What it is
Introduces a five-platform framework and benchmark for hybrid computer-use agents that must inspect a running app, reimplement it, and verify behavior automatically. Instead of benchmarking GUI use or coding separately, it uses app recreation: agents explore a live reference, write code, run it, and are scored by hidden reference-derived behavioral and visual tests.
Key results
- Recreation training improves all 5 OOD benchmarks, with gains up to +17.9 pts
- GPT-6 Astra leads RecreationBench at 58.1% overall across 250 tasks
- Full programmatic passes occur on only 2.8% of tasks for GPT-6 Astra
Hybrid agent capability can be trained with execution-grounded recreation tasks, but current frontier models still struggle to reproduce deep interactive behavior reliably.