Asana optimizes browser agent workflows to run 76x cheaper using GPT-6.1 Sol

Article Summary

Asana optimized its browser agent workflow on GPT-6.1 Sol, achieving a 76x reduction in costs and 5x faster execution compared to its original production setup on Model B. Through a 144-run study directed by StackAI CTO Frank Hidalgo using GPT-6 Astra, the team identified inefficiencies in history caching and screenshot management, reducing research time from an estimated two months to one week. The final optimized configuration, which features extended history caching and batched screenshot removal, resulted in an average cost of $0.47 per run with a duration of approximately four minutes.

AI Analysis

## Core Assessment
Asana’s shift to GPT-6.1 Sol for its browser agent isn't a miraculous leap in model intelligence; it is a masterclass in aggressive state management. By moving from a "forgetful" agent that re-processed redundant screenshots and history to a batched, cache-heavy architecture, the team slashed costs by 76x and latency by 5x. The primary bottleneck was never the model's reasoning capacity but the inefficient "re-feeding" of history. By caching context and batching screenshot disposal—retaining 20 images before purging—they hit a sweet spot where 89% of input volume could be served via cache at 5% of standard costs. At $0.47 per run, this is a legitimate production win, but it highlights how much "intelligence" we currently burn on re-reading our own past mistakes.

## Impact and Advice
* **Engineering Leads:** Audit your agent's "context window hygiene." If your agents are re-ingesting full page history or screenshots at every turn, stop optimizing the prompt and start optimizing the data pipeline.
* **Product Managers:** If a workflow feels sluggish or expensive, track the "token churn"—how often the agent re-processes the same data. Reducing unnecessary re-reads is often cheaper and more effective than switching to a smaller model.

## Bottom Line
True optimization here came from refactoring the code to make experiments testable, not just from the model itself—a reminder that an automated tool is only as efficient as the architecture it’s forced to navigate.

Explore this story further