On Oct. 8, Asana’s StackAI team reported that a browser-agent redesign cut GPT-6.1 Sol’s estimated model cost on a catalog task from $1.97 to $0.47 per run. The agent kept its growing page history stable long enough to reuse cached input, with screenshots removed in batches and a larger history budget.
A browser agent resends its tools, instructions and the page text and screenshots it has collected each time it asks a model what to do next. Prompt caching reuses a request’s unchanged opening sequence, or prefix. Asana’s system was dropping the previous screenshot and trimming old text at nearly every step, changing that sequence and preventing cache reuse beyond the fixed instructions and tools.
The team added the growing history to the cache and kept screenshots until it had as many as 20, then removed older ones in a batch. That left the request unchanged for most steps. It also raised the history budget from 120,000 to 480,000 characters, reducing text trimming. Asana tested six history and caching policies at both budget sizes on four models, with three runs per setting.
The task was to visit a public book catalog, collect six details for each of 32 books and return a table with the prices and totals. In its best conditions, each model saw all 192 facts and returned the correct answer. In the best Sol condition, 89% of input tokens were read from cache, at 5% of the standard input rate. Average estimated cost was $0.47 per run, which took four minutes.
The report’s 76-fold figure compares the optimized GPT-6.1 Sol workflow with the original setup on another company’s Model B. On Model B itself, the best tested policy brought estimated cost down from at least $36.21 to $1.24 per run; some baseline runs hit the step limit, making $36.21 a lower bound. For Sol, the like-for-like comparison holds the larger 480,000-character history budget constant and cuts cost from $1.97 to $0.47. The report calls the cross-model comparison descriptive because the two phases used different request settings.
The main study covered one browser agent and one catalog task, with three runs per condition. A 12-run follow-up kept every screenshot; on this task, the authors said pruning was unnecessary. It lowered cost per call by about 17% on Model B and GPT-6.1 Sol, and about 5% on Model C, but does not establish lower total costs for longer jobs. Asana says the changes have shipped in StackAI; the study did not measure savings on customers’ production workloads.
Sources
- Asana, “How Asana used Codex to optimize browser agent costs and runtime” (technical report, Sept. 30, 2026), especially §§0, 2 and 4.
- StackAI by Asana’s Oct. 8 study summary.
- Illustrative photo of a 2020 MacBook Air M1 showing Wikipedia, not Asana’s test interface: Padgriffin, Wikimedia Commons, licensed under CC BY-SA 4.0.
