StackAI, part of Asana, used GPT-6 Astra in Codex to inspect its browser agent and ran 144 controlled trials across four models, two history budgets and six caching and screenshot policies. The best policy expanded history to 480,000 characters, cached stable history and let screenshots accumulate to 20 before trimming them in a batch. Optimized GPT-6.1 Sol averaged $0.47 and roughly four minutes per run, 76 times cheaper and five times faster than the original Model B production setup; all 18 runs returned the correct answer. The result comes from a fixed 32-book collection task and Asana's own cost estimates, so it does not guarantee the same gain for other agents, sites or pricing structures.