Asana cuts browser-agent model costs 76-fold in tests

Asana used GPT‑6 Astra in Codex to optimize a browser agent running on GPT‑6.1 Sol, reaching an average estimated model cost of $0.47 and roughly four minutes per run. Its 144-run study tested four models on the same catalog task, and Asana has released the browser-navigation changes in StackAI. [9] Why it matters: The study shows that agent economics depend on how browsing history and screenshots are retained and cached, not just which model is selected. On GPT‑6.1 Sol alone, the workflow changes reduced estimated cost from $1.97 to $0.47 per run, separating an engineering gain from the larger cross-model comparison. [9] Key insights: The original agent cached fixed instructions and tool definitions but repeatedly sent its growing page-text and screenshot history at full price. [9] | The selected policy let screenshots accumulate to 20 before retaining only the latest one, keeping earlier history stable for longer caching intervals. [9] | In the optimized GPT‑6.1 Sol workflow, 89% of input came from cache at 5% of the uncached price. [9] | A larger browsing-history budget increased answer-producing runs from three of 18 to all 18 on GPT‑6.1 Sol, with every returned answer correct. [9] Cheatsheet facts: What happened: The optimized agent averaged $0.47 in estimated model costs and about four minutes per run, reported as 76-fold cheaper and fivefold faster than the original Model B setup. [9] | Why: Caching more browsing history and changing screenshot retention reduced repeated full-price inputs. [9] | Who pays: Asana reduced estimated operating costs; workflow changes alone took GPT‑6.1 Sol from $1.97 to $0.47 per run. [9] | The catch: The 144-run study used one task: collecting six fields for each of 32 books from a public demo catalog. [9]