An agent optimized another agent's workflow
On 9 October 2026, OpenAI published a case study describing how Asana used GPT-6 Astra in Codex to optimize a browser-agent workflow running on GPT-6.1 Sol. Asana's StackAI platform lets customers build no-code workflows that navigate websites, fill forms and gather data. At scale, small inefficiencies in these workflows compound.
Asana's StackAI CTO, Frank Hidalgo, directed GPT-6 Astra in Codex to run a systematic experiment: four models, two history budgets (120k and 480k characters), six caching and screenshot policies, three runs per configuration--144 runs total. The task was collecting six fields for each of 32 books from a public demo catalog, representative of customer workloads.
The best configuration on GPT-6.1 Sol used the larger history budget and a policy that accumulated screenshots up to 20 before keeping only the most recent one. This preserved earlier history for longer stretches. With this setup, GPT-6.1 Sol averaged $0.47 estimated model cost and about four minutes per run.
The optimization was history management, not model substitution
The 76x cost reduction (from at least $36.21 on the original Model B setup to $0.47 on optimized GPT-6.1 Sol) came from workflow changes, not simply switching models. On Model B, the optimization alone reduced cost 29x (to $1.24). Moving to GPT-6.1 Sol with the optimized workflow added another 2.6x improvement.
On GPT-6.1 Sol alone, the new caching and screenshot policy cut cost 4x (from $1.97 to $0.47). The larger history budget also improved reliability: with the smaller budget, only 3 of 18 runs produced an answer; with the larger budget, all 18 succeeded, each with the correct answer.
Asana says GPT-6 Astra in Codex compressed an estimated two months of manual optimization research into one week. The company now uses GPT-6 Astra in Codex to test product features before release--moving through the platform, trying inputs, and reporting bugs for human QA reviewers.
One workload, four models, lab conditions
The study covers one task (book-catalog extraction on StackAI), four unnamed models (A, B, C, and GPT-6.1 Sol), and a controlled experimental setup. The "Model B" baseline is not identified. The 76x figure compares the original production setup on Model B to the optimized GPT-6.1 Sol configuration. It does not mean every browser agent will see 76x savings. The task is specific, the models are not all named, and the cost figures are OpenAI's estimated model costs, not full infrastructure costs.
Teams running browser agents should audit history and screenshot policies
The practical takeaway is structural: history budget and screenshot retention policy are high-leverage knobs for browser-agent cost and reliability. Teams using any model for browser automation should test whether accumulating screenshots (with periodic truncation) preserves context better than aggressive truncation, and whether larger history budgets reduce failure rates. The experiment design--systematic grid search over policies, directed by a capable model in Codex--is itself a reusable pattern.