Asana has reduced the cost of its browser agent by 76 times and improved speed by 5 times using GPT-6.1 Sol, according to a 144-run study. The optimization allows the company to offer more capable models to customers at lower operating costs.
TL;DR
- Asana's browser agent now runs 76x cheaper and 5x faster using GPT-6.1 Sol.
- The optimization enables Asana to provide more advanced models to customers while reducing operational expenses.
- Asana plans to incorporate similar experimentation tools into its platform for easier testing and evaluation.
What happened
Asana conducted a 144-run study using GPT-6.1 Sol and other frontier models to optimize its browser agent's workflow. The optimized workflow averaged $0.47 in estimated model costs and about four minutes per run, a significant improvement from the original setup.
The study identified inefficiencies in the browser agent's caching and history management, leading to the development of an optimized workflow that extended caching to the agent's browsing history and improved text retention.
Asana's StackAI CTO, Frank Hidalgo, PhD, directed GPT-6 Astra in Codex to investigate the agent, test improvements, and compare results. The process, which would have taken one to two months manually, was completed in about a week.
Why it matters
The optimization allows Asana to offer more capable models to customers at lower costs, making advanced automation more accessible. This can drive greater adoption and usage of Asana's StackAI platform.
For developers and startups, this demonstrates the potential of using advanced AI models to optimize workflows and reduce costs. It highlights the importance of efficient history management and caching in AI-driven automation.
Investors may see this as a positive sign of Asana's innovation and ability to leverage advanced AI models to enhance its product offerings and reduce operational expenses.
Key facts
- Asana's optimized browser agent workflow costs $0.47 per run, a 76x reduction from the original setup.
- The optimized workflow is 5x faster, with an average runtime of about four minutes per run.
- The study involved testing GPT-6.1 Sol and three other frontier models, with the optimized workflow emerging on GPT-6.1 Sol.
- GPT-6 Astra in Codex completed the optimization process in about a week, a task estimated to take one to two months manually.
- The optimized workflow improved the number of successful runs from three of 18 to all 18 with the larger history budget.
- Asana plans to incorporate similar experimentation tools into its platform for easier testing and evaluation.
- The optimization enables Asana to provide more advanced models to customers while reducing operational expenses.
Context
Asana's StackAI platform allows customers to build workflows that navigate websites, fill out forms, and gather information without writing code. At Asana's scale, small inefficiencies in these workflows can add up significantly.
The use of GPT-6 Astra in Codex to optimize the browser agent demonstrates the potential of advanced AI models to improve efficiency and reduce costs in automation tasks.
This optimization is part of a broader trend in the AI industry to leverage advanced models for improving product offerings and reducing operational expenses.
