Asana Cuts Browser Agent Costs 76x With GPT-6.1 Sol

Asana cut model costs 76x in browser tests with GPT-6.1 Sol by using Codex with GPT-6 Astra, while also boosting execution speed to 5 times its previous rate.

Frank Hidalgo, StackAI CTO at Asana (StackAI is a platform Asana acquired), used GPT-6 Astra in Codex to analyze the browser agent, test improvements, and compare outcomes. The research involved 144 run tests using GPT-6.1 Sol alongside three other frontier models, referred to as Models A, B and C.

The finalized GPT-6.1 Sol workflow achieved an average estimated model cost of 0.47 dollars per run, completing tasks in approximately four minutes. Compared to the baseline configuration built on model B, the optimized workflow was 76x cheaper and 5x faster.

For model B specifically, the configuration changes brought down the estimated model cost per run from a minimum of 36.21 dollars to 1.24 dollars, a 29x reduction. Focusing solely on GPT-6.1 Sol, a larger history capacity paired with updated caching and screenshot strategies decreased per-run costs from 1.97 dollars to 0.47 dollars. This cost reduction occurred because 89% of inputs were served from cache, at 5% of the uncached price.

The investigation tested history capacities set at 120,000 and 480,000 characters paired with six caching and screenshot strategies across four models, running each combination three times. The top-performing setup allowed screenshots to accumulate up to 20 before trimming back to retain only the newest capture. With the larger history budget, the number of GPT-6.1 Sol runs that produced an answer rose from 3 of 18 to all 18, each with the correct answer.