Automation Revolution: How Asana Reduced AI Costs by 76x with GPT-6.1 Sol

Operational efficiency is the "Holy Grail" for tech companies operating at scale. For Asana, a leading work management platform, small bottlenecks in automation flows can turn into astronomical costs as the user base grows. Recently, the company proved that artificial intelligence is not just for creating content, but for optimizing itself, achieving impressive results: a 76-fold reduction in modeling costs and a 5-fold increase in the speed of its browser agent.

The Challenge of Large-Scale Automation

Asana, through its StackAI platform—acquired strategically to integrate execution across systems—allows clients to create workflows that navigate websites, fill out forms, and extract data autonomously. The problem? The complexity of these flows. When dealing with thousands of automated executions per day, any inefficiency in using the Large Language Model (LLM) API multiplies rapidly.

Frank Hidalgo, CTO of StackAI at Asana, proposed an ambitious challenge: to make the platform's browser agent more agile and accessible. The task, which would traditionally take one to two months of manual engineering work, was completed in just one week thanks to collaboration with GPT-6 Astra in Codex.

The Role of GPT-6 Astra in Development

The strategy adopted by Hidalgo demonstrates an emerging paradigm: "agent-assisted development." Instead of analyzing code line by line, the CTO used GPT-6 Astra to:

1. Map the source code: The agent explained how each model request was constructed.
2. Identify inefficiencies: The system discovered that the agent was storing fixed instructions in the cache but failed to ignore the growing history of screenshots and text from visited pages. This forced the system to resend a massive volume of redundant data with every request, increasing processing costs.
3. Intelligent refactoring: Since the original structure did not allow for controlled experiments, GPT-6 Astra refactored the code so that a single backend could support parallel workflows, each with customized settings.

Results: GPT-6.1 Sol vs. The Competition

In a rigorous comparative study of 144 executions, Asana put the new GPT-6.1 Sol head-to-head with other top models on the market (referred to as Models A, B, and C). The numbers are striking:

• Financial Efficiency: The average cost per flow dropped to $0.47, a 76x reduction compared to the previous model.
• Performance: Execution speed jumped to approximately four minutes per run, 5x faster than the production standard at the time.

These gains were not purely accidental. The agent proposed three fundamental changes: extending the cache to the browsing history, increasing text retention, and implementing batch deletion of screenshots instead of removing them at every individual step.

The Future of Human-Agent Collaboration

Arnab Bose, CPO at Asana, summarized the feat with a sentence that defines the current moment in technology: “This is what human and agent teams look like in practice.” The process was not entirely automated by an AI, nor was it carried out entirely by humans. There was orchestration:

• The Engineer: Defined the strategic direction and success parameters.
• The Agent (GPT-6 Astra): Executed the research, testing, and implementation of the improvements.
• The Result: Brought to production via Command, ensuring that the innovation reached the final customer in record time.

Conclusion: What This Means for the Market

This milestone reached by Asana signals a profound shift for software companies. The ability to use AI to optimize the costs of its own infrastructure—especially in the field of browser agents—democratizes access to complex automations. If previously the financial viability of a browsing bot was restricted by AI token costs, the new era of model optimization, exemplified by GPT-6.1 Sol, promises to transform inefficient workflows into highly profitable and fast solutions.

Asana is not just selling a management tool; it is proving that the next frontier of productivity will be built by engineers who know how to command armies of digital agents.