OpenAI's Asana case study prices 1.5 weeks of engineer time against a five-year staffing plan

Enzyme and React Testing Library are not two dialects of the same test. An Enzyme test operates on a component instance; RTL asserts against the rendered DOM, so the test sees the whole rendered page rather than the component it was written for ^[6]. Assertions have to be re-expressed in terms of what a user can see, and for complex user journeys the two approaches diverge sharply ^[7]. Asana went from the first to the second ^[4]. Both sides of the comparison are describing genuinely awkward work.

The units are where it comes apart. The arithmetic behind the $6M is four engineers at circa $300K a year for five years, according to Gergely Orosz's reading of the case study ^[5], and it multiplies out cleanly to the published figure ^[16]. That is also about 1,040 engineer-weeks of budgeted labour ^[17]. The other side of the comparison is 1.5 weeks of engineering effort ^[3], and the write-up does not say how many engineers those 1.5 weeks cover, which is a curious omission in a comparison this lopsided.

The $12K is model and infrastructure spend, not the cost of the work ^[2]. Charge the engineer time at the salary the $6M itself assumes and you add roughly $8,650, for about $20,700 all in ^[18]. Against an advertised gap of $5.99M ^[19], that is noise either way.

Airbnb's numbers let you test the rate the plan assumed. Its manual estimate was 1.5 engineering years for 3,500 Enzyme component test files ^[8]. At 250 working days a year, that implies about nine files per engineer-day ^[14], the top of the five-to-ten range Orosz guesses Asana's planners used ^[20]. Push Asana's implied 20 engineering years through Airbnb's rate and the plan is sized for roughly 46,700 test files ^[15]. If Asana's suite is nowhere near that, the five years measured appetite rather than throughput. Orosz puts it plainly: you estimate a project like this when it is work you really do not want to do ^[13].

Airbnb also shows where the effort actually sits. Its team built loops that retried failed migrations, and once those existed 75% of files went through in four hours ^[9]. The remaining quarter needed a purpose-built refactor pipeline, which took the total to 97% over four days, with engineers finishing the last 3% by hand over a week ^[10]. That was March 2025, on Claude 3.7 Sonnet ^[11]. The bulk was hours; the tail was a week of humans.

Orosz still finds two weeks credible for Asana, a year on from Airbnb ^[12]. The outcome can stand while the denominator gets audited. Whether a saving like this transfers depends on who authored the counterfactual and whether they wanted the work, the unit both figures are denominated in, and what the smaller figure leaves out. Calendar weeks set against engineer-years fails on the second before you get to the third.