OpenAI Skips Specialist Medical Models, Betting Instead That General Models Will Keep Getting Stronger
Specialized medical models will quickly be outpaced by cutting-edge general models, so OpenAI chooses instead to bake specialist data from hundreds of doctors into each generation of GPT, rather than building a separate medical-specific model.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Medical AI's real value lies between appointments, not in replacing them
American primary care waits are 40+ days; specialist appointments are weeks to months, and only 15 minutes long. Most people leave with questions unasked. ChatGPT gets 300+ million weekly health questions—mostly not asking to replace doctors, but to fill the gap: understanding lab results, preparing questions, digesting medical advice. What's scarce isn't medical information but the person managing the time between visits.
— Karan SinghalSpecialist Medical Models Get Outpaced—So Build General Ones Instead
Singhal built specialist models before and knows they get rapidly overtaken by general cutting-edge models because industry scales so fast. OpenAI does the reverse: feed specialist data from doctor networks into every training stage of each GPT generation, letting the model capture both the scaling rewards of general models and specialist medical training. The bet is that continuous model improvement beats betting on a standalone medical model.
— Karan SinghalMedical Consensus Is Harder Than It Looks; Ask Five Doctors, Get Eight Answers
OpenAI initially wanted a few doctors to define how the model should handle edge cases—but discovered it doesn't work: the same question to five doctors gets eight different answers. The team built a network spanning dozens of specialties and practicing physicians across countries to aggregate consensus rather than rely on one or two voices. When users want reassurance but the condition is actually dangerous, the team puts safety over user satisfaction.
— Karan SinghalClinical Guidelines Travel Slowly—5 to 17 Years From Paper to Practice
Medical knowledge travels a chain: research paper → clinical guidelines (updated yearly) → how doctors actually practice. The final step typically lags 5–17 years; newest guidelines take time to filter into daily work. AI's value isn't inventing new knowledge but compressing that lag from years to weeks or hours—bringing the latest literature and guidelines directly to the doctor or patient at the moment of care, rather than waiting for knowledge to slowly seep into clinical practice.
— Karan SinghalIntegrating Epic EHR Means Breaking Down Healthcare's Data Silos
A major healthcare barrier is fragmented data and workflows. Many hospitals still run their core EHR on aging virtual machines with no ready AI hooks. OpenAI's partnership with Epic lets doctors call Epic data directly from ChatGPT and embed ChatGPT's sidebar into Epic workflows. UCSF was among the first hospitals to land this integration, serving as proof that fragmentation can be dismantled.
— Karan SinghalCounterintuitive But Correct: AI Challenging Doctors on Antibiotic Overuse
Borrowing from AlphaGo's Move 37 against Lee Sedol—the unintuitive play that later became legendary—Singhal saw a medical parallel in a clinical doctor collaboration with Panda Health: the model repeatedly pushed doctors not to over-prescribe antibiotics for the case in front of them, but to think about antibiotic resistance as a larger system problem. Cases that integrate a decade of longitudinal records and predict what the human brain can't calculate will become increasingly common.
— Karan SinghalNew Models Cost 34 Times Less and Perform Better on Medical Benchmarks
Singhal sets the $5.3 trillion or more healthcare spend (larger than the world's third-largest economy's GDP) against another curve: AI's latest and cheapest API model outperforms what was the strongest model a month ago on OpenAI's hardest clinical benchmark, HealthBench Professional, while costing 34 times less. He bets that once model capability crosses the threshold patients and doctors will trust, costs for that threshold will keep plummeting like they always have.
— Karan SinghalThe Next 5–10 Years: AI's Biggest Health Impact Will Be in Drug Discovery
Asked what looks like science fiction today but will ship in 5–10 years, Singhal didn't point to chatbots but to drug discovery: AI generating new drug targets and molecules, and AI screening and validating clinical trial candidates. Candidate molecules far outnumber trial slots; AI can rank priority and move trials faster and cleaner. More personalized new drugs will likely be AI's single biggest health impact in coming years.
— Karan SinghalIn their own words · checked verbatim
I love this idea of a kind of safety net for both patients and clinicians that is just kind of there in the care journey and not always interfering or being the middleman, but really thinking about the ways in which it can augment the existing experience and build on it.
Karan Singhal11:11
health is a place where, you know, it moves to the speed of trust.
Karan Singhal15:19
So it turns out if you ask five physicians or something, this any question, you'll get like eight different answers.
Karan Singhal17:25
this takes, depending on the study, like five to 17 years for that penetration to actually happen, which is massive.
Karan Singhal21:32
there's no future where, you know, any one company is delivering all the medical care, right?
Karan Singhal31:45
But in hindsight is this kind of like brilliant and insightful move that humans are now just beginning to understand and learn from the patterns of and things like this.
Karan Singhal38:53
it performed better than 5.6 soul, which was literally our best model like a month ago… it performs better despite being 34 times cheaper.
Karan Singhal44:01
Figures
| US annual healthcare spending | $5.3 trillion or more | 43:00 |
| US healthcare workforce | 23 million people | 43:00 |
| ChatGPT weekly health questions | 300+ million users | 1:02 |
| ChatGPT weekly active users | 1.2 billion | 24:32 |
| Americans without a regular primary care doctor | 100+ million | 5:05 |
| Average wait for primary care appointment | 40+ days | 5:05 |
| Lag from new clinical guidelines to routine practice | 5–17 years | 21:32 |
| Latest cheapest model vs. previous strongest model | 34 times cheaper; better performance on HealthBench Professional | 44:01 |
Glossary
- Move 37
- AlphaGo's enigmatic move against Lee Sedol that initially seemed flawed but proved brilliantly insightful—now used to describe AI judgments that defy human intuition yet prove correct.
- HealthBench Professional
- OpenAI's medical model benchmark, co-developed with physicians, measuring performance on high-difficulty clinical scenarios.
- A1C
- Glycated hemoglobin; a measure of a diabetic patient's long-term blood sugar control.
- EHR
- Electronic health record; a hospital system for storing patient medical history and clinical notes.
How to listen
Founders and investors focused on health tech, digital health startups, or hospital AI deployment, plus anyone wanting to understand how AI actually embeds in clinical workflows.
Skip the opening mission statement and closing thanks and disclaimers—low information density.