An October 9 OpenAI case study describes Asana's work to reduce model costs and runtime for a StackAI browser agent. Its 144-run study used one task: collecting six fields for each of 32 books from a public demonstration catalog. That scope matters when interpreting the headline.
OpenAI says the optimized workflow on GPT-6.1 Sol averaged $0.47 in estimated model costs and roughly four minutes per run. It describes this as 76 times cheaper and five times faster than the original production setup using an unnamed Model B. These are test results for that workflow, not a published retail price or savings for every browser task.
The study varied history budgets, caching and screenshot retention across four models, with three runs per configuration. The optimized policy retained more history and allowed screenshots to accumulate before removing older ones. The article says the changes were reviewed and moved into production.
The comparison includes both workflow changes and a model change. For Model B alone, optimization lowered estimated cost from at least $36.21 to $1.24 per run. On GPT-6.1 Sol alone, the article says the new policy lowered costs from $1.97 to $0.47. The 76-times headline should not be described as the isolated effect of switching models.
The original notes that some baseline runs hit a step limit before finishing, so baseline means are lower bounds. It says every optimized-workflow run completed the task with the correct answer. Correct completion on this catalog task does not establish reliability on payments, private accounts or unrelated websites.
Limits: this is a vendor-published customer study, not an independent broad benchmark. Estimated model costs are not the full cost of running a service, and three-run means do not guarantee a fixed runtime or bill. Keep the task, comparison workflow and capped baseline together when quoting the result.