---
title: "The 8.5 cent employee"
description: "OpenAI says GPT-6 Luna earned 91.8% of rubric points on health admin tasks at 8.5 cents each. A five minute whiteboard video reads it against the first-party pages."
publishedAt: 2026-11-17
author: Ena Pragma
url: https://enapragma.co/field-notes/ep-intel-003-the-8-5-cent-employee
video: https://media.enapragma.co/videos/field-notes/ep-intel-003/ep-intel-003-v2-1080p.mp4
videoCaptions: https://media.enapragma.co/videos/field-notes/ep-intel-003/ep-intel-003-v2.en.vtt
tags: ["ep-intel", "ai-benchmarks", "computer-use", "healthcare-ai", "ai-verification"]
---

On Nov 16, 2026 OpenAI said its GPT-6 model Luna earned 91.8% of the available rubric points on Stanford's HealthAdminBench at 8.5 cents per task. Those are OpenAI's own numbers. Luna's token price is on OpenAI's price list: $0.10 per million input tokens and $0.50 per million output tokens. The 91.8% is partial credit. Stanford's benchmark counts a task only when every step passes, and in April its best agent finished 36.3% of whole tasks while getting 78.4% of the steps right. OpenAI did not publish a whole-task rate.

In the video, a hand-drawn character named Tally walks to each source page and marks the exact words. Every number is the source's own, and vendor numbers are labelled as the vendor's.

## Three yardsticks, three different bars

Stanford, Anthropic and OpenAI each report a different thing on this benchmark, so the three numbers do not sit on one scale.

<Yardsticks caption="Each bar has its own 0 to 100 track. The three figures measure different things and are not one scale.">
<Yardstick source="Stanford · Apr 10, 2026" value="36.3" unit="of whole tasks" note="The best agent in the paper, Claude Opus 4.6 CUA. The same agent got 78.4% of the steps right. A task counts only if every step passes." />
<Yardstick source="Anthropic · Jun 9, 2026" value="51.9" unit="of whole tasks" note="Anthropic's own run on its own port of the benchmark, one trial. Anthropic says the result is not directly comparable to the published leaderboard." />
<Yardstick source="OpenAI · Nov 16, 2026" value="91.8" unit="of rubric points" tone="partial" note="Luna, run and reported by OpenAI. The Stanford paper does not use the term rubric points, and OpenAI does not define it." />
</Yardsticks>

The Stanford paper says all verifiers return strict binary pass or fail outcomes, with no partial credit. OpenAI's post reports a share of points and does not give the whole-task number.

## What 8.5 cents is, and is not

For scale, the Bureau of Labor Statistics puts the median pay for medical secretaries and administrative assistants at $45,930 a year (May 2025). The CAQH 2025 Index puts the average cost to a medical provider of a manual transaction at $8.03, against $2.65 for an electronic one. That average covers all transaction types, not prior authorization alone.

OpenAI does not say how it counted the 8.5 cents. The Stanford paper has no per task dollar figure for these agents. Nothing in the post says review time or repeat attempts are in that number.

## What the announcement did not say

- What a rubric point is, which benchmark version OpenAI used, or how many runs it made.
- How the 8.5 cents was computed, or what Astra costs per task.
- The share of whole tasks that Luna or Astra finished.
- That Stanford ran or checked the scores. The post calls HealthAdminBench "an independent benchmark from Stanford researchers" and does not say Stanford scored the GPT-6 models.
- A safety number for Luna or for portal work. OpenAI's safety figure, 99% of 325 tasks rated safe, covers Astra on clinician inbox tasks.

## What the test leaves out

Stanford's authors took out the hard parts of real portals: multi-factor sign in, CAPTCHAs and session timeouts. The portals are static, with no interface drift or changing payer rules, and each task runs once. The authors warn that premature deployment risks operational errors with financial or clinical consequences.

OpenAI's own researcher told Fierce Healthcare: "In some cases, the model might overreach and might take an action without getting the proper sign-off or authorization."

## What OpenAI's own customer did

OpenAI names CVS Health as a production customer, and says CVS Health uses validation checks and human oversight, with screenshots and action records. It gives no backlog size, error rate or cost.

CVS Health's open source kit takes a different route from running a model on every run. An agent discovers a workflow once, and the working steps are compiled into ordinary Playwright code. The kit's README says asking a model to reason through the same steps on every run "adds inference cost and variability to work that ordinary code can perform." The codified workflow "runs without model calls or an API key."

## EP's take

1. The price is real and low enough that a no-API portal chore that costs a person minutes can now be tried for pennies, but the 91.8% is a partial-credit score on a simulated test, so do not plan around it.
2. Before buying anything, pick one repetitive portal task, run it twenty times on your own system, and count only the runs where every field is right and a check outside the screen agrees.
3. Where the steps never change, do what OpenAI's own customer did: let the agent learn it once, turn it into plain code, and keep a person on anything that cannot be undone.

## How it was made

The narration is a synthetic voice and Tally is a hand-drawn character. Every number and page on screen comes from the source named beside it. Two cards are typeset from the page text rather than captured from the live page, and the video labels them as typeset: the BLS wage table (bls.gov blocks automated capture) and the Fierce Healthcare quote, which is an outside report.

Related field notes: [Astra for Law is real. The benchmark a buyer actually needs is not on the page.](/field-notes/astra-for-law-is-real-the-benchmark-a-buyer-needs-is-not-on-the-page) and [Why a high AI benchmark score doesn't mean a reliable agent](/field-notes/ai-benchmark-scores-reliability).

<Transcript title="Transcript">

**0:00** On November 16, OpenAI said that its GPT-6 model, Luna, can do health care admin work for 8.5 cents a task.

**0:09** Here is the sentence, straight from OpenAI's own post.

**0:14** Luna earned 91.8 percent of the available points, at 8.5 cents per task.

**0:19** Points. Not whole tasks.

**0:23** First, the price. Luna lists at $0.10 per million input tokens.

**0:28** And $0.50 per million output tokens.

**0:31** Astra, the bigger model, lists at $10.00 and $50.00.

**0:36** OpenAI says Astra scored 94.3 percent.

**0:41** Now, what was actually measured? The test is called HealthAdminBench.

**0:46** It comes from Stanford, with Stanford Health Care and Kinetic Systems.

**0:51** It has 135 tasks, broken into 1698 checkable steps.

**0:59** They run on 4 simulated screens, and all of the data is synthetic.

**1:04** The screens are an EHR, two payer portals and a fax system.

**1:09** 60 tasks are prior authorizations. 60 are appeals and denials. 15 are equipment orders.

**1:19** Here is the catch. Stanford grades strictly: a task counts only if every step passes.

**1:25** The paper says all verifiers return strict binary pass or fail, with no partial credit.

**1:32** In April, the best agent got 78.4 percent of the steps right.

**1:37** But only 36.3 percent of whole tasks.

**1:41** In June, Anthropic ran the same benchmark on its own port, and reported 51.9 percent of whole tasks.

**1:49** It says that is not directly comparable to the leaderboard.

**1:52** OpenAI's 91.8 percent is a different thing. It counts rubric points.

**1:58** The Stanford paper does not use that term, and OpenAI does not define it.

**2:03** So three labs report three different yardsticks.

**2:07** And OpenAI did not publish the whole task number.

**2:11** Now, what does 8.5 cents actually mean? Start with a person.

**2:17** The Bureau of Labor Statistics puts a medical secretary's median pay at $45,930 a year, in May 2025.

**2:28** CAQH's 2025 index puts the average manual transaction for a medical provider at $8.03.

**2:37** The electronic one costs $2.65.

**2:41** That covers all transaction types, not prior authorization alone.

**2:45** OpenAI does not say how it counted the 8.5 cents a task.

**2:50** Stanford's paper has no per task dollar figure for these agents.

**2:54** And nothing in the post says review time or repeat attempts are in that number.

**2:59** Next, what the test leaves out.

**3:01** Stanford's authors took out the hard parts of real portals: multi factor sign in, human verification puzzles, and session timeouts.

**3:10** The portals are static, with no layout drift or changing payer rules, and each task runs once.

**3:17** The authors warn that premature deployment risks operational errors with financial or clinical consequences.

**3:24** OpenAI's 99 percent safety figure covers Astra, on clinician inbox tasks.

**3:30** It does not cover Luna, and it does not cover portal work.

**3:35** And OpenAI's own researcher told Fierce Healthcare the model might overreach, and take an action without the proper sign-off.

**3:43** So what is the honest way to use this? CVS Health, the customer OpenAI names, built its open source kit differently.

**3:52** An agent discovers a workflow once. The working steps are then compiled into ordinary Playwright code.

**3:59** The README says asking a model to reason through the same steps on every run adds inference cost and variability.

**4:08** The compiled workflow runs without model calls or an API key.

**4:14** So here is EP's take. The price is real, and low enough to try a portal chore that costs a person minutes.

**4:22** But 91.8 percent is a partial credit score on a simulated test, so do not plan around it.

**4:29** Before you buy anything, pick one repetitive portal task, and run it 20 times on your own system.

**4:36** Count only the runs where every field is right, and a check outside the screen agrees.

**4:42** Where the steps never change, do what OpenAI's own customer did: let the agent learn it once, and turn it into plain code.

**4:50** And keep a person on anything that cannot be undone.

**4:53** The sources are on screen and in the description. Every number here is the source's own.

</Transcript>

## Sources

- [OpenAI, Advancing computer-use agents for healthcare teams, Nov 16, 2026](https://openai.com/index/advancing-computer-use-for-healthcare-teams/)
- [OpenAI API pricing (current docs page, no on-page date, captured Nov 16, 2026)](https://developers.openai.com/api/docs/pricing)
- [Stanford, HealthAdminBench (arXiv 2604.09937), Apr 10, 2026](https://arxiv.org/abs/2604.09937)
- [HealthAdminBench site (captured Nov 16, 2026)](https://healthadminbench.stanford.edu/)
- [Anthropic, Claude system card, Jun 9, 2026](https://www.anthropic.com/claude-fable-5-mythos-5-system-card)
- [CAQH, 2025 Index, Executive Report, Feb 19, 2026](https://index.caqh.org/hubfs/Index/CAQH_IndexReport_2025_Executive%20Report_v6.pdf)
- [U.S. Bureau of Labor Statistics, Secretaries and Administrative Assistants, Occupational Outlook Handbook, May 2025 wages; typeset card in the video, bls.gov blocks automated capture](https://www.bls.gov/ooh/office-and-administrative-support/secretaries-and-administrative-assistants.htm)
- [CVS Health, cua-agent-kit README, latest commit Nov 12, 2026](https://github.com/cvs-health/cua-agent-kit)
- [Fierce Healthcare, CVS Health, Cigna and Ambience collaborating with OpenAI to test computer use, Nov 16, 2026 (outside report; typeset card in the video)](https://www.fiercehealthcare.com/ai-and-machine-learning/hlth26-cvs-health-cigna-and-ambience-collaborating-openai-test-computer-use)
