OpenAI prices healthcare back-office work at 8.5 cents a task, undercutting Ambience Healthcare
OpenAI's new healthcare agents run prior-authorization work at 8.5 cents a task, and Menlo Ventures says the scribe market's new dollars are going to Microsoft and Epic. Ambience's $37,500-per-doctor receipt is now the only thing it sells that the model layer does not.
Vincent Jiangtwo-minute read
A machine opened an insurer's website, read a patient's chart, filled in a prior-authorization request, and then did the thing that matters: it stopped, flagged the ambiguous record, and waited for a clinician's approval, logging every click on the way. That agent belongs to Ambience Healthcare~$1.25B — Ambience Healthcare, private, latest valuation ~$1.25B, the San Francisco company that drafts clinical notes for health systems, and OpenAI$1.18T — OpenAI, private, latest valuation $1.18T put it on stage on November 16 as one of four deployment partners for its new healthcare computer-use agents, alongside CVS Health−3.40% — CVS Health, down 3.40 percent today, Cigna−1.15% — Cigna, down 1.15 percent today Healthcare and Abridge$6.69B — Abridge, private, latest valuation $6.69B 1.
The back office just got a price tag
The number OpenAI attached is the story. On HealthAdminBench, a Stanford-built benchmark of prior authorizations, claim appeals and equipment orders, GPT-6 Luna scored 91.8% of available rubric points at 8.5 cents per task, 98% less than its larger sibling GPT-6 Astra, which scored 94.3% 14. Administrative work that vendors sell as software suites is now priced like electricity.
Luna trails Astra by 2.5 points on admin work and costs 98% less
- GPT-6 Astra
- GPT-6 Luna
Data
| GPT-6 Astra | GPT-6 Luna | |
|---|---|---|
| Clinical safety (325 tasks) | 99% | — |
| Physician helpfulness (646 reviews) | 93% | — |
| HealthAdminBench (prior auth) | 94.3% | 91.8% |
That lands on Ambience at an awkward moment. Its validated edge is a receipt, not a demo: a KLAS study published November 3 found about $37,500 of extra annual productivity per emergency physician at Memorial Hermann 8. But the prior-auth workflow it demoed this week is administrative automation, and OpenAI's same launch put Ambience's rivals Assort Health and Infinitus Relay inside ChatGPT for Healthcare as plugins, next to Availity, Stedi, TriZetto and UpToDate 23.
The money is moving, and not to the scribes
Menlo Ventures, reporting the same day, put healthcare AI spending at $3.7 billion in 2026, 2.5 times the prior year, with ambient scribing at $950 million 6. Established vendors took 68% of new documentation spend, Microsoft−1.38% — Microsoft, down 1.38 percent today Dragon Copilot and Epic's Art lifting their share from 33% to 46% 6. The listed winners are Microsoft, Epic and the model layer itself.
Microsoft and Epic now take 46% of new documentation spend, up from 33%
- Microsoft + Epic46%46%
- Other established vendors22%22%
- All other vendors32%32%
Data
| Slice | Value | Share |
|---|---|---|
| Microsoft + Epic | 46% | 46% |
| Other established vendors | 22% | 22% |
| All other vendors | 32% | 32% |
The honest caveat: these are benchmark scores, not clinical outcomes, and HealthAdminBench counts full success only when every subtask passes; OpenAI's own health AI chief has acknowledged agents can still overreach 5. Menlo's survey also carries a conflict, as the firm invests in companies it ranks 6.
What to watch is whether Ambience prices the audited workflow and its outcome contracts rather than the task, because the task now costs 8.5 cents 18. Whoever owns the receipt owns the margin 18. Deep dive in the sources below for the full picture.
For earlier coverage of OpenAI's healthcare push, see our October 29 piece on Anima.
Get Health by email, free
AI in medicine, pharma and biotech. In our daily email, every weekday.






Reader comments
0 comments