Asana says tests of an optimized GPT-6.1 Sol browser-agent workflow averaged $0.47 and about four minutes per run.
Asana says it has reduced the estimated model cost and runtime of a browser-agent workflow by changing how the agent handles browsing history, screenshots, and cached inputs. OpenAI published the results on October 9, 2026. In the reported tests, the optimized workflow using GPT-6.1 Sol averaged $0.47 and about four minutes per run—76 times cheaper and five times faster than the original production setup on another model, which the study labels Model B.
The work began with StackAI CTO Frank Hidalgo using GPT-6 Astra in Codex to investigate the agent and run experiments. Asana’s StackAI platform lets customers build workflows that navigate websites, fill out forms, and collect information. Hidalgo estimated that the investigation would have taken one to two months manually; the team completed it in about a week.
The study tested a task that collected six fields for each of 32 books from a public demo catalog. The team compared four models and tested different history budgets, caching approaches, and screenshot-removal policies. The selected configuration gave the agent more room to retain browsing history, cached that history, and removed screenshots in batches rather than at every step.
The changes addressed a cost problem in the earlier design: each request resent the accumulated page text and screenshots, while frequent edits to that history made caching less effective. On Model B, the revised workflow cut estimated cost from at least $36.21 to $1.24 per run. The optimized GPT-6.1 Sol workflow averaged $0.47. On Sol alone, the changes reduced cost from $1.97 to $0.47; the company says 89% of input tokens in the optimized runs came from cache.
The larger history budget also improved completion in the tests: GPT-6.1 Sol returned correct answers in all 18 runs, compared with three of 18 under the smaller budget. Asana says it has released the browser-navigation changes in StackAI and is developing ways to repeat these evaluations when configuring agents.