Asana says it cut the estimated model cost of a browser-agent workflow to $0.47 per run, 76 times below its original setup, while reducing execution time by a factor of five. The company reached those figures in a 144-run study rather than across all customer workloads.
The agent collected six fields for each of 32 books from a public demonstration catalog. OpenAI’s GPT-6 Astra, working through Codex, helped Asana inspect and refactor the system before comparing four models and multiple context policies. The team found that the original agent repeatedly changed its history by removing old screenshots and trimming text, preventing effective prompt caching.
The best configuration raised the text-history budget and allowed screenshots to accumulate before deleting them in a batch. That kept more of the request prefix stable, making cached input reusable. Runs on GPT-6.1 Sol then averaged about four minutes.
Asana estimates that investigating and implementing the changes took roughly one week instead of one or two months by hand. Those gains are specific to this browsing task and pricing setup, but the underlying lesson is broader: context management can matter as much as model choice when an agent repeatedly sends a growing history.