Benchmarks built for the endgame.
The hardest benchmarks are interactive. We put frontier agents inside games and real-world applications where success requires perception, reasoning, memory, planning, and action in environments that are 3D, non-deterministic, reactive, and constantly changing.
APK Arena
50+ hand-built phone-use levels that measure how well a model can interact with a phone. Each level is based on a real problem we encountered while running agents in production.
Emerald Bench
One continuous Pokémon Emerald playthrough per model. A test of planning, memory, and recovery over an entire game.
Game Gym
Agents are dropped into real games where they must perceive, reason, plan, adapt, and act in real time. These 3D, reactive, and non-deterministic environments require true general intelligence to complete. Saturating this benchmark is what we would consider AGI.