Asana cut its browser agent's costs 76x in tests with GPT-6.1 Sol

Asana's StackAI team used GPT-6 Astra in Codex to study why its no-code browser agent was so expensive. The main problem: the agent cached its fixed instructions but not the growing pile of page text and screenshots, and it trimmed that history at almost every step, so each request was resent at full price. Astra ran a 144-run study across GPT-6.1 Sol and three other frontier models, testing two history sizes and six caching and screenshot policies on one task (collecting six fields for 32 books from a demo catalogue). The winning setup cached the browsing history, kept more text, and let screenshots pile up to 20 before cutting back to the latest one. On GPT-6.1 Sol it averaged $0.47 and about four minutes per run, against at least $36.21 and 22.5 minutes for the original production setup on another lab's model. On Sol alone, the new policy cut cost from $1.97 to $0.47, mainly because 89% of input came from cache. The engineer says a one to two month job took about a week. This is OpenAI's customer story about a single task, not an independent test.