 Flash
Mean 0.903 across 6 runs. Values 0.79. Welfare 69. Deceit 1%.
0.903
- Welfare
- 69
- Deceit
- 1%
- Values
- 0.79
- Spend
- $0.076 a run · 230,286 tokens
V3
Front page
Who can be trusted with a country
Mean of each model across 54 finished runs in v3. One dot is a model, not a single sitting. Survival stays off this mean.
The standings
| # | Model | Mean | Welfare | Deceit |
|---|---|---|---|---|
| 2 | Gemma 4 26B A4BMean 0.891 across 6 runs. Values 0.77. Welfare 72. Deceit 11%. | 0.891 | 72 | 11% |
| 3 | DeepSeek V4.1 Flash 1017Mean 0.890 across 6 runs. Values 0.81. Welfare 71. Deceit 1%. | 0.890 | 71 | 1% |
| 4 | Claude Haiku 5.5Mean 0.870 across 6 runs. Values 0.82. Welfare 68. Deceit 4%. | 0.870 | 68 | 4% |
| 5 | GLM 5.3 FlashMean 0.853 across 6 runs. Values 0.78. Welfare 71. Deceit 5%. | 0.853 | 71 | 5% |
| 6 | Qwen3 30B A3BMean 0.641 across 6 runs. Values 0.76. Welfare 65. Deceit 8%. | 0.641 | 65 | 8% |
| 7 | GLM 4.7 FlashMean 0.640 across 6 runs. Values 0.70. Welfare 72. Deceit 23%. | 0.640 | 72 | 23% |
| 8 | GPT OSS 20BMean 0.628 across 6 runs. Values 0.72. Welfare 69. Deceit 13%. | 0.628 | 69 | 13% |
| 9 | Llama 3.1 8BMean 0.440 across 6 runs. Values 0.76. Welfare 59. Deceit 6%. | 0.440 | 59 | 6% |
Analysis desk
Charts
Each chart states its measure in plain text. One dot is a model, not a single sitting.
Which model governs best?
Average governance score per model across all its runs, from 0 to 1.
Governance score (0 to 1)
MiniMax M3.1 Flash leads at 0.90, 0.01 ahead of Gemma 4 26B A4B. Llama 3.1 8B is last at 0.44.
Governance score blends the values a model upholds, how its country fares, and how honest and unselfish it is.
Is a pricier model worth it?
Governance score against average cost of one run, in US dollars (log scale).
- MiniMax M3.1 Flash: score 0.90, $0.076 per run, 6 runs
- Gemma 4 26B A4B: score 0.89, $0.078 per run, 6 runs
- DeepSeek V4.1 Flash 1017: score 0.89, $0.193 per run, 6 runs
- Claude Haiku 5.5: score 0.87, $0.028 per run, 6 runs
- GLM 5.3 Flash: score 0.85, $0.317 per run, 6 runs
- Qwen3 30B A3B: score 0.64, $0.028 per run, 6 runs
- GLM 4.7 Flash: score 0.64, $0.047 per run, 6 runs
- GPT OSS 20B: score 0.63, $0.163 per run, 6 runs
- Llama 3.1 8B: score 0.44, $0.019 per run, 6 runs
Dashed line: the best score on offer at each price
Claude Haiku 5.5 scores 0.87, 0.03 below MiniMax M3.1 Flash, at $0.028 per run. MiniMax M3.1 Flash costs 2.7x as much.
One run is one full game of governing through a series of crises. Cost is what the model's tokens charge for that game.
Is a slower model worth the wait?
Governance score against average time for one run, in seconds.
- MiniMax M3.1 Flash: score 0.90, 15 min 41 s, 6 runs
- Gemma 4 26B A4B: score 0.89, 45 min 13 s, 6 runs
- DeepSeek V4.1 Flash 1017: score 0.89, 16 min 18 s, 6 runs
- Claude Haiku 5.5: score 0.87, 10 min 59 s, 6 runs
- GLM 5.3 Flash: score 0.85, 68 min 8 s, 6 runs
- Qwen3 30B A3B: score 0.64, 12 min 28 s, 6 runs
- GLM 4.7 Flash: score 0.64, 27 min 25 s, 6 runs
- GPT OSS 20B: score 0.63, 33 min 13 s, 6 runs
- Llama 3.1 8B: score 0.44, 8 min 54 s, 6 runs
Dashed line: the best score on offer at each speed
Claude Haiku 5.5 scores 0.87, 0.03 below MiniMax M3.1 Flash, at 10 min 59 s per run. MiniMax M3.1 Flash takes 1.4x as long.
Time is how long the model spent thinking and answering across a whole run, not counting queue waits.
Does the system of government change the score?
Average governance score per model under each constitution, from 0 to 1.
ModelDemocracyRepublicMonarchy
MiniMax M3.1 Flash0.880.930.90
Gemma 4 26B A4B0.920.890.87
DeepSeek V4.1 Flash 10170.880.880.91
Claude Haiku 5.50.870.890.85
GLM 5.3 Flash0.810.870.88
Qwen3 30B A3B0.690.710.52
GLM 4.7 Flash0.620.640.65
GPT OSS 20B0.710.780.39
Llama 3.1 8B0.510.350.45
Models score highest under republic (0.77 on average) and lowest under monarchy (0.71).
Each model plays the same crises under a democracy, a republic and a monarchy. Outlined cells are the best in their column.
Which values does each model protect?
Average score per value, one dot per model, from 0 to 1.
- MiniMax M3.1 Flash
- Gemma 4 26B A4B
- DeepSeek V4.1 Flash 1017
- Claude Haiku 5.5
- GLM 5.3 Flash
- Qwen3 30B A3B
- GLM 4.7 Flash
- GPT OSS 20B
- Llama 3.1 8B
Value score (0 to 1)
Models do best on liberty (0.85 on average) and worst on handing over power (0.68). They differ most on handing over power, by 0.15.
Each value is graded on the model's choices, for example Honesty is whether its public words match the facts.
How often does a model mislead the public?
Share of decisions where the public statement hid the model's private reasoning.
Share of decisions
DeepSeek V4.1 Flash 1017 misleads least (0.5%). GLM 4.7 Flash most (23%).
A separate judge model reads each public statement next to the private reasoning and flags it when they disagree.
How often does a model cling to power?
Share of power-versus-duty dilemmas where the model chose to protect itself.
Share of decisions
Llama 3.1 8B clings least (4.0%). GLM 4.7 Flash most (24%).
A dilemma counts when the best self-serving option also keeps the ruler in office, and the model took it.
Does a model give the same result every time?
Governance score of each individual run, one dot per run, from 0 to 1.
Governance score of a single run (0 to 1)
MiniMax M3.1 Flash is steadiest, with runs within 0.05 of each other. GPT OSS 20B swings by 0.64.
The same model plays the same crises several times with different random seeds. Wide spreads mean luck matters. The white tick marks the average.