A vendor that rations is saying something about cost, and OpenAI's stated reason for holding Astra back at first is that it needs to scale capacity for what a company spokesperson called "a very large model" [11]. That is the most useful sentence in the launch for anyone drafting an automation budget, because it prices the capability in the currency that is actually scarce, serving capacity rather than model quality, and because there is no published cost per session and no error rate on repetitive clerical work anywhere in the announcement [23].
Read the benchmark spread precisely. Astra's ARC-AGI-3 result sits 90.8 points above the model OpenAI shipped before it, roughly 12.6 times the score [19], and about 3.3 times Anthropic's Claude Opus 5 [20]. ARC-AGI-3 is meant to test reasoning applied to situations a model has not seen before [24], which is nearly the opposite of the work being budgeted: the same purchase-order layout again and again, where the expensive failure is a confidently wrong field that nobody catches until reconciliation. The launch demonstration itself was a person instructing a computer by voice in a video that Fortune described as seamless albeit staged [4].
Notice what a finance team that underwrites next year's clerical headcount on this model is quietly deferring: permissions and audit logs, the unglamorous plumbing that decides whether an agent is allowed near a general ledger at all. Anthropic pioneered computer use with a beta in 2024 and Perplexity shipped a system that navigates a virtual computer in February, and the concept is still not a mainstream part of working on a computer for most people [15]. Two years of availability without adoption is evidence that the binding constraint sat downstream of the model.
If the reason computer use never landed was simply that the models were not good enough, then the 90.8 points are the whole story and the integration objection is a lagging indicator that will look silly by summer. More likely, both hold, and sequencing decides the budget: the capability arrives now, the permissions arrive whenever IT and finance finish arguing about who signs off when an agent submits the wrong form. Brockman told reporters it is not unreasonable to feel we are in the AGI era and reasonable to call this the first such model [16]; OpenAI once defined AGI as a system performing all economically valuable work as well or better than humans [17], which is a labor-cost claim, and labor-cost claims get settled by invoices rather than scores.
A published price per computer-use session alongside an error rate on a named real workflow would make the substitution arithmetic possible for the first time, and it would move me. A general rollout that is not capacity-gated would move me too, telling you OpenAI thinks the marginal session is cheap enough to sell in volume. Until one of those lands, the honest line item is a pilot.
