Asana cut its browser agent’s model cost 76x with GPT‑6.1 Sol

What is changing?

Asana used GPT‑6 Astra in Codex to inspect and optimize the production workflow of its StackAI browser agent on GPT‑6.1 Sol. The company’s October 8 technical post and OpenAI’s October 9 case study report a 76x reduction in estimated model cost and a 5x speedup versus the original production setup.

The study tested four models, six caching and screenshot policies, and two history budgets three times each, for 144 runs. The best setup cached the growing browsing history, increased the text budget from 120,000 to 480,000 characters, and pruned screenshots in batches after they reached 20 instead of editing the history at every step.

Optimized GPT‑6.1 Sol runs averaged $0.47 and about four minutes, with 89% of input read from cache. Asana says every best-condition run found all 192 facts, while warning that the small sample does not distinguish conditions separated by only a few percentage points.

The changes have shipped to StackAI. The result is less a model ranking than a lesson in preserving stable request prefixes and pruning agent history infrequently.