Asana said a test of StackAI’s AI browser automation cut estimated model costs to one-seventy-sixth of those in its previous production setup and made processing five times faster. In an optimized configuration using GPT-6.1 Sol and improved browsing-history management, each run took about four minutes and cost an average of $0.47.
The 76-Fold Cost Reduction Wasn’t Just About Switching Models
The key point about these figures is that they compare not the performance of models alone, but the results of combining a model with a processing setup. A browser agent is an AI that reads web pages and carries out tasks such as gathering information and filling out forms. Through its acquisition of StackAI, Asana offers automation for this kind of work.
In the test, the team gathered six pieces of information about each of 32 books from a public demo catalog. It ran 144 tests across four models and multiple configurations. With the initially used Model B, the estimated model cost per run was at least $36.21, but simply redesigning the processing setup brought it down to $1.24. An optimized configuration using GPT-6.1 Sol lowered the cost further, to $0.47.
In other words, costs could be reduced to about one-twenty-ninth of the original level even while still using Model B. Behind the headline figure of a 76-fold reduction, improvements to the processing design before choosing a model played a major role.
The Key to Savings Was the Browsing History Fed to the AI
In the previous setup, fixed instructions and similar inputs were cached, but the growing history of page text and screenshots was not. Caching is a way to reuse inputs that have already been processed. Without being able to reuse the history, the system had to resend past information every time it called the AI, incurring a charge for it each time.
On top of that, old images were deleted and text was shortened at almost every step. Because the history was frequently rewritten, adding caching alone was unlikely to make much difference. Removing too much information could also cause the agent to revisit pages it had already read.
Asana raised the history-retention limit from 120,000 to 480,000 characters and made browsing history eligible for caching. The system now accumulates up to 20 screenshots before keeping only the most recent one. Rather than cleaning up the history each time, it does so in batches, extending the time the history can be reused without changes.
Keeping More Information Can Also Lower Costs
Giving AI less information should make it cheaper. These results show that this idea alone cannot explain the cost of an agent. Retaining and reusing information, and avoiding unnecessary checks, can sometimes make the overall process less expensive.
With GPT-6.1 Sol, optimizing caching and image management while expanding the history-retention limit brought the estimated model cost per run down from $1.97 to $0.47. After optimization, 89% of the input came from the cache, and that portion cost 5% of what it would have cost without caching.
The amount of history retained also affected task completion. With GPT-6.1 Sol, answers were generated in 3 of 18 runs when the retention limit was lower, compared with all 18 runs when it was higher. All 18 answers in the test were correct.
AI Was Also Used to Test Proposed Improvements
Separately from GPT-6.1 Sol, which operated the browser, GPT-6 Astra in Codex was used for research and improvements. It examined the code structure, identified problems with history management, reworked the code for experiments, and tested multiple configurations in parallel.
That does not mean humans were uninvolved. Frank Hidalgo, StackAI’s CTO, reviewed the proposals and selected which changes to test. Experiment records and outputs were also reviewed in separate model sessions and kept in Asana’s development platform, Command. Changes went through tickets and pull requests before being deployed to production.
Hidalgo said that research he estimated would have taken one to two months by hand was completed in about a week. This time comparison is based on his estimate, not on a controlled test in which humans performed the same work alone.
Evaluate Cost and Completion Rates Together
The $0.47 figure is an estimated model cost, not the total cost including labor and system operations. In addition, some baseline runs hit their execution limits before completing, making the figures of at least $36.21 and at least 22.5 minutes lower bounds. The main comparison figures are averages of three runs per configuration, and the same cost reduction cannot be assumed for every web task.
Among the optimized configurations, Model C cost $0.66 and took 3.5 minutes, while GPT-6.1 Sol cost $0.47 and took about four minutes. Sol was less expensive for this task, but Model C ran faster—the rankings differed for cost and speed.
Asana has incorporated these changes into StackAI’s browser automation feature. It plans to build testing into the platform’s evaluation tools so customers and internal teams can compare cost, execution time, and answer quality across configurations.
Comments