OpenAI says Asana cut the estimated model cost of a browser-agent task to 47 cents per run, while making it about five times faster than its previous production setup. The case study shows how changes to browsing history and screenshot handling, alongside GPT-6.1 Sol, produced the gains.
OpenAI Watch analysis
What happened
Asana’s StackAI team used GPT-6 Astra in Codex to investigate a browser agent that collected six fields from each of 32 books in a public catalogue. In a 144-run study, it found that the agent repeatedly sent its growing browsing history at full price and discarded useful information too often.
The team tested changes including caching more of that history, increasing the amount of text retained and removing screenshots in batches rather than at every step. OpenAI says the best workflow averaged $0.47 and about four minutes per run on GPT-6.1 Sol. That is 76 times cheaper and five times faster than the original production setup, which used another model. Asana says every run with the optimised workflow completed the task and returned the correct answer. Read OpenAI’s case study.
Why it matters
The useful lesson is not simply that a newer model did the job. On GPT-6.1 Sol, the workflow changes cut estimated cost fourfold, from $1.97 to $0.47 per run. OpenAI says 89% of input tokens in the optimised setup came from cached input, charged at a lower rate.
History management also affected whether the agent produced an answer: with the larger history budget, all 18 tested runs returned the correct answer, compared with three of 18 using the smaller budget. These are results from a small, task-specific study, not a promise that every browser agent will suddenly become a bargain.
Our read
This is a concrete example of AI-assisted engineering finding savings in the plumbing, not just swapping in a shinier model. Keeping useful page history intact and trimming screenshots less often sounds less glamorous than “agentic transformation”; it also appears to have mattered more to this task.
The results are promising, but the comparison has limits: the 76-fold saving is against the old setup on a different model, and the reported averages are based on three runs per configuration. Teams should treat the methods as worth testing on their own workloads, not as a ready-made forecast for their next invoice.
What to watch
- Whether Asana publishes results from a wider range of browser-agent tasks.
- Whether customers can compare cost, runtime and answer quality when configuring agents.
- How often the same history and screenshot changes improve performance outside this catalogue task.
Discussion spark: When an AI agent gets cheaper through better workflow design, should teams prioritise that work before switching to a newer model?
Sources and evidence
- Asana cuts model costs 76x in browser tests with GPT-6.1 Sol – OpenAI (9 October 2026, 17:53 UTC)
OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.