
OpenAI’s $0.085 Healthcare Benchmark Still Needs a Payroll Number
The models can do useful work cheaply. The investment case rests on how much paid human work they actually retire.
CVS Health says a computer agent cleared a months-long document-sending backlog over a weekend, according to a 16 November 2026 account. Another handles system steps for tens of thousands of consults daily. The paperwork is moving, and the missing labor economics matter precisely because the agents are already working. We would pass on paying a premium for unverified earnings from those agents: CVS describes throughput without a net labor ledger. Reduced payroll, avoided hiring and increased capacity are different financial outcomes, even when the same agent produces them. A useful agent and a profitable agent belong in different valuation models.
Cheap model attempts make supervised healthcare automation attractive. The investment claim rests on human time returned per finished workflow. The Completion Tax is the incremental expense per successfully completed production case above the benchmark model attempt cost per task, required to deliver a correct, authorized, reviewed workflow. It equals fully loaded cost per successfully completed production case minus benchmark-attempt cost, including human review, corrections, integration and runtime expense allocated to each successfully completed case. That additional bill has to fit inside the productivity gain. Treating a model attempt as finished work books savings before you earn them.
OpenAI's 16 November 2026 healthcare evaluation puts a GPT-6 Luna administrative task at $0.085, 98% below its GPT-6 Astra model evaluation cost. Luna earned 91.8% of available HealthAdminBench rubric points against Astra's 94.3%, with both models at maximum reasoning effort. That is an attractive trade: far cheaper attempts for fewer rubric points. Rubric points measure evaluation performance, rather than finished production jobs. The hospital still has the full workflow to fund. When you capitalize $0.085 as a complete workflow cost, most of the value sits in your spreadsheet.
The U.S. Bureau of Labor Statistics puts the 2025 median wage of medical records specialists at $24.59 an hour. OpenAI's $0.085 evaluation cost divided by the wage per minute ($24.59 an hour divided by 60) is 0.207 minutes, or about 12.4 seconds of wages. The attraction is real: modest human time saved covers the bare model input. A CTO evaluating repetitive work wants that gap to be wide; an investor wants to know how often it survives the actual review queue. That wage is a broad proxy, and the calculation excludes correction, review and implementation. Profit depends on eliminated payroll, avoided hiring or monetized capacity exceeding the all-in cost per successfully completed case.
CVS offers a tempting scale story. Its FY2025 Form 10-K lists $402.067bn in revenue and $44.977bn in operating expenses. In a hypothetical scenario where an agent genuinely avoids one wage-minute, $24.59 divided by 60 gives approximately $0.4098 in avoided wages, leaving $0.4098 minus the $0.085 evaluation cost, or $0.3248, before other expenses. That hypothetical assumes paid labor genuinely disappears rather than simply becoming available for other work. CVS discloses neither the completed-workflow count nor the attributable labor reduction needed to multiply this unit example into company-wide savings. You see the temptation, but a large expense line becomes investment value only when some of it disappears.
OpenAI provides a second reason to take the advocates seriously. Its clinical tests found 99% of 325 reviewed tasks rated safe, and 93% of 646 physician helpfulness assessments rated the work "very helpful" or "extremely helpful." A synthetic patient demonstration has an agent prepare an MRI order, a clinical note, a reply and an insurer authorization request for human review. Supervised assistance already has a plausible operational case. Yet 646 assessments describe outputs, and $0.085 measures an evaluation input. The investable number is net labor minutes saved per completed, approved case.
CVS Health's production account makes the strongest opposing case: supervised automation can create value through increased capacity or avoided hiring without eliminating a worker. Lagan Khare, its Executive Director, Engineering, says "Less time on administrative work means more time for care. That’s a win for our teams and everyone we serve." That deployed throughput establishes useful operational activity. Human-supervised agents could pay for themselves by accommodating demand without additional staff or allowing existing employees to complete more productive work while current payroll remains intact. Testing that economic case requires counterfactual staffing, successfully completed volume and incremental contribution. Without those measures, the potential benefit remains outside an earnings valuation. Twelve-point-four seconds of avoided wages offsets only the reported $0.085 benchmark cost, before review, corrections, integration or runtime expense.
An independent study submitted on 10 April 2026 describes the problem in completed jobs. Across 135 simulated administrative tasks comprising 1,698 verifiable subtasks, GPT-5.4 CUA succeeded on 82.8% of subtasks and finished 26.7% of whole tasks. Steps can succeed while whole workflows stay unfinished, adding expense above the model attempt before a completed case produces economic value. The highest whole-task success rate in that study was 36.3%, from Claude Opus 4.6 CUA. This concerned older models and simulated interfaces, with different tasks and configurations from GPT-6; it measured neither current GPT-6 production completion nor human rework costs. The failure mode still matters because an unfinished case requires further work before its labor saving can be booked. Each incomplete job eats into the saving that made the cheap attempt attractive.
Look at how the clinical demonstration is arranged. The agent reads synthetic visit and imaging information, drafts a note, order and patient reply, and hands them to a clinician; an authorization specialist reviews the request for the insurer. OpenAI describes configurable approvals, access controls and audit logs for production users. The arrangement is designed to speed staff while retaining human approval. At CVS, the incremental expense above the model attempt includes human-review wages, correction costs, integration spending and runtime charges that the disclosed activity leaves unpriced. You own the benefit only when the economic value of completed cases exceeds those costs.
Cheaper authorizations and appeals change the incentives on both sides: providers and payers could each process more transactions as individual tasks become less expensive. A growing queue of faster machine actions could leave the labor budget stubbornly intact, because an operator saves cash through fewer paid minutes per completed workflow rather than an increased volume of actions.
A corporate purchaser can demand completed-case counts, reviewer minutes and the full deployment bill before signing. For shareholders, public disclosures establish which operational gains deserve an earnings multiple. OpenAI supplies the capability; CVS operates the workflows; the illustrated clinical use still reserves approval for people. The public numbers leave savings unassigned, even though each participant has an obvious reason to claim some of the gain. Valuing model efficiency as hospital earnings assumes a cost split nobody has disclosed. The Completion Tax, measured as the extra completed-workflow expense above the model attempt, determines whether that gain reaches earnings.
We would pass on capitalizing unverified healthcare automation earnings while recognizing the operational value of supervised agents. CVS's fiscal 2026 Form 10-K would prove us wrong if it documented successfully completed, reviewed consult volume and verified economic value per case from eliminated payroll, avoided hiring or monetized capacity exceeding all-in deployment cost per successfully completed case, including review, corrections, integration and runtime. You would then have a measurable unit of productivity to capitalize. Until then, $0.085 measures the evaluation attempt; the investment premium waits for the time sheet.