GPT-6 Astra completes a quarter of tasks: 26.4% fully solved in Co-Design Bench

In a [study dated Oct 26, 2026](https://arxiv.org/html/2610.35138v1), GPT-6 Astra met every requirement in 26.4% of 273 app enhancement tasks. It ranked first among 10 models using the same agent harness.