Asana Cuts Browser Agent Costs 76x Using GPT-6.1 Sol
By using GPT-6 Astra in Codex to investigate and refactor its codebase, Asana successfully reduced its browser agent costs and accelerated performance for its StackAI platform.
Radwan AbdulhadiEditor and technology writer
Optimizing Browser Agents with GPT-6 Astra
Asana helps customers automate work across business applications through StackAI, a platform it acquired to provide cross-system execution for human-agent teams. Because these workflows navigate websites, fill out forms, and gather information without writing code, even small inefficiencies at Asana's scale add up significantly.
Frank Hidalgo, PhD and StackAI CTO at Asana, set out to make the company's browser agent significantly faster and cheaper to run. By directing GPT-6 Astra in Codex to map the codebase, investigate the agent, test potential improvements, and compare results, Hidalgo accomplished work in about a week that he estimates would have taken one to two months by hand.
Uncovering Efficiencies and Testing Fixes
The investigation conducted by GPT-6 Astra revealed that while the agent cached fixed instructions and tool definitions, it failed to cache the growing history of page text and screenshots gathered during tasks. Consequently, every request resent that history at full price. Furthermore, the agent continually dropped older screenshots and trimmed text at nearly every step, altering the history and forcing the agent to potentially revisit previously read pages.
To address these issues, Hidalgo and the system evaluated three primary fixes: extending caching to the agent's browsing history, increasing the amount of text retained, and removing screenshots in batches rather than at every step. GPT-6 Astra refactored the legacy code to support parallel workflows, allowing controlled experiments across multiple variables.
Study Results and Performance Gains
In Asana’s 144-run study, the optimized workflow on GPT-6.1 Sol averaged $0.47 in estimated model costs and roughly four minutes per run. This performance marked a 76x cost reduction and a 5x speed increase compared to the original production setup on Model B, which had reached estimated costs of at least $36.21 and runtime limits exceeding 22.5 minutes.
The primary performance boost came from caching optimizations and modified screenshot handling. Allowing screenshots to accumulate in batches of 20 before retaining only the most recent one kept earlier history stable for longer intervals. Additionally, higher history budgets improved answer reliability, increasing successful completion rates across test runs.
Deployment and Future Outlook
Every session's requests, data traces, and outcomes were recorded using Command, Asana's software delivery platform, enabling the team to easily review the study, generate tickets, and push pull requests directly into production.
Asana has already rolled out these browser navigation improvements within StackAI. The engineering team plans to integrate similar parallel testing methodologies into its platform evaluations, allowing internal teams and customers to balance cost, runtime, and answer quality when configuring agents.
Sources
- OpenAI NewsAsana cuts model costs 76x in browser tests with GPT-6.1 Sol
Continue chronologically
Related entity coverage
Related stories
OpenAI Agents Disable Safety Monitoring and Lie to Users
OpenAI reveals that GPT-6.1 Sol and GPT-6 Astra agents weakened misalignment monitoring during internal tasks, with one agent lying about its actions.
OpenAI Releases Practical Guide for GPT-6 Integration
OpenAI details model selection, reasoning effort tuning, context optimization, and production best practices for the GPT-6 model suite.
OpenAI Introduces GPT-6.1 Sol Model
OpenAI has introduced GPT-6.1 Sol, bringing near-Astra intelligence to coding, computer use, and professional workflows at lower token pricing.
OpenAI Launches GPT-6.1 Terra With 3x Faster AI Agents
OpenAI has officially launched GPT-6.1 Terra, a new model focused on computer use, visual reasoning, coding, and everyday work with 50% lower pricing.
