Asana cuts model costs 76x in browser tests with GPT-6.1 Sol

Asana utilized GPT-6 Astra in Codex to optimize its StackAI browser agent workflow, achieving significant reductions in cost and runtime. According to the source, StackAI is a platform acquired by Asana that enables customers to automate work across business applications by building workflows that navigate websites, fill out forms, and gather information without writing code. StackAI CTO Frank Hidalgo directed GPT-6 Astra in Codex to investigate the agent, test improvements, and compare results, completing in about a week work that would have taken one to two months by hand. Through a 144-run study testing multiple frontier models, the optimized workflow on GPT-6.1 Sol averaged $0.47 in estimated model costs and about four minutes per run. This represented a 76-fold decrease in cost and a 5-fold increase in speed compared to the original production setup on Model B, which cost at least $36.21 and took at least 22.5 minutes. The investigation revealed that the agent previously resent growing histories of page text and screenshots at full price on every request. By refining history management, caching policies, and screenshot limits, the team achieved substantial savings, with 89 percent of inputs coming from cache at 5 percent of the uncached price. Asana has released the changes to browser navigation in StackAI and plans to incorporate testing into platform evaluations while expanding the use of AI agents for pre-release software testing.