OpenAI says computer-use agents can now carry MRI prior authorizations, EHR inbox work, and payer-portal steps across real healthcare software — with physician reviewers rating 99% of evaluated clinical tasks as safe. The November 16, 2026 healthcare computer-use post and the companion ChatGPT for Healthcare update package benchmarks, customer pilots, plugin expansions, and HIPAA deployment requirements in one launch window.
For builders, the headline is not "agents exist." It is measured safety and cost on workflows hospitals already pay humans to click through, while Google's Gemini for Healthcare preview pursues a parallel industry-agent strategy.
TL;DR
| Metric | What OpenAI reported |
|---|---|
| Clinical safety (EHR inbox) | 99% of 325 reviewed tasks rated safe |
| Helpfulness | 93% very or extremely helpful (646 ratings) |
| HealthAdminBench (admin) | 94.3% rubric (GPT-6 Astra); 91.8% (GPT-6 Luna) at ~8.5¢/task |
| CVS production | Cleared months-long document backlog over a weekend |
| Plugins | 20+ healthcare partners in ChatGPT for Healthcare |
What computer use means in a hospital stack
An outlined hand lowers a cream tile into a compartmented tray, illustrating computer-use agents completing structured healthcare UI steps
Healthcare IT is a maze of EHRs, payer portals, scheduling tools, and fax-era handoffs. OpenAI defines computer-use agents as models that click, type, and switch applications the way a human admin would.
The canonical story in OpenAI's post is an MRI prior authorization: clinical facts live in the EHR, but the insurer wants them retyped in a separate portal. OpenAI's demo (on synthetic patient data) shows an agent drafting the order, encounter note, and patient reply for clinician review, then preparing the payer submission for a specialist.
That is the same API gap ChatGPT's Epic EHR integration started to close — but aimed at UI surfaces vendors never expose.
Benchmarks: clinical inbox and admin rubrics
EHR inbox clinical tasks
OpenAI evaluated GPT-6 Astra on multi-step EHR inbox workflows shaped by physician input. Tasks mixed longitudinal clinical reasoning with complex UI navigation.
Physicians judged clinical safety: could errors plausibly harm patients? 99% of 325 reviewed tasks passed. On helpfulness, reviewers chose very helpful or extremely helpful on 93% of 646 assessments.
Those numbers are OpenAI's study, not FDA clearance. Treat them as vendor-reported usability and safety screening, then run your own eval on your order sets and policies.
HealthAdminBench for administrative work
For prior authorizations, appeals, and DME orders, OpenAI cites HealthAdminBench from Stanford researchers.
| Model | Rubric score | Cost note (per OpenAI) |
|---|---|---|
| GPT-6 Astra | 94.3% | Higher capability tier |
| GPT-6 Luna | 91.8% | ~98% cheaper; ~8.5¢ per task |
Both models ran at maximum reasoning effort in OpenAI's tests. The Luna row is the cost headline for high-volume admin automation.
Customer deployments OpenAI named
CVS Health
CVS uses computer use for steps APIs miss: document sending, referrals, and specialist consult workflows.
OpenAI claims one agent cleared a months-long document backlog over a weekend. Another handles tens of thousands of consult-related system steps per day, with validation, human oversight, screenshots, and action logs.
CVS is open sourcing Agent Kit, built with OpenAI's Agents SDK, to package repeatable healthcare agent patterns. Data stays in CVS-controlled infrastructure and is not used to train foundation models, per OpenAI.
Cigna Healthcare
Cigna care managers previously jumped across 10+ systems for member context. OpenAI says computer use can aggregate context in seconds, citing CIO Madhu Nutakki on an eight-week path from concept to deployment.
Abridge and Ambience
Abridge plans inpatient workflows that merge conversations, EHR facts, and staff context for discharge barriers. Ambience demonstrates prior-auth prep with stop-and-ask behavior when charts are ambiguous — a pattern security-minded teams should copy for any clinical agent.
These sit beside Apple's OpenEvidence health study and Google's connector strategy as competing enterprise_health agent stacks.
ChatGPT for Healthcare: plugins and clinical search
The second post adds three pillars:
- Computer-use access in governed workspaces (same underlying capability).
- Improved trusted clinical search with better citations — blind physician reviews across 1,300+ assessments reportedly rose on eight dimensions, including factual correctness and source authoritativeness.
- 20+ healthcare plugins from partners such as Redox, Stedi, Turquoise Health, UpToDate Expert AI, Metriport, Luma Health, TriZetto, and Oracle HDI.
OpenAI illustrates multi-plugin prompts: Luma plus TriZetto for appointment confirmation and eligibility, Metriport plus UpToDate for post-discharge medication monitoring, Turquoise plus Stedi for MRI site cost comparison.
That design mirrors MCP-style orchestration but inside a HIPAA-eligible ChatGPT contract lane.
Plugin roster: what each partner adds
OpenAI's second post names more than 20 healthcare plugins. Grouped by job, they cover:
| Lane | Partners named by OpenAI | Typical task |
|---|---|---|
| EHR and chart access | Elation Health, Oracle HDI, Metriport, PenguinAI | Pull structured records or summaries into chat |
| Scheduling and access | Luma Health, Ferry Health, Amion, Assort Health | Confirm visits, close referral loops, route staff |
| Payer and revenue | Stedi, Turquoise, Availity Extend, TriZetto, Subsalt | Eligibility, benefits, negotiated rates |
| Clinical knowledge | UpToDate Expert AI, IMO Health, KMED | Evidence lookup with citations |
| Data pipes | Redox, Notable, Photon | Integration middleware between SaaS systems |
ChartGPT — OpenAI's EHR-embedded ChatGPT surface — is where these plugins are meant to run without switching browsers. If your health system already standardized on Redox or Stedi, the launch is an integration announcement as much as a model upgrade.
Cost modeling for admin agents
OpenAI's 8.5 cents per task figure for GPT-6 Luna on HealthAdminBench is the hook for finance teams. Rough math for planning (not a quote):
- 10,000 prior auths per month at 8.5¢ ≈ $850 in model spend before platform fees, humans in the loop, and validation labor.
- The same volume at frontier Astra pricing would cost more per task but might reduce rework if error rates differ.
Run a two-week pilot on synthetic or de-identified queues before you multiply headlines across your real denial volume. Compare against Baseten's efficient frontier inference guidance if you self-host orchestration.
Voice and multimodal entry points
The ChatGPT for Healthcare update shows Voice starting an MRI follow-up that hands off to computer use. That matters for clinicians who will not type long prompts between rooms.
Voice plus computer use increases audit surface: every spoken command should map to logged UI actions. Align retention with your BAA and state privacy rules before enabling voice on shared workstations. Pilot voice on a single ward with explicit patient-consent signage before hospital-wide rollout.
HIPAA, BAAs, and audit expectations
A green mechanical gripper works inside a transparent dome, illustrating governed sandbox access for healthcare computer-use agents
OpenAI repeats deployment guardrails:
- Role-based access and audit logs in ChatGPT for Healthcare.
- Configurable approvals and application allowlists per workflow.
- BAA-covered services only when PHI is involved, plus retention and security settings on connected systems.
- No training on organization data for foundation models.
Teams evaluating agents should treat computer use as privileged automation: it can read screens humans see, including messages and financial identifiers.
How to try it
| Audience | OpenAI's path |
|---|---|
| Health systems | Contact OpenAI sales; enable computer use in ChatGPT Work or Codex |
| U.S. clinicians | Free ChatGPT for Clinicians account with computer use where enabled |
| Developers | OpenAI API computer-use guide for custom healthcare apps |
Existing ChatGPT Epic integration customers should ask whether computer use covers payer portals their Epic APIs do not.
Search and procurement FAQs
Is ChatGPT for Healthcare the same as ChatGPT Enterprise? OpenAI markets a governed healthcare workspace with BAAs, role-based access, and healthcare plugins. Confirm SKU names on the contract you sign.
Can we buy through AWS Marketplace? OpenAI's healthcare post says teams can purchase via AWS Marketplace private offer in addition to direct sales — useful if your cloud procurement rail already runs through Amazon.
Do plugins require computer use? No. The November 16 plugin expansion works in conversational workflows even when UI automation stays disabled by policy.
Building your own healthcare agent: checklist
Teams copying the CVS or Ambience pattern should verify:
- BAA scope covers every API, plugin, and computer-use environment that touches PHI.
- Allowlisted applications — agents should not browse the open web from clinical workstations.
- Screenshot and action logs retained long enough for your compliance calendar.
- Stop-on-ambiguity rules like Ambience's prior-auth demo — never auto-submit when the chart conflicts.
- Separate eval sets for clinical versus admin workflows; OpenAI's 99% safety figure applies to physician-reviewed inbox tasks, not every order type you run.
What this does not solve
- Liability and licensure when an agent drafts orders — human sign-off remains mandatory in OpenAI's demos.
- Payer denial appeals at scale — benchmarks are rubric scores, not appeal win rates.
- Cross-border deployment — announcements emphasize U.S. clinician programs and HIPAA framing.
Related reading
- Google Gemini for Healthcare preview
- ChatGPT Healthcare Epic EHR integration
- OpenAI Ironclad contracting agents computer use
- Apple health AI usage via OpenEvidence
- MCP connectors for health data in chat tools
- ChatMate remote prompt execution on M365 Copilot
Figures and partner names reflect OpenAI's November 16, 2026 healthcare announcements; verify availability and BAA terms on your contract.
