Browser agent costs cut 76x! One week with Asana and GPT-6 Astra

Hi, it's Shii-chan! OpenAI shared a story of Asana making its browser agent much cheaper and faster, so let me walk you through it.

OpenAI News openai.com

What was announced?

OpenAI News published "Asana cuts model costs 76x in browser tests with GPT-6.1 Sol". StackAI, which Asana acquired, lets customers build workflows that navigate websites, fill out forms, and gather information without writing code. Its CTO directed GPT-6 Astra in Codex to investigate the browser agent, test improvements, and compare results.

Why it matters

At Asana's scale, small inefficiencies in these workflows add up to real cost. The post argues that lower costs let Asana offer customers more capable models.

What changes

The optimized workflow on GPT-6.1 Sol averaged an estimated $0.47 per run and about four minutes. That's 76x cheaper and 5x faster than the original production setup on Model B. The CTO estimates work that would take one to two months by hand took about a week.

Dive Deep

What GPT-6 Astra found:

• The agent cached its fixed instructions and tool definitions, but not the growing history of page text and screenshots, so every request resent that history at full price
• The agent dropped older screenshots and trimmed text at nearly every step, which changed the history each time, so caching the history alone wouldn't have helped

The three fixes tested:

• Extend caching to the browsing history
• Increase the amount of text it can retain
• Remove screenshots in batches rather than at every step

The study had 144 runs: history budgets of 120,000 and 480,000 characters, six caching and screenshot policies, three runs each on four models. The task was collecting six fields for each of 32 books from a public demo catalog. The best policy let screenshots accumulate to 20 before cutting back to the most recent one, keeping earlier history unchanged for longer.

The numbers:

• On Model B, estimated cost fell from at least $36.21 to $1.24 per run (29x). Some original runs hit the step limit, so $36.21 is a lower bound
• On GPT-6.1 Sol, the new policy cut cost from $1.97 to $0.47 per run (4x), with 89% of input coming from cache at 5% of the uncached price
• With the larger history budget, runs that produced an answer went from 3 of 18 to all 18, each correct

Asana has already released the changes to browser navigation in StackAI and is now using GPT-6 Astra in Codex to test product features before release.

Wrap-up

• Rethinking history caching and screenshot removal was the key
• $0.47 and about four minutes per run on GPT-6.1 Sol
• 76x cheaper and 5x faster than the original setup, in the test conditions
• Astra handled the experiment setup and study, while humans set direction and reviewed

It's a great case study for engineers running agents in production and worrying about cost.