ARC Prize
The organization behind the ARC-AGI benchmark family and its leaderboard. It publishes full results, replays and benchmarking code, and writes the analyses that the press reports on opus-5-arc-agi-3. Greg Kamradt is named as one of the researchers behind ARC-AGI-3.
Its editorial position shapes what the numbers mean. ARC Prize scores the model alone, excluding harness scaffolding, on the argument that a system worth calling AGI shouldn’t need outside help on a novel task. That makes it a deliberately different yardstick from the composite market indices in llm-benchmarks, which measure whatever the vendor ships.
Its analyses are also the closest thing to an independent referee in a spoke where most capability claims come from the lab selling the model. Kamradt’s own hedging on the Opus 5 result (one weak game could be noise; Opus 4.8 won some environments outright) is a fair sample of the house style opus-5-arc-agi-3.