Asana Cuts Browser-Agent Test Costs 76-Fold, Giving Repeated Reading a Smaller Expense Account
A book-catalog task exposed costly history handling. Workflow changes and a model switch brought estimated model spend to $0.47 per run in Asana's tests.
Automation ·
Thirty-two books, six fields apiece, and a browser agent capable of making the research bill more interesting than the catalog. Asana's published experiment found its original setup spent at least $36.21 in estimated model costs per run; some runs hit the step limit before finishing. The agent was resending its growing browsing history at full price while frequent edits to that history frustrated reuse.
The team changed how much history it retained and how it handled cached content and screenshots. Combined with a switch to GPT-6.1 Sol, the optimized workflow averaged $0.47 and roughly four minutes per run. That is the reported 76-fold cost comparison. Keeping the original model and optimizing the workflow alone delivered a 29-fold reduction. I appreciate a study that separates buying a different brain from repeatedly charging for reminding the current one where it has been.
These are model-cost estimates for the tested task, not a promise to divide every customer's bill by 76. Still, the catalog provided an unusually helpful performance review. Before replacing the researcher for being slow, it was worth checking whether the replacement kept paying to reread its own notes.
Based on: How we cut a browser agent's cost 76x and made it 5x faster by keeping its cache intact, Asana, October 8, 2026; Asana cuts model costs 76x in browser tests with GPT-6.1 Sol, OpenAI, October 9, 2026.