Browser-based AI agents can repeatedly send a growing history of page text and screenshots to the model as they work.
Asana, a work-management software company, has found that changing how one of its agents handled the history cut estimated model costs 29-fold without changing models.
The agent is part of StackAI, Asana’s platform for building no-code workflows that navigate websites and collect information. The company’s 144-run study tested four AI models, two limits on retained history and six caching policies. Each configuration performed the same task: collecting six details about each of 32 books from a public demonstration catalog.
It claims that the problem was how the agent handled information it had collected.
The agent cached its fixed instructions but repeatedly sent its browsing history back to the model at full price.
It also removed older screenshots and trimmed text at nearly every step. Those edits disrupted caching because the model provider could reuse only the unchanged beginning of each request.
Asana tested three changes: caching the browsing history, increasing the amount of text the agent could retain from 120,000 to 480,000 characters and removing screenshots in batches. On the model originally used in production, the best configuration reduced estimated model charges from at least $36.21 to $1.24 per run. Average runtime fell from at least 22.5 minutes to 4.7 minutes.
A later follow-up tested keeping every screenshot rather than removing them in batches. On the same catalog task, that approach cost less per model call than the best batch-pruning settings on three models. Asana said pruning could still be necessary for longer workflows or smaller model context limits.
Another model lowered test costs
Asana’s study recorded still lower costs when the optimized browser agent was tested on GPT-6.1 Sol. OpenAI reported an average estimated cost of $0.47 per run, 76-fold below the original setup. Runs took about four minutes, with 89% of the input served from cache.
At $0.47, Sol’s estimated model cost was about 2.6 times lower than the optimized result on the original model. Both models had the same listed standard input and output prices, but Sol’s cache-read price was lower. Asana cautioned that differences in testing protocols limited direct comparisons between models.
The larger history allowance also affected whether the agent returned an answer. With the 120,000-character limit, GPT-6.1 Sol did so in three of 18 runs. With the 480,000-character limit, it returned the correct price total in all 18. However, the agent returned a summary rather than the complete table requested, so it did not fully satisfy the task’s instructions.
Production savings remain unmeasured
The reported costs were estimates calculated from model token usage and listed prices, excluding server and browser-sandbox expenses. The study covered one browser agent and one catalog task, with three repetitions per condition. Some baseline runs also reached their step limit before finishing, making their reported costs lower bounds rather than complete-run measurements.
Asana said it had shipped the browser-navigation changes to StackAI and was developing tools to compare agent costs, runtime and answer quality more systematically. Its published results did not show whether comparable savings carried over to other production workloads.