Asana optimized its automated browser agent workflow on OpenAI’s GPT-6.1 Sol model, achieving a 76x reduction in estimated model costs and a 5x increase in execution speed. The browser agent infrastructure, acquired through platform StackAI, allows enterprise customers to build automated web navigation workflows without writing code.
To identify inefficiencies, StackAI CTO Frank Hidalgo used GPT-6 Astra in Codex to map the agent’s codebase and analyze model requests. The analysis revealed that while fixed instructions were cached, growing page text and screenshot histories were re-sent at full price during every step. Astra refactored the code to run parallel controlled tests across multiple configurations.
The resulting optimal workflow extended caching to browsing history, expanded text character budgets, and altered screenshot retention policies. The streamlined workflow averaged $0.47 in estimated model costs and four minutes per execution across a 144-run benchmark study.
Why it matters
Demonstrates how dynamic context caching and media trimming can cut agentic LLM operational costs by 76x.
Shows AI coding assistants effectively refactoring legacy codebases and conducting benchmark tests autonomously.
Provides an optimization template for startups deploying cost-sensitive web-scraping and browser automation agents.
Source: openai.com



