Asana Cuts Its Agent Cost on GPT-6.1 Sol by 76%

Asana announces significant savings on the execution of its browser agent, now 76 times less expensive and 5 times faster. The cost per run drops to $0.47, with durations around four minutes, and gains related to better cache utilization. These changes are integrated into StackAI, following an accelerated testing campaign via GPT-6 Astra.

Optimization Reduces Cost per Execution to $0.47

On Model B, optimization has reduced the estimated cost per execution from $36.21 to $1.24, a 29-fold decrease. By applying the optimized workflow on GPT-6.1 Sol, the cost drops to $0.47 per execution while still producing the correct response. Adjusting the historical budget alone cut costs by 4, from $1.97 to $0.47, and each call became about 3 times cheaper thanks to 89% of entries served from the cache at 5% of the out-of-cache price. On average, execution lasts about four minutes. Overall, the workflow on GPT-6.1 Sol is 76 times less expensive and 5 times faster than the initial production setup, a cost reduction also observed during browser testing. Asana has already integrated these changes into StackAI's navigation and is developing tools to facilitate the repetition of these experiments.

What Was Driving Up Costs and the Tested Fixes

After leveraging GPT-6 Astra to map the codebase and analyze the structuring of model queries, Frank Hidalgo, CTO of StackAI at Asana, took charge of the optimization. He noticed that the agent was not caching the progressive accumulation of page text or screenshots, leading to additional costs with each request. Three improvement avenues were chosen: expanding the cache to include browsing history, increasing the amount of text stored, and performing bulk deletions of screenshots instead of doing so at each step.

A Testing Campaign Focused on Budgets and Cache Policies

The experiments were conducted with history limits set at 120,000 and 480,000 characters, as well as six different strategies for cache and screenshot management; each combination was evaluated three times across four distinct models. The most effective strategy involved retaining up to 20 screenshots before keeping only the last one, which allowed for extended retention of previous history.

Astra to Accelerate Iterations and Next Steps

Thanks to GPT-6 Astra integrated with Codex, Asana was able to experiment and apply these optimizations in a week, whereas such work would have taken between one and two months if done manually. According to Frank Hidalgo, these adjustments make it possible to provide models that are both faster and more efficient while ensuring controlled operating costs, thereby removing the financial barriers that previously limited model choices for clients. The team is considering adding these experiments to platform evaluations to measure and compare costs, execution times, and relevance of responses when configuring agents. Frank Hidalgo now believes that human attention, rather than speed of delivery, is the main bottleneck, and he estimates that Asana is moving closer to a context where each engineer pilots a fleet of agents.