DIRTY STATIC / DISPATCH News17 NOV 2026 • 7 MIN READ
A large computer screen glowing caution yellow in a brutalist concrete hospital room, showing abstract medical records, with hazard stripes on the floor
NewsINTELLIGENCE FOR THE MACHINE AGE

OpenAI Adds Healthcare Agents, Twenty Plugins

OpenAI published two linked posts on November 16 that put agents inside clinical and payer software, upgrade clinical search in ChatGPT for Healthcare, and add more than twenty plugins from partners across records, scheduling, eligibility, and clinical reference. This report covers what the agents do, what OpenAI says its own evaluations found, and what remains unproven.

Start reading ↓

OpenAI put agents inside healthcare software on November 16. The agents read screens, click, type, and move between applications that care teams already use. They work in electronic health records and in payer portals, the insurer sites where staff file requests and check coverage.

That is a different job from answering a medical question. An answer sits in a chat window for a clinician to judge. An agent that clicks through a record system takes steps in that system. OpenAI is selling the second job now, with controls and audit logs as part of the pitch.

The Announcement

OpenAI published two posts that work as one launch. One covers the agents. The other covers ChatGPT for Healthcare, with stronger clinical search and the plugin expansion.

The agents handle referrals, prior authorization, discharge planning, and outreach between visits, according to OpenAI. Prior authorization is the insurer approval step that often must happen before a scan or a treatment can go ahead. In a demo built on synthetic patient data, an agent reviewed visit notes and records after a patient reported back pain, then drafted an order, a note, and a patient reply, and prepared an authorization request for a specialist to review.

Access is split by audience. Healthcare teams and developers can use the capability through the OpenAI API and ChatGPT for Healthcare. Existing customers enable the agents in ChatGPT Work and Codex through a workspace administrator. Eligible and verified clinicians in the United States can use a free ChatGPT for Clinicians account, with improved search and select plugins.

Important Details

Two models sit behind the work. GPT-6 Astra is the larger model in OpenAI reporting. GPT-6 Luna is the cheaper model aimed at high volume administrative work. OpenAI ran both at maximum reasoning effort in the benchmark work it reported.

ItemWhat OpenAI reported
Launch dateNovember 16, 2026
Where agents are availableOpenAI API, ChatGPT for Healthcare, ChatGPT for Clinicians
Tasks namedReferrals, prior authorization, discharge planning, outreach
Tasks rated safe99 percent of 325 tasks in record inboxes
Helpfulness ratingVery or extremely helpful in 93 percent of 646 assessments
Astra score on HealthAdminBench94.3 percent of rubric points
Luna score on HealthAdminBench91.8 percent of rubric points
Luna cost on that benchmark8.5 cents per task, 98 percent less than Astra
New pluginsMore than twenty, enabled by a workspace administrator

HealthAdminBench is a benchmark from Stanford researchers for common administrative work such as prior authorizations. The scores above are as OpenAI reported them. OpenAI did the reporting, and the figures are not a clearance from a regulator and not a measure of patient outcomes.

HealthAdminBench scores reported by OpenAI, share of rubric points

Astra94.3 percent
Luna91.8 percent

GPT-6 Astra and GPT-6 Luna at maximum reasoning effort, as OpenAI reported on November 16, 2026. Luna cost 8.5 cents per task in that work.

Clinical search changed in the same launch. OpenAI says it improved how the product finds, evaluates, and cites medical evidence. In a blind evaluation, physicians reviewed answers to real questions across specialties. In more than 1,300 reviews, they rated the updated experience higher on all eight measures OpenAI tracked, including factual correctness, safety, and source authoritativeness. Those are physician ratings of answers, again as OpenAI reported them.

The plugins bring outside tools into the same conversation, including through ChartGPT, which places ChatGPT inside the record workflow. OpenAI named Aiva, Amion, Assort Health, Availity Extend, Elation Health, Ferry Health, IMO Health, Infinitus Relay, KMED, Luma Health, Metriport, Notable, Oracle HDI, PenguinAI, Photon, Redox, Stedi, Subsalt, Trizetto, Turquoise, and UpToDate Expert AI.

The examples chain plugins together. Metriport can retrieve a discharge summary from care elsewhere while UpToDate Expert AI supplies guidance on monitoring after discharge. Aiva can open a hospital policy on insulin pumps while Amion finds the covering specialist. IMO Health, Stedi, and Turquoise can match a procedure code for a knee scan, check eligibility, and compare negotiated rates at two hospitals.

Desks in a concrete office with monitors glowing caution yellow showing abstract insurance forms and patient charts, with tall stacks of paper

What The First Partners Are Doing

OpenAI named Abridge, Ambience, Cigna Healthcare, and CVS Health as organisations developing, piloting, or deploying the agents. The detail varies a lot by name, and the strongest claims come from CVS Health.

OpenAI says one CVS Health agent cleared a backlog of document sending that had built up over months, clearing it over a weekend. Another handles system steps behind tens of thousands of specialist consults each day. CVS Health uses validation checks and human oversight, with screenshots and action records, according to OpenAI. CVS Health is also open sourcing an Agent Kit built with the OpenAI Agents SDK.

Cigna Healthcare is using the models to assemble context for member outreach. Case managers once gathered that context from more than ten clinical and administrative systems, including benefits, coverage, providers, and appointments. OpenAI says the agents can now bring that material together in seconds. That time claim is OpenAI account of Cigna work, not an outside measurement published with data.

The other two names are earlier in the work. Abridge is developing agents for discharge barriers in a planned inpatient use, including access to medication and coordination of care at home, with a clinician brought in when a decision needs clinical judgment. Ambience showed an agent that prepares an authorization request, stops when a record is unclear, and waits for approval before it submits.

Governance sits under all of this. For protected health information, organisations must use services covered by a Business Associate Agreement with OpenAI, apply required security and retention settings, and extend those protections to connected systems. ChatGPT for Healthcare offers role based access controls and audit logs, and OpenAI says organisation data there is not used to train its foundation models. Organisations can set which applications an agent may open, add review and approval steps, and pause an agent.

The launch also builds on earlier ground. Coverage of the launch places it after the September 1, 2026 Epic integration, which brought authorised patient context into ChatGPT in supported deployments.

What’s Next

Confirmed steps are narrow. The agents and the plugins are available now, subject to workspace settings. Administrators choose which plugins to enable, and clinicians in the United States can sign up for the free clinician account. OpenAI has invited more healthcare companies to build plugins, and CVS Health Agent Kit gives other builders a starting point that OpenAI says is open source.

What is not established is just as clear. OpenAI has not published a list price for ChatGPT for Healthcare in the posts covered here. It sells the product through its sales team and through private offers on AWS Marketplace, according to coverage of the launch. The evaluation figures are OpenAI own, the partner results are described by OpenAI without underlying data, and no results here measure whether patients did better.

Watch what health systems publish on their own. A system that names a workflow, a review step, and a correction rate is saying something checkable. A logo on a launch post is not. The work these agents aim at is real, repetitive, and expensive, which is why the pitch lands. Whether a clicking agent should do that work inside a live record will be settled in production, under audit, one workflow at a time.

Facts in this report were checked on November 17, 2026 against the two OpenAI posts published on November 16, 2026 and the independent coverage cited below. Evaluation figures and partner results are attributed to OpenAI where they appear.

Sources

■

END OF TRANSMISSION

Stay curious. Question everything.

CONTINUE THE DESCENT

MORE FROM DIRTY STATIC ↘