Asana cuts browser agent costs 76x and boosts speed 5x using GPT-6.1 Sol

Asana reduced the estimated model cost of its browser agent by 76x and increased speed by 5x by optimizing its workflow on GPT-6.1 Sol. The company used GPT-6 Astra in Codex to identify inefficiencies in StackAI, a platform acquired to help customers automate work across business applications. The optimized workflow now averages $0.47 per run and takes about four minutes, compared to at least $36.21 and 22.5 minutes on the original production setup using Model B.

Identifying Inefficiencies with GPT-6 Astra

Frank Hidalgo, PhD, Asana’s StackAI CTO, directed GPT-6 Astra to investigate the browser agent’s codebase. The agent discovered that while fixed instructions and tool definitions were cached, the growing history of page text and screenshots was resent at full price with every request. Additionally, the agent dropped older screenshots and trimmed text at nearly every step, which prevented effective caching and forced the agent to revisit pages it had already read.

Optimizing Caching and History Budgets

Hidalgo selected three fixes for testing: extending caching to browsing history, increasing retained text, and removing screenshots in batches rather than at every step. GPT-6 Astra refactored the code to support parallel workflows and conducted a full study across four models, including GPT-6.1 Sol and three other frontier models (Models A, B, and C). The best-performing policy allowed screenshots to accumulate to 20 before cutting back to the most recent one, combined with a larger history budget of 480,000 characters.

Cost and Performance Results

The optimization reduced estimated model costs for Model B from at least $36.21 to $1.24 per run, a 29x reduction. The optimized workflow on GPT-6.1 Sol was 2.6x cheaper still, at $0.47 per run. On GPT-6.1 Sol alone, the new caching policy reduced costs 4x from $1.97 to $0.47, as 89% of the input came from cache at 5% of the uncached price. Every run in the optimized workflow completed the task and returned the correct answer.

Impact on Agent Reliability and Development

Increasing the history budget for GPT-6.1 Sol improved reliability significantly. With a smaller budget, only 3 of 18 runs produced an answer; with the larger budget, all 18 runs produced the correct answer. Asana has released these changes to browser navigation in StackAI and is using GPT-6 Astra to test product features before release, aiming to shift the bottleneck from shipping speed to human attention.