Asana cut its AI browser agent's costs 76 times in internal tests after fixing how the agent reused data and switching to OpenAI's GPT-6.1 Sol model, the companies said.
Key Points:
- Asana's optimized agent averaged $0.47 per run on GPT-6.1 Sol, down from at least $36.21 on a rival model.
- The workflow changes alone cut costs 29 times before any model switch.
- The figures come from one narrow test and have not been independently audited.
Asana StackAI Test
OpenAI published the results in an Oct. 9 case study, a day after Asana engineers released the 144-run study behind them.
The tests covered a browser agent inside StackAI, an automation platform Asana acquired, which navigates websites, fills out forms and gathers information without code. Each run asked it to collect data on 32 books.
Running on a rival model, the original production setup cost at least $36.21 and took at least 22.5 minutes per run, since some runs hit a step limit before finishing. The optimized version on GPT-6.1 Sol averaged $0.47 and about four minutes, and every run returned the correct answer.
OpenAI's GPT-6 Astra, working inside the Codex coding tool, found that the agent resent its growing history of page text and screenshots at full price on nearly every call. Astra then tested fixes that cached the history, let screenshots pile up to 20 before cutting back to one and quadrupled the history limit. On GPT-6.1 Sol, 89% of the input then came from cache.
Also Read: ChatGPT’s Intelligent UI Puts Charts And Calculators In Answers For 1.2B Users
Frank Hidalgo On Costs
Frank Hidalgo, StackAI's chief technology officer, said the research would have taken him one to two months by hand but took about a week. Cost once limited which models Asana could offer customers for this work, he said, and a leaner agent lets it offer a faster model while cutting operating costs.
An analysis by AI routing firm OrcaRouter argued that most of the savings came from fixing the workflow, not from switching models.
The same changes cut costs 29 times on the rival model. GPT-6.1 Sol supplied a further 2.6-fold cut at the same list price, a gap the analysis tied to how well the model used its cache.
No outside party has audited the figures.
It is not Asana's first public Codex result. On Aug. 7, its engineers said Codex agents moved the company's front-end tests off an aging library in about two weeks, a job they had expected to take roughly five more years. The bill came to about $12,000.
Read Next: Can TRON Shield $95B In Stablecoins From Quantum Attacks? Sun Says It’s Ready
