Asana's browser agent costs were put on display in a customer story OpenAI published on October 9, 2026: after an optimization round, the browser agent in StackAI, Asana's no-code workflow platform, ran on GPT-6.1 Sol at an average estimated model cost of $0.47 per run in about four minutes — 76x cheaper and 5x faster than the original production setup on another lab's model, called Model B in the story. The number worth studying is not the multiple, but how it was achieved.
Where the money went: resending the whole history at every step
StackAI lets customers build workflows without code, sending agents to open websites, fill in forms and gather information. StackAI CTO Frank Hidalgo wanted the browser agent faster and cheaper, and instead of auditing it by hand he directed GPT-6 Astra in Codex to map the codebase, propose fixes and run controlled comparisons — work he estimates would have taken one to two months manually, done in about a week. What Astra found was specific: the agent cached its fixed instructions and tool definitions, but not its growing browsing history of page text and screenshots, so every step resent everything it had already seen to the model at full price. Worse, it dropped older screenshots and trimmed text at nearly every step, so the history kept changing and could not be cached even in principle — and facts it discarded could force it to revisit pages it had already read.
Three fixes and a 144-run comparison
Hidalgo picked three of Astra's proposals to test: extending caching to the browsing history, retaining more text (comparing history budgets of 120,000 and 480,000 characters), and removing screenshots in batches (the best policy let them accumulate to 20 before cutting back to the most recent one, keeping earlier history unchanged for longer). The 144-run study had each configuration collect six fields for each of 32 books from a public demo catalog, across GPT-6.1 Sol and three anonymous models. The result needs unpacking: with workflow changes alone and no model switch, Model B's cost per run fell from at least $36.21 (a lower bound — some original runs hit the step limit unfinished) to $1.24, roughly 29x. Moving to the optimized workflow on Sol brought it to $0.47, another 2.6x. Most of the 76x, in other words, came from the workflow itself; the model switch added the rest.
How much history you keep also decides whether the task finishes
The same study produced a result harder than price: on Sol with the smaller history budget, only 3 of 18 runs produced an answer; with the larger budget, all 18 finished with the correct answer. Looking at Sol alone, the new caching and screenshot policy cut cost from $1.97 to $0.47 per run, each call about 3x cheaper because 89% of input came from cache at 5% of the uncached price. Asana has shipped the changes in StackAI's browser navigation and plans to build this kind of comparison into the platform's evaluation tools.
The caveats deserve equal billing: this is a vendor-published single-customer case, costs are estimates, the comparison models are anonymous, the task is one type, and 76x is not a benchmark any browser agent can expect to reproduce. But it leaves every agent-building team a check they can run today: when a multi-step agent keeps getting more expensive, do not rush to switch models — count how much of each request is content being resent at full price. How history is kept, cached and pruned is often the real bill.
