A browser agent, optimized by another agent
StackAI’s task was to collect six fields for each of 32 books from a public catalog. The original browser workflow could spend at least $36.21 in estimated model costs and take 22.5 minutes or longer. After changes to history caching, retained page text and screenshot cleanup, the GPT‑6.1 Sol version averaged $0.47 and around four minutes.
OpenAI says GPT‑6 Astra in Codex inspected the code, proposed changes, organized the test harness and ran a 144-run comparison across four models and six caching/screenshot policies. Engineers selected the fixes and reviewed the results.
The important detail is repeatability
The optimized workflow returned the correct answer in all 18 reported Sol runs, compared with three of 18 in the smaller history-budget configuration. The benchmark task was a browser data-collection workflow, and each configuration had three repetitions; the result should be read as a specific engineering study, not a universal agent cost guarantee.
The practical lesson is that agent performance depends on memory and workflow design as much as model choice. Keeping useful browsing history stable made caching more effective and reduced the need to revisit pages. StackAI says the resulting browser-navigation changes have shipped, and it is developing ways to repeat this kind of cost/runtime/quality evaluation.
Source published 2026-10-09. Coverage is based on the maker’s announcement and demonstration.