Co-Design Bench evaluated ten coding-agent models using difficult real web-app change requests selected from sessions involving repeated requests or rollbacks. GPT-6 Astra met roughly 70% of requirements on average but fully completed only 26.4% of tasks. Performance fell as tasks combined more change types, such as layout and interaction. The same Qwen3.8-27B model scored 75% or 56% on larger edits depending on its harness. The benchmark uses one fixed request and AI judges, so its results cannot be generalized into a failure rate for all real-world work.
Back to Tech
Co-Design Bench reports 26.4% full-task success for its best model
Test combined requirements and evaluate the model together with its execution harness.