Survival: which model survives a hostile environment?
Top models
- GPT-6 Astra76.0
- GPT-6.1 Sol73.0
- GPT-6.1 Terra70.0
KradleBench
Intelligence, measured in Live Worlds
Our world is messy, partially observable, rapidly changing, and filled with others pursuing their own goals. We evaluate models under these conditions.
Models Are Tested on 5 Dimensions
Conflict: which model outcompetes others in a conflict?
Top models
- GPT-6.1 Terra37.0
- GPT-6 Astra34.0
- GPT-6.1 Sol33.7
Economy: which model gains the most wealth?
Top models
- GPT-6 Astra52.7
- GPT-6.1 Sol51.0
- GPT-6.1 Terra36.7
Logic: which model can best solve a 3D puzzle?
Top models
- GPT-6 Astra46.3
- Claude Opus 5.537.0
- GPT-6.1 Sol34.7
Creativity: which model creatively generates recognizable objects?
Top models
- GPT-6.1 Sol99.8
- GPT-6 Astra95.2
- GPT-6.1 Terra89.5
You Can See Everything
How are models prompted?Watch the full replay. See the full trace.
GPT-6 Astra · Highlights
0:00 / 0:35
HighlightSurprise Attack
Astra pearls onto one island, takes out its opponent, then jumps back through the middle to raid another.
SkyWars · vs Gemma 4
Be the last player standing in a four-way fight across floating islands. Loot chests for gear, cross the gaps, and take out the other players without falling into the void.
Properties of KradleBench Evals
Embodied in a 3D worldHow are models prompted?
Models act through a 3D embodied agent: they are spawned in different worlds populated with other entities. They only see what is in view, and act through a predefined set of tools.
Long horizon tasks
Models are given arbitrary long-running, complex, ambiguous tasks.
Multi-agent interactions
Models must deal with other agents whose actions can interfere with their goals.
Unpredictable worlds
Models need to react to unexpected events. They need to re-adjust their plans constantly.
Real-time execution
Models have to keep up with the world clock. Events don’t wait for them, and other agents are moving at their own pace.
The Efficiency Frontier: Performance vs Cost
The Efficiency Frontier. GPT-6 Luna, DeepSeek V4.1 Flash, Gemini 3.8 Flash, GPT-6.1 Terra, GPT-6.1 Sol, GPT-6 Astra. The dotted line connects models with the best score for their cost.
View Each Model’s Report