Asana reduces browser‑agent expenses 76‑fold using GPT‑6.1 Sol

Significant Cost and Speed Improvements

Asana reported that its browser‑agent workflow, powered by GPT‑6.1 Sol, now runs about 76 times cheaper than its prior setup.

The optimized workflow costs an estimated $0.47 per run and completes the task in roughly four minutes.

This represents a five‑fold speed increase compared with the original production model, identified as Model B.

Model B and GPT‑6.1 Sol are priced the same, while Model A is half the price of GPT‑6.1 Sol.

In a 144‑run study, Asana evaluated GPT‑6.1 Sol alongside three other frontier models labeled Models A, B and C.

The study measured each model under varying history budgets and caching policies to determine the most efficient configuration.

Results showed that a policy allowing up to 20 accumulated screenshots before pruning delivered the best performance on GPT‑6.1 Sol.

Combined with a larger text‑history budget, this policy formed the optimized workflow now in production.

How the Optimizations Were Discovered

Frank Hidalgo, PhD, Asana’s StackAI CTO, tasked GPT‑6 Astra in Codex with mapping the browser‑agent codebase.

GPT‑6 Astra identified that the agent repeatedly sent its full browsing history and screenshots with each request, incurring high token costs.

The agent also trimmed text and dropped older screenshots at nearly every step, preventing simple caching of history.

Hidalgo selected three remedial actions: extending caching to include browsing history, increasing the retained text length, and batching screenshot deletions.

GPT‑6 Astra first ran quick experiments to gauge variable impact before refactoring the code for parallel workflow testing.

The full study explored two history budgets—120,000 and 480,000 characters—and six combinations of caching and screenshot policies.

Each configuration was executed three times on each of the four models, producing a comprehensive performance matrix.

The task used in testing required extracting six data fields for each of 32 books from a public demo catalog, mirroring typical StackAI customer jobs.

Implications for Enterprise Automation

By cutting model costs dramatically, Asana can scale its no‑code automation platform, StackAI, without inflating cloud spend.

Faster runs enable more responsive workflows, improving user experience for enterprises that rely on web‑based data collection.

The study demonstrates that AI‑assisted code analysis can replace months of manual engineering effort with a week‑long automated process.

Arnab Bose, CPO at Asana, highlighted the collaboration: “This is what teams of humans and agents look like in practice.”

“An engineer set the direction, GPT‑6 Astra ran the experiments, and the results went through Command to production.”

“This demonstrates how Asana brings human and agent teams to life.”

The quote underscores Asana’s strategy of pairing human oversight with AI‑driven optimization.

Such a model could influence other enterprise software providers seeking to reduce operational AI costs.

However, the findings are based on a controlled benchmark; real‑world workloads may exhibit different patterns.

Overall, the work validates the practical benefit of using frontier models like GPT‑6.1 Sol in production pipelines.

Why This Matters: The cost reduction makes high‑performing AI agents financially viable for large‑scale enterprise automation, opening the door for broader adoption of AI‑driven workflows.