With GPT‑6 Astra running experiments in Codex, Asana optimized its GPT‑6.1 Sol-based browser agent workflow, reducing running costs to 1/76th (76x cheaper) and speeding up execution by 5x.
Through StackAI, a platform Asana acquired, Asana helps customers automate work across business applications. With StackAI, customers can build workflows that browse the web, fill out forms, and gather information without writing any code. At Asana’s scale, even minor inefficiencies in these workflows are magnified.
Dr. Frank Hidalgo, CTO of StackAI at Asana, set out to make the browser agent faster and more cost-effective. He tasked GPT‑6 Astra in Codex with investigating the agent, testing optimizations, and comparing results. By his estimate, work that would have taken a human one to two months was completed in roughly a week.
Asana’s 144-run study tested GPT‑6.1 Sol alongside three other frontier models (referred to here as Models A, B, and C). The optimized workflow powered by GPT‑6.1 Sol achieved an estimated average model cost of just $0.47 per run in about 4 minutes—a 76x reduction in cost and a 5x increase in speed compared to the original production setup on Model B.
“This is what human-agent teams look like in practice. An engineer set the direction, GPT-6 Astra ran the experiments, and the findings were deployed to production through Command. It’s a great example of how Asana is making human-AI teams real.” — Arnab Bose, Chief Product Officer at Asana
Identifying Browser Agent Inefficiencies with GPT‑6 Astra
To move quickly, Hidalgo first had GPT‑6 Astra in Codex comb through the codebase and explain how the agent structured each model request. GPT‑6 Astra discovered that while the agent cached static instructions and tool definitions, it didn’t cache the growing history of page text and screenshots; as a result, every request was resending this history at full price.
The agent was also discarding older screenshots and trimming text at almost every step. Because each modification altered the history, simply caching the history wouldn’t solve the problem, and losing these factual details could force the agent to revisit pages it had already seen.
Compressing a Two-Month Study into One Week with GPT‑6 Astra
Hidalgo reviewed the fixes proposed by GPT‑6 Astra and selected three to test:
- Extending prompt caching to the agent’s browsing history
- Increasing the volume of text it could retain
- Pruning screenshots in batches rather than at every step
GPT‑6 Astra first ran quick tests to determine which variables mattered most. Because the original codebase wasn’t designed for controlled experiments, it then refactored the code so a single frontend and backend could support multiple workflows in parallel, each with its own independent settings.
Astra conducted the full comparative study: testing history budgets of 120,000 and 480,000 characters alongside 6 caching and screenshot pruning strategies, running each configuration 3 times across four models (see table below). The best-performing strategy allowed up to 20 screenshots to accumulate before pruning down to only the most recent one. This kept the early history stable for longer between cleanups. Combined with the larger history budget, this strategy became the final optimized workflow. Each configuration performed the exact same task: collecting 6 fields for each of 32 books from a public demo catalog, representing a typical StackAI use case for some Asana customers.
| Model | Description | Price |
|---|---|---|
| Model A | A smaller, cheaper model from another frontier lab, released in Fall 2025 | Half the price of GPT‑6.1 Sol |
| Model B | The original model used in production, from the same lab as Model A, released in Summer 2026 | Same price as GPT‑6.1 Sol |
| Model C | An updated version of Model B, released in Fall 2026 | Same price as GPT‑6.1 Sol |
| GPT‑6.1 Sol | OpenAI’s model | Baseline price |
GPT‑6 Astra ran these workflows, inspecting requests, usage logs, and outputs, with separate model sessions auditing the work. Requests, traces, and results from each session were logged in Command, Asana’s software delivery platform, allowing the team to review the full study afterward. Through Command, these findings were converted into tickets, which then generated pull requests and deployed the changes to production.
“By hand, this would have taken one to two months. In Codex with GPT-6 Astra, it took about a week: I could set a
/goalbefore bed and wake up to the results.” — Dr. Frank Hidalgo, CTO of StackAI at Asana
Cutting Model Costs to Under $0.50 per Run
For Model B, the optimizations dropped the estimated model cost from at least $36.21 (several original runs hit step limits before finishing) to $1.24 per run—a 29x reduction. The optimized workflow on GPT‑6.1 Sol was even cheaper at $0.47, a further 2.6x decrease. Every run in the optimized workflow successfully completed the task and returned the correct answer.
Averages across 3 runs. ≥: Baseline includes runs that hit step limits, so the average is a lower bound.
The two rightmost columns compare the optimization results for Model B. In the same study, Model B was run in Phase 1, while Model C and Sol 6.1 were run in Phase 2 (dashed line).
On GPT‑6.1 Sol alone, paired with the larger history budget, the new caching and screenshot pruning strategy reduced costs 4x, from $1.97 to $0.47 per run. Because 89% of inputs came from cache (with cache hits priced at just 5% of the uncached rate), cost per call dropped by roughly 3x. Runs were also dramatically faster: the original setup on Model B took at least 22.5 minutes, compared to roughly 4 minutes for the optimized workflow on GPT‑6.1 Sol.
Averages across 3 runs, error bars indicate standard deviation (SD whiskers). ≥: The average includes runs that hit step limits or timed out, so true values are at least this high.
Charts rendered with a blue theme. See 480k budget charts for caching effects.
Run markers and SD whiskers approximated from source images; underlying per-run values and standard deviations unavailable.
Averages across 3 runs, error bars indicate standard deviation (SD whiskers). ≥: The average includes runs that hit step limits or timed out, so true values are at least this high.
Charts rendered with a blue theme. See 480k budget charts for caching effects.
Run markers and SD whiskers approximated from source images; underlying per-run values and standard deviations unavailable.
The study also revealed how history management directly impacts whether an agent can return an answer. Giving GPT‑6.1 Sol more room to retain browsing history led to a dramatic increase in successful runs: with the smaller history budget, only 3 of 18 runs answered; with the larger budget, all 18 runs returned the correct answer. For Hidalgo, the business value lies in offering customers faster, more capable models while keeping operational costs sustainable.
“Previously, cost limited which models we could offer customers for this kind of workload. By making the agent more efficient, we can offer better, faster models to customers while lowering operational costs.” — Dr. Frank Hidalgo, CTO of StackAI at Asana
Scaling Experiments and Product Testing
Asana has already rolled out these browser navigation improvements in StackAI and is building tools to make replicating similar experiments easier. Over time, the team plans to incorporate such testing into the platform’s evals so customers and internal teams can compare cost, runtime, and answer quality when configuring agents.
“Shipping velocity is no longer the bottleneck—human attention is. We’re rapidly approaching a world where every engineer is a PM leading a fleet of agents.” — Dr. Frank Hidalgo, CTO of StackAI at Asana
Asana is now using GPT‑6 Astra in Codex to test product features before release: Astra interacts with the platform, tries different inputs, and reports bugs to human QA reviewers. Hidalgo sees this as the foundation of a new software development lifecycle, where fleets of cloud-based agents test features in parallel.