Asana has cut model costs 76x and sped up its browser agent 5x, using GPT-6 Astra in Codex to run experiments on GPT-6.1 Sol. Asana helps customers automate work across business applications through Stack AI, a platform it acquired. Using Stack AI, customers can build workflows that navigate websites, fill out forms, and gather information without writing code. At Asana's scale, small inefficiencies in these workflows add up, so Stack AI CTO Frank Hidalgo, Ph D, set out to make the browser agent faster and cheaper to run. He directed GPT-6 Astra in Codex to investigate the agent, test improvements, and compare the results. Work he estimates would have taken one to two months by hand took about a week.
Asana's 144-run study tested GPT-6.1 Sol and three other frontier models. The optimized workflow on GPT-6.1 Sol averaged $0.47 in estimated model costs and about four minutes per run, 76x cheaper and 5x faster than the original production setup on Model B. Arnab Bose, CPO at Asana, said the work showed how an engineer could set direction while GPT-6 Astra ran the experiments and results moved through Command to production.
The investigation found that the agent cached its fixed instructions and tool definitions but not the growing history of page text and screenshots, so every request resent that history at full price. Hidalgo tested three changes: extending caching to browsing history, retaining more text, and removing screenshots in batches. The best policy let screenshots accumulate to 20 before trimming to the most recent one, keeping earlier history unchanged for longer stretches.
Each configuration collected six fields for each of 32 books from a public demo catalog. Every session's requests, data traces and results were recorded in Command, so the team could review the study afterward. Findings became tickets, then pull requests, and the changes went to production.
Source: OpenAI Blog · Summarized by HeadlinesBriefing