Asana slashes AI costs 76x using GPT-6.1 Sol and codex

Overview

Asana has announced a significant breakthrough in operational efficiency, revealing that its browser-based automation agents can now run at 76 times lower costs using OpenAI’s latest GPT-6.1 Sol model. The company also reported a fivefold increase in speed during comparative browser tests, marking a pivotal moment in the practical application of large language models for enterprise workflow automation.

The results were achieved through StackAI, the platform Asana acquired to help customers automate tasks across business applications without writing code. By leveraging GPT-6 Astra within Codex, Asana optimized its internal workflows to handle complex, multi-step browser interactions more efficiently than ever before.

“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.” — Arnab Bose, Chief Product Officer at Asana

The Scale of the Optimization

At Asana’s enterprise scale, minor inefficiencies in automated workflows accumulate into substantial costs. The company’s StackAI platform enables users to build workflows that navigate websites, fill out forms, and extract information autonomously. However, as the volume of these tasks grew, so did the computational expense.

To address this, Asana’s leadership initiated a comprehensive review of their browser agent’s performance. The goal was straightforward yet ambitious: reduce the cost per run while simultaneously improving execution speed. The outcome exceeded initial expectations, with the optimized workflow on GPT-6.1 Sol dropping the estimated model cost to just $0.47 per run.

This figure represents a dramatic decline from the original production setup, which relied on Model B. While the optimized workflow on Model B alone reduced costs by 29 times (from over $36 to $1.24 per run), switching to GPT-6.1 Sol provided an additional 2.6 times reduction. Consequently, the final cost landed at approximately 76 times cheaper than the baseline.

How GPT-6 Astra Transformed the Workflow

The catalyst for this improvement was not merely a change in model but a fundamental rethinking of how the agent manages data. Frank Hidalgo, PhD, CTO of StackAI, directed GPT-6 Astra in Codex to investigate the underlying architecture of the browser agent. This AI-driven audit revealed a critical inefficiency: the agent cached its fixed instructions and tool definitions but failed to cache the growing history of page text and screenshots it collected during browsing.

Every time the agent made a new request, it resent the full history at full price. Additionally, the agent aggressively trimmed older screenshots and text at nearly every step, causing it to lose context and potentially revisit pages it had already analyzed.

Hidalgo reviewed Astra’s proposed fixes and selected three key strategies:

• Extending caching to include the agent’s entire browsing history.
• Increasing the text retention limit to preserve more context.
• Removing screenshots in batches rather than at every single step.

GPT-6 Astra then refactored the codebase to support parallel workflows, allowing for controlled experiments with different variables. It tested history budgets of 120,000 and 480,000 characters across six different caching and screenshot policies. The study involved 144 runs in total, comparing GPT-6.1 Sol against three other frontier models.

Results and Production Deployment

The optimized policy allowed screenshots to accumulate up to 20 before trimming back to the most recent ones. Combined with a larger history budget of 480,000 characters, this approach kept earlier history unchanged for longer periods. On GPT-6.1 Sol alone, these changes reduced costs by four times—from $1.97 to $0.47 per run.

Performance metrics showed that 89% of input data was served from cache at just 5% of the uncached price, making each API call roughly three times cheaper. Speed also improved dramatically, with runs completing in about four minutes compared to the original 22.5 minutes or more on Model B. Every optimized run successfully completed the task and returned the correct answer, eliminating the previous issue of capped or unfinished executions.

The entire study, including requests, data traces, and results, was recorded in Command, Asana’s software delivery platform. This allowed the team to review the work thoroughly before turning findings into tickets and pull requests for production deployment.

“This would have taken me one to two months by hand. With GPT-6 Astra in Codex, it took about a week: I’d set a /goal before going to bed and review the results in the morning.” — Frank Hidalgo, PhD, StackAI CTO at Asana

Implications for Enterprise AI

Asana’s success story highlights the evolving dynamic between human engineers and AI agents. By delegating the experimental design and execution to GPT-6 Astra, the human team could focus on strategic direction rather than manual coding and testing. This hybrid approach significantly compressed the timeline from concept to production.

The comparison across models also sheds light on the broader landscape of frontier AI. While Model A and Model C were evaluated, GPT-6.1 Sol emerged as the superior choice for this specific workflow due to its balance of cost and performance. The ability to drop costs below $0.50 per run makes advanced browser automation viable for a much wider range of customer use cases.

As enterprises continue to integrate AI into their operational fabric, Asana’s optimization serves as a case study for leveraging AI to improve AI systems. It demonstrates that even small efficiencies in agent workflows can yield massive savings at scale, paving the way for more capable and accessible automation tools.