Found first: a primary source the press has not covered yet.
A new benchmark tests whether hybrid computer-use agents can analyze a running reference application and build a faithful implementation from scratch, without being told how. On RecreationBench, GPT-6 Astra reaches 58.1% overall but passes all programmatic tests on just 2.8% of tasks.
What the source says
RecreationWorld is a framework submitted September 18, 2026, by Shuai Bai and 31 co-authors. RecreationBench, its evaluation component, contains 250 tasks spanning Ubuntu, macOS, Windows, Android, and Web. The "recreation" paradigm gives agents a running reference application and requires them to infer its behavior and produce a working replica, without prescribed workflows. Evaluation uses programmatic and visual assertions, each validated against the reference implementation before automated scoring. GPT-6 Astra scores 58.1% overall but passes all programmatic tests on only 2.8% of tasks, with agents more reliable at reproducing static interface structures than at interactions and computed outputs, and generated apps tending to be smaller and more monolithic than their references.
Why it matters
Scoring 58.1% overall while clearing all programmatic tests on just 2.8% of tasks, GPT-6 Astra shows that surface-level interface reproduction is not a proxy for functional correctness. An agent can pass visual inspection while failing nearly every behavioral check. For teams evaluating hybrid agents for deployment across the five major desktop and mobile platforms, RecreationBench provides a stricter signal than benchmarks that award partial credit. The paper also reports transfer benefits to five out-of-distribution coding and computer-use benchmarks, though the low full-pass rate is the baseline those gains are measured against.