Asana removed Enzyme, an obsolete testing system that was blocking its frontend upgrades, in about two weeks. Model and infrastructure costs came to roughly $17,000. The plan it replaced was estimated at five years and $8.5 million.
OpenAI published the case this week. The same week, Gartner predicted that inference costs per agentic workflow will rise more than fivefold through 2028. Both can be true, and the gap between them is where the useful part sits.
What Asana actually did
The setup was smaller than the numbers suggest. A five sentence prompt, up to four Codex agents working in parallel, each in a separate copy of the codebase, and an engineer checking progress twice a day and approving every proposed change, according to OpenAI.
Total engineering effort was about a week and a half, spread across two calendar weeks. Asana reported that simpler instructions worked better than a more elaborate setup, which is the opposite of how most teams approach a first agent project.
The comparison has an asterisk
The $17,000 is a real cost. The $8.5 million is not: it is Asana’s own estimate of a staffing plan that never ran. OpenAI published both figures in United States dollars, converted here at the Reserve Bank rate. Setting money spent against a budget nobody spent produces a ratio near 500 to 1, and that ratio is doing a lot of work in the headline.
The honest reading is narrower and still remarkable. A job the business had declined to fund for years became cheap enough to attempt. Chief technology officer Amritansh Raghav said as much: not every years long project collapses into weeks.
Gartner says the opposite is coming
Gartner expects inference costs per agentic workflow to rise more than fivefold through 2028, and calls it the inference paradox: the rate of innovation is outpacing the cost curve, as reported by ARN. Cheaper tokens are being eaten by workflows that use far more of them.
Senior director analyst Will Sommer put it plainly: product leaders cannot rely on more efficient token economics to rationalise AI costs, because each generation of capability needs more tokens, and often more expensive ones.

Why Australian businesses should pay attention
The two stories describe different work. Asana ran a bounded, mechanical migration with a clear finish line and a human approving every change. Gartner is describing agents that run continuously inside a product, reasoning and negotiating and checking themselves.
One is a project with an end date. The other is an operating cost that compounds. Budgeting for the second using the numbers from the first is how an AI strategy ends up underwater by month four.
For Australian teams the near term opportunity is the first kind. Every codebase here has a migration nobody funded: a dead framework, an unsupported library, a test suite nobody trusts. Those now carry a price a chief financial officer can look at without flinching.
Which jobs this actually fits
The pattern that worked at Asana is specific and worth copying: a mechanical change across many files, a definition of done a machine can check, parallel agents on separate copies, and a human reviewing at a fixed cadence. Teams already running their own agents have most of the pieces.
What it is not is a template for open ended product work. The same week that produced this case also produced the research on agents in groups, where agents with conflicting goals sabotaged each other. The difference is the finish line.
The number worth remembering is not 500 to 1. It is that a five sentence prompt beat an elaborate setup, and that a human still approved every change.
