Asana found that caching browser history and expanding the agent’s context window slashed estimated model charges from at least $36.21 to $1.24 per run, before changing the underlying model.
Asana has reported that a relatively small change in the way its browser agent handled page history cut estimated model costs by 29-fold before any model switch was considered, underlining how much of the expense in browser-based AI work can come from orchestration rather than model choice alone. The work, which Asana said was carried out on its StackAI no-code workflow platform, examined how an agent gathered six facts about each of 32 books from a public demonstration catalogue and compared four models, two history limits and six caching policies across 144 runs.
The company said the main issue was not the model’s reasoning ability, but the way the agent continually resent its browsing history. Fixed instructions were cached, yet the accumulated page text and screenshots were repeatedly sent back at full cost. As the agent advanced, it also kept trimming older screenshots and shortening text, which reduced the chance that the model provider could reuse the unchanged opening portion of each request. In Asana’s testing, changing that behaviour so the browsing history was cached, extending the retained text window from 120,000 to 480,000 characters and batching screenshot removal reduced estimated charges on the production model from at least $36.21 to $1.24 per run. Average runtime also fell from at least 22.5 minutes to 4.7 minutes.
A later check suggested that screenshot pruning was not always necessary. When Asana tested keeping every screenshot instead of removing them in batches, that approach was cheaper per model call than the best pruning setup on three of the models it examined. The company said pruning may still be useful for longer workflows or for models with smaller context limits, where unbounded history can become impractical.
The strongest results came when the optimised browser agent was run on GPT-6.1 Sol. According to OpenAI’s account of the work, Asana recorded an average estimated cost of $0.47 per run, which was 76 times lower than the original setup, with runs taking about four minutes and 89% of input served from cache. Asana said the larger history allowance also improved reliability: with the smaller 120,000-character limit, Sol returned an answer in only three of 18 runs, while the larger limit produced the correct total in all 18. Even so, the agent returned a summary rather than the full table requested, so it still fell short of the task specification.
Asana said the figures were estimates based on token usage and published model prices, and did not include server or browser-sandbox costs. The study covered one browser agent and one catalogue task, with three repetitions per condition, so the company cautioned that the results should not be treated as a universal benchmark for production workloads. Some baseline runs also hit their step limits before finishing, meaning those costs were lower bounds rather than complete-run totals. Asana said it has already shipped the browser-navigation changes to StackAI and is building tools to compare agent cost, runtime and answer quality more systematically.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
