Asana has significantly reduced the operational costs of its browser automation tools by optimizing workflows on GPT-6.1 Sol, achieving a 76x cost reduction compared to its previous production setup. The company utilized GPT-6 Astra within the Codex environment to investigate and test improvements to its StackAI platform, which enables customers to build no-code workflows that navigate websites and gather information. This optimization highlights how SI agents can be used to refine the efficiency of other SI models in enterprise deployments.
What Happened
Frank Hidalgo, PhD, Chief Technology Officer of StackAI at Asana, directed GPT-6 Astra in Codex to analyze the browser agent’s codebase and identify inefficiencies. The investigation revealed that the agent was caching fixed instructions but not the growing history of page text and screenshots, causing every request to resend this history at full price. Additionally, the agent frequently dropped older screenshots and trimmed text, which altered the history and prevented effective caching.
Hidalgo selected three fixes to test: extending caching to the browsing history, increasing the retained text amount, and removing screenshots in batches rather than at every step. GPT-6 Astra conducted a 144-run study comparing GPT-6.1 Sol and three other frontier models (referred to as Models A, B, and C). The optimized workflow on GPT-6.1 Sol averaged $0.47 in estimated model costs and about four minutes per run. This represents a 76x cost reduction and a 5x speed increase compared to the original production setup on Model B, which cost at least $36.21 per run.
The study found that the best-performing policy allowed screenshots to accumulate to 20 before cutting back to the most recent one, combined with a larger history budget of 480,000 characters. This configuration ensured that 89% of the input came from cache at 5% of the uncached price. Every run in the optimized workflow completed the task and returned the correct answer, whereas the smaller history budget on GPT-6.1 Sol resulted in only three of 18 runs producing an answer.
Why It Matters
The efficiency gains allow Asana to offer faster, more capable SI models to customers while maintaining sustainable operating costs. Previously, cost constraints limited which models could be deployed for these workloads. By reducing the cost per run from over $36 to under $0.50, the company can enhance the user experience without prohibitive financial overhead.
The process also highlights the evolving role of SI agents in software development. Hidalgo noted that what would have taken one to two months by hand was completed in about a week using GPT-6 Astra. "This is what teams of humans and agents look like in practice. An engineer set the direction, GPT-6 Astra ran the experiments, and the results went through Command to production," said Arnab Bose, Chief Product Officer at Asana. Asana is now using GPT-6 Astra to test product features before release, navigating the platform and reporting bugs for human review.
The Bottom Line
Asana has released the optimized browser navigation changes to StackAI and plans to integrate similar experimentation tools into the platform’s evaluations. The company views this approach as a foundation for a new software development lifecycle, where cloud agent sessions test features in parallel, shifting the bottleneck from shipping speed to human attention.
"Shipping speed is no longer the bottleneck; human attention is. We’re close to a world where every engineer is a PM leading a fleet of agents," said Hidalgo.