Asana Retooled Its Browser AI and Cut Model Costs to 1/76th
GPT-6 Astra reworked how the job gets done, and GPT-6.1 Sol executed it. In a test that gathered book data, estimated model cost came to $0.47 per run, with each run taking about four minutes.
Asana says it has reduced both model costs and run times by changing how its AI handles tasks on websites. According to a case study published by OpenAI on October 9, the newly retooled GPT-6.1 Sol setup carried an estimated model cost of 1/76th that of the previous production setup. Run time fell to one-fifth. Asana says it applied the changes to StackAI’s browser navigation feature.
What GPT-6 Astra and GPT-6.1 Sol Each Did
Two different models were used: one for the improvement work and one for actual execution. GPT-6 Astra was tasked with analyzing code in Codex, running experiments, and finding inefficiencies. GPT-6.1 Sol was tested as the model that actually runs the reworked browser workflow. The cost reduction figures come from the setup that used GPT-6.1 Sol as the execution model.
StackAI is a service that lets customers set up website navigation, form filling, and information gathering tasks without writing code. The target of this round of improvements was the AI agent that handles tasks on its own inside the browser. The team measured the cost and time involved when a user hands off the job of moving through web pages and collecting information.
How the Cost Dropped to 1/76th
The test covered four models across a total of 144 runs. Each setup was assigned the task of gathering information on 32 books from a public demo book catalog, with six fields to collect per book. Among the comparison models, Model A cost half as much as GPT-6.1 Sol, while Model B and Model C were priced the same. Model B was the previous production model.
The newly retooled GPT-6.1 Sol setup averaged $0.47 and about four minutes per run. The 1/76th cost and one-fifth time figures are measured against the previous Model B production setup. Model B’s estimated cost per run also fell, to $1.24, once the same workflow changes were applied to it. In other words, this comparison bundles together the effect of switching execution models and the effect of changing how the work is done.
The changes centered mainly on how task history and screens are handled. The team broadened the scope for saving and reusing records of web pages the agent had already visited, and increased the volume of text it can retain. Screenshots went from being cleared out at every step to being cleared in batches, several at a time. After the history retention limit was raised, the number of runs in which GPT-6.1 Sol produced an answer rose from 3 out of 18 to all 18, and every one of those answers was correct.
Will the Same Savings Show Up in Real Work?
OpenAI presented these costs as estimates rather than actual billed amounts. It noted that some of the previous Model B runs hit step limits and failed to finish, which may have pulled the averages for the old cost and time figures lower than they really were.
The change users will notice applies to tasks that involve handing off web navigation and information gathering. For the same job, costs and time went down because both the execution model and the working method changed together. In the test with higher history retention, the rate at which the agent produced an answer also improved. The open question is whether the gains seen in book catalog collection carry over to the other web tasks customers run.