**Asana** and OpenAI said that after revamping how its AI browser agent reuses data, Asana switched to OpenAI’s GPT-6.1 Sol model, cutting the cost per run by 76 times in internal tests.
Key points
The optimized Asana agent averages $0.47 per run on GPT-6.1 Sol, down sharply from at least $36.21 when using competing models.
Simply revamping the workflow structure before changing models had already cut costs by 29 times.
The figures come from tests conducted under specific, narrow conditions and have not been independently audited by an external organization.
Asana tests StackAI
OpenAI disclosed the results in a case study dated October 9. The study was based on a 144-run test that Asana engineers published the previous day.
The test used a browser agent in StackAI, the automation platform acquired by Asana. The agent can navigate websites, fill out forms, and gather information without writing code. It was instructed to collect information on 32 books in each run.
In the existing production environment, using a competing model cost at least $36.21 per run and took at least 22.5 minutes. Some runs hit the step limit midway and could not be completed. By contrast, the optimized version using GPT-6.1 Sol cut the average cost to $0.47 per run and the time to about four minutes, returning the correct answer in every run.
After analyzing the browser agent’s behavior inside the Codex coding tool, OpenAI’s GPT-6 Astra found that the agent was paying full price by resending its history—which consisted of page text and screenshots—almost unchanged with nearly every call. Astra then tested improvements such as caching and reusing the history, accumulating up to 20 screenshots before reducing them to one, and increasing the history length limit fourfold. After GPT-6.1 Sol was introduced, 89% of input tokens were supplied from the cache.
Read also: ChatGPT’s intelligent UI embeds charts and calculators directly into answers used by 1.2 billion people
Frank Hidalgo explains the cost structure
Frank Hidalgo, StackAI’s chief technology officer (CTO), explained that research work that would have taken one to two months if done manually was completed in about a week. He said the high operating costs of browser agents had limited the range of models they could offer customers. By building a “lighter” agent, they were able to use a faster model while also lowering overall operating costs.
In a blog analysis, AI routing company OrcaRouter noted that much of the cost savings came from fixing the workflow itself, rather than switching models.
Changing only the workflow reduced costs by 29 times with the competing model. After that, switching to GPT-6.1 Sol cut costs by a further roughly 2.6 times at the same list price. OrcaRouter analyzed that this gap was related to how efficiently GPT-6.1 Sol uses caching.
However, all of these figures were calculated by Asana and OpenAI themselves; no official audit or verification by an independent third party was conducted.
This is not Asana’s first publicly disclosed use of Codex. On August 7, Asana engineers said they had used a Codex agent to complete the migration of frontend tests from an obsolete library in about two weeks—a task previously estimated to take nearly five years. The total cost at the time was about $12,000.
Next article: Can TRON fend off a quantum attack on its $95 billion in stablecoins? Sun says it’s “ready”