Asana Reports 76-Fold Cost Reduction for Browser Agent After Workflow Changes
In an internal study, Asana describes how cache and history changes, together with deploying GPT-6.1 Sol, reduced the cost of a browser agent test run to $0.47.
Asana browser agent was the subject of an internal optimization study that Asana published on October 8 and OpenAI published as a case study on October 9, 2026. The company says that after workflow changes and the use of the GPT-6.1 Sol model, the average test-run cost fell to $0.47 and its duration to approximately four minutes.
Asana compares these figures with the original deployment on an unidentified “Model B,” for which it reports at least $36.21 and at least 22.5 minutes per run. From this, it derives approximately 76-fold lower costs and a fivefold shorter duration. However, this is not an independent benchmark or an isolated comparison of the models themselves: both a workflow change and a switch to GPT-6.1 Sol contributed to the result.
Asana browser agent: cache and growing history
The optimization involved the StackAI browser agent, whose workflow was analyzed and modified by GPT-6 Astra in Codex. Asana focused mainly on prompt-cache usage and the agent’s continually growing interaction history.
The changes included retaining cache-suitable history, bulk removal of screenshots, and a larger history budget. With browser agents, context can grow with each navigation or task-execution step. If a relevant portion of the input remains unchanged, the model interface may handle it as a cache read rather than repeatedly processing the entire input.
Asana says that cache reads accounted for 89% of the input in the optimized run. According to the company, the resulting costs and latency are therefore determined not only by model choice but also by how the system manages context and repeatedly used data.
The test included 144 runs
The company tested six cache and history policies at two history sizes and on four models. It performed three runs under each condition, for a total of 144 runs. It then added a 12-run test. The published results therefore provide insight into a specific test task and configuration, not the universal properties of all browser agents or language models.
Asana also notes that its study is intended primarily to show general patterns. Given the scale of the measurement, it does not consider differences of several percentage points reliable. In addition, the company did not identify the original Model B in the comparison underlying the stated 76-fold cost reduction.
This limits the ability to assess externally what share of the improvement resulted from caching and history pruning, what share from the model change, and what share from other test parameters. The claim cannot be read as evidence that GPT-6.1 Sol is itself 76 times cheaper than competing models.
Asana has already deployed the changes
Asana says it has already deployed the browser-navigation optimizations to StackAI. Further validation of the results would require reproductions on other browser-agent tasks, a larger number of runs, and more detailed data on the compared models and measurement variance. It will also be important whether the company publishes production metrics beyond the test task.
Sources
- OpenAI – Confirms the publication date, the stated 76× and 5× results, the $0.47 cost, and the described use of GPT-6 Astra in Codex.
- Asana – Confirms the methodology involving 144 runs and a subsequent 12-run test, the scope of the comparison, the study’s limitations, and deployment of the changes to StackAI.
Verified and updated: October 10, 2026 06:28
