Asana's StackAI team published a 144-run browser-agent study on October 8 reporting a 76× cost reduction on a catalog-scraping task, moving from at least $36.21 per run on its production model to $0.47 per run on OpenAI's GPT-6.1 Sol. The headline is the model swap. The actual finding is quieter: 29× of the gain came from a prompt-cache fix that has nothing to do with which model you run, and only 2.6× came from switching to Sol.

The task itself is unglamorous. An agent visits a bookstore site, collects six fields across 32 books, and returns 192 facts. StackAI ran four models (Sol plus three unnamed as A, B, and C), six caching and screenshot policies, two history budgets, and three runs per condition. A 12-run follow-up on October 1 stress-tested the result.

Here's what the baseline agent was doing wrong. On the original production model, call it Model B, it burned through at least 271 page revisits per run and took upwards of 22.5 minutes to finish. The request history was being edited in place on every step. That meant the model's cache, which charges cache reads at roughly 0.05× to 0.1× of the standard input rate on the models tested, almost never hit.

Making the history append-only dropped Model B's cost from $36.21 to $1.24. That's the 29× fix. Everything after is incremental.

Screenshot handling matters almost as much. StackAI let the agent accumulate 20 screenshots and then pruned back to 1, which left roughly 19 consecutive calls reusing the same unchanged prefix. Raising the history budget from 120,000 to 480,000 characters cut page revisits from 271 to just 2 on Model B. Models that trimmed history mid-run broke the cache again: Model C first trimmed at call 10, Model A at call 64. At the 120,000-character budget, GPT-6.1 Sol produced successful answers on 3 of 18 runs; Model C managed 0 of 18.

With the pipeline fixed, GPT-6.1 Sol served 89% of input from cache at its $0.10-per-million cached rate against a $2.00 uncached rate and $10.00 output rate, cache reads at 5% of uncached. Runtime fell to about 4 minutes. Model C hit 4.7 minutes at 1.3× the cost. A no-pruning follow-up (capped at 100 screenshots, never reached) ran roughly 17% cheaper per call on Model B and Sol, 5% on Model C.

Frank Hidalgo, StackAI's CTO, said the manual version of this study would've taken one to two months. Codex and Astra ran it in about a week.

The structural read is that browser-agent economics in late 2026 are dominated by cache discipline, not model choice. The $35 you thought you were paying GPT was mostly a bill for re-sending yesterday's conversation.

Sources

https://openai.com/index/asana-browser-agent/
https://www.stackai.com/blog/how-stackai-by-asana-used-astra-codex-and-command-to-reduce-browser-agent-costs
https://assets.asana.biz/asset/fb86cc8e-6611-4404-b7c6-172692a724bb/How-Asana-used-Codex-to-optimize-browser-agent-costs-and-runtime.pdf
https://www.orcarouter.ai/blog/gpt-6-1-sol-browser-agent-76x
https://www.globaltechnewsdesk.com/2026/10/asana-reports-76-lower-estimated-gpt-61.html