Asana’s browser-agent study offers a useful distinction for a studio considering automation: making a run cheaper and producing the required deliverable are separate tests. The company investigated how its agent retained page text and screenshots. Its results concern a development task, rather than the cost of delivering a creative campaign.
The report, dated September 30, describes 144 runs across four models. The task was to gather six fields for each of 32 books from a public demonstration catalog. Asana varied history retention and caching, with three runs per condition. This is a documented company experiment; The Human Input has not reproduced it.
What the 76x comparison includes
Asana reports that its optimized workflow on GPT-6.1 Sol cost 76 times less than the development baseline on another model, identified as Model B. That comparison combines changes to the agent with a model switch. Optimizing Model B alone reduced estimated model cost 29-fold. The report’s comparison within Sol, holding the larger history budget fixed, found a fourfold reduction.
The engineering change was to preserve more of the browsing record between calls so previously sent input could be reused from cache. Removing screenshots at every step had repeatedly changed that record. Asana instead tested keeping them until a larger batch could be removed. Those settings describe its experiment, rather than a universal configuration for a studio’s agent.
Check the answer as well as the bill
The full report draws an important boundary around the result. Sol returned the correct price total when it answered, but gave a short summary rather than the requested full table. Reading the reference facts and returning the expected total did not establish that the complete requested document had been delivered.
The costs are estimates from model usage, excluding server and sandbox costs. The report also notes that its best settings were selected after the runs and that shared computing resources limit timing comparisons. A company-reported result on one catalog task cannot establish an equivalent saving on an image archive, a supplier portal or client work.
Keep the production acceptance test separate
For a hypothetical studio task that collects asset metadata, define the required records, fields and final format before evaluating a run. Inspect the returned file against that reference. A correct summary can still leave the producer to reconstruct the delivery sheet.
Record failures and review time alongside model usage and infrastructure costs. Retain only material the team is permitted to store, and keep limits on a run that drifts or loops. Asana’s study provides a reason to examine how an agent retains its evidence; it does not replace a test of the studio’s own deliverable.
Prepared with AI assistance. The Human Input retains editorial responsibility.