Asana says cache changes cut a browser agent's cost 76x

Asana says it cut the estimated model cost of one browser-agent workflow from at least $36.21 to $0.47 per run, while reducing runtime from at least 22.5 minutes to about four minutes. The change was not a new model. It kept the agent's growing page history stable long enough for prompt caching to work, then pruned screenshots in batches instead of rewriting the history on every step. The result came from a 144-run study across GPT-6.1 Sol and three other frontier models, plus a 12-run follow-up. Asana says every best-condition run reached the same 192 facts, but also notes that the study used only three or four runs per condition and that its baseline included capped runs. The practical lesson is to measure cache reads, history churn, cost, runtime, and answer quality together. A cheaper agent that fails more tasks is not cheaper. Sources: OpenAI: https://openai.com/index/asana-browser-agent/ Asana: https://asana.com/inside-asana/cut-browsers-agent-cost