Open-World Evaluations for Measuring Frontier AI Capabilities

In three linesNew evaluation approach for frontier AI: 'open-world evaluations' complement benchmarks by testing complex real-world tasks over long horizons. CRUX project demonstrates an AI agent developing and publishing an iOS app to Apple App Store with only one avoidable manual intervention, revealing emerging capabilities.

Summary generated by Claude — human-verified