A model 34 times cheaper beats last month's best on health

1. Floor and ceiling — ▶ 2:06

   More than 300 million people use ChatGPT weekly for health and wellness, and the vast majority are free users, most outside the US. OpenAI treats health as a concrete way to "make upside happen": raise the floor with trustworthy, abundant health intelligence for free or nearly free, and raise the ceiling with things not possible before. Usage clusters between visits. Specialist waits run weeks or months, then a 15-minute slot leaves patients with "six more questions". The bet is that AI fills the gaps between appointments.

2. Specialized training, general model — ▶ 6:20

   Healthcare data, including data from physicians across the globe, is added at every stage of training for every model shipped in ChatGPT over the last year. Singhal built specialized models in a previous role. He says a forked, smaller healthcare-only model would "fall behind the frontier relatively quickly" because general improvement moves so fast. OpenAI instead hitches a ride on scale while integrating its own evals and training data. The result is a model that gets both frontier reasoning and medical specialization.

3. Safety net, not middleman — ▶ 10:52

   In Greg Brockman's story, a patient double-checked a drug in ChatGPT, spotted an interaction with her medical history, and her doctor admitted he lacked time to review it. Singhal wants that safety net for patients and clinicians without interference. OpenAI's Penda Health study deployed a clinician copilot that proactively flagged things like medication interactions. The patient-side version is proactive: if a diabetic's A1C rises, it flags a medication review. He is explicit that this is no replacement for having a doctor.

4. Legacy systems suit agents — ▶ 13:09

   Healthcare should not expect an MCP connector everywhere tomorrow. It runs on legacy software, and many primary EHRs run on a virtual machine in the cloud. Singhal sees that as a quirk that actually makes things easier for agents using computer use, and calls the sector "absolutely ripe for agents and computer use". Fragmentation and disconnected data are the main barriers between patients and care. Agents could handle scheduling, insurance coverage and billing errors across that disjointed system.

5. 5 physicians, 8 answers — ▶ 16:45

   Asking 5 physicians a question yields 8 different answers, so OpenAI did not leave uncertain cases to a few in-house clinicians. It built a global network of hundreds of experts across dozens of specialties, languages and countries, and aggregates their views into evaluations and training data. When expert consensus collides with a user wanting comfort, it is "a prioritization of safety": evals track cases where the model must say to seek care even if unwelcome. Mental health uses the same method, with even more disagreement.

6. The 5-17 year lag — ▶ 20:19

   Medical consensus flows from papers to annual clinical guidelines, which work like flowcharts, to actual practice. The last step is the massive bottleneck. Studies put guideline penetration into everyday practice at 5 to 17 years. Singhal wants to compress that to weeks, days or even hours by putting the latest literature in front of clinicians during a 15-minute visit. Where science is truly unsettled, models are trained to recognize their own uncertainty and hedge, and HealthBench includes evaluations of that hedging.

7. Lowering activation energy — ▶ 23:31

   A good model fails if the activation energy to use it is too high. Singhal's three levers are better models, the right harness or product, and deployment through partnerships. OpenAI spent roughly a year, starting about 2 years ago, on security, HIPAA compliance and privacy protections so health systems could trust an enterprise offering. Epic now works inside ChatGPT, and ChatGPT sits as a sidebar inside Epic's workflow. The free tier was brought to frontier performance so the same quality reaches everyone.

8. Quarterback, not care provider — ▶ 30:56

   Medical care is a bundle of diagnosis, empathy and legal liability, and Singhal says AI will not provide all of it. No single company will deliver all care, so OpenAI wants the last mile to look like "the last 100 meters" for partners. It will act as "quarterback" for navigation and scheduling. AdventHealth reports roughly 80% lower administrative workflow costs. He expects visits to shift toward empathy, bedside manner, physical examination and history-taking, and expects both sides to feel safer with AI present.

9. Searching for Move 37 — ▶ 38:19

   Singhal's team wants a "Move 37" for health: a prediction humans would not make that proves brilliant in hindsight. He admits it is very early. The Penda Health study offers a small case. Clinicians were prone to overusing antibiotics, and the model repeatedly flagged it, weighing global resistance over the local case. The larger bet is models reading a 10-year longitudinal history no human can hold in their head, and agents mining EHR and claims data that sits unused in boxes.

10. 34x cheaper, spend unchanged — ▶ 44:06

    OpenAI's cheapest frontier API model beat the best model from a month earlier on HealthBench Professional, a benchmark of difficult clinician workflows, while being 34 times cheaper. The bet is that cost at a given level of health intelligence plummets. But Singhal says it has not yet had a macro effect on the $5.3 trillion in US spend. That needs integration and deployment in documentation, back-office and chronic care. The provider-payer bot arms race is short-term pain, to be replaced by a shared ground-truth handshake.

11. Prevention, robotics, drugs — ▶ 50:36

    For procedural care, Singhal expects the biggest macro factor in the next year or two to be proactive prevention that avoids expensive procedures later. Robotics and care is longer-term, "not a next few years kind of thing", given technology and reimbursement hurdles. Over 5 to 10 years he expects the biggest health impact from AI-guided generation of targets and molecules, and AI-guided prioritization and validation. Candidate drugs already outnumber clinical trial slots, and he predicts validation, not ideas, will be the bottleneck.

Abundant intelligence does not, on its own, shrink a healthcare bill. Singhal is blunt about this, and it is the least glamorous and most useful part of his argument. The price of capability is collapsing. OpenAI's cheapest frontier model on the API beat the best model the company had a month earlier on HealthBench Professional, a benchmark built from difficult clinician workflows, and it cost 34 times less. The bet is that any fixed level of health intelligence gets cheaper on a steep curve, and that the level clinicians and patients will trust is arriving now. Yet he says that intelligence has not had a macro effect on the $5.3 trillion US spend. Compression, if it comes, comes from integration, workflows and deployment in administration, documentation, back-office work and chronic care management. Cheap intelligence is a prerequisite, not the cure.

The foundation for the bet is an architecture choice. OpenAI does not build a separate medical model. It adds specialized healthcare data, including data from physicians around the world, at every stage of training of its general models. Singhal built specialized models in an earlier role and says the forked version would "fall behind the frontier relatively quickly". A healthcare-only model has to keep pace with a frontier that moves very fast. A general model with a health layer rides that curve for free. The same logic drives distribution. Most of the 300 million weekly health users are on the free tier, and most are outside the US.

Capability without trust goes unused, so the next question is how anyone decides what a good answer is. Medicine offers no clean ground truth. Ask 5 physicians a question and you get 8 answers. A handful of in-house clinicians adjudicating edge cases would bake in their quirks. OpenAI built a network of hundreds of physicians across dozens of specialties, languages and countries, and aggregates them into evaluations and training data. The same method now runs in mental health, where licensed experts disagree even more. When consensus collides with what a user wants to hear, the rule is "a prioritization of safety". Evals track the cases where the model must say to seek care even if the user does not want to hear it. The stated aim is a shared ruler for the whole industry, not a private one.

Consensus has a second problem: it arrives late. Papers feed annual clinical guidelines, which work like flowcharts, which then feed actual practice. The last step is the choke point. Studies Singhal cites put the time for guidelines to reach everyday practice at 5 to 17 years. His target is weeks, days or even hours, by putting the latest literature in front of the clinician during a 15-minute visit. Where the science is genuinely unsettled, the answer is calibrated hedging. Models are trained to recognize their own uncertainty and say so, and HealthBench includes hedging evals. The Penda Health study shows the pattern live. A clinician copilot flagged medication interactions as a safety net. It also repeatedly flagged antibiotic overuse, pushing clinicians to weigh global resistance over the local case.

Then comes the barrier Singhal keeps returning to: activation energy. A good model is useless if using it is hard. That is why OpenAI spent roughly a year, starting about 2 years ago, on security, HIPAA compliance and the privacy protections health systems demand. It is why ChatGPT now sits inside Epic as a sidebar, and Epic can be used inside ChatGPT. And it is why he is relaxed about legacy software. Healthcare will not have an MCP connector tomorrow, and some primary EHRs run on a virtual machine in the cloud. He treats that as an opening, not a wall: it is "absolutely ripe for agents and computer use". AdventHealth reports roughly 80% lower administrative workflow costs from this kind of deployment.

The boundary he draws around OpenAI's role is firm. Health information and medical care are different products. Care is a bundle of diagnosis, triage, empathy and legal liability. Singhal says AI cannot and will not provide all of it, and no single company will deliver all medical care. OpenAI wants to be the quarterback for navigation, scheduling and logistics, and to make the last mile look like "the last 100 meters" for partners who deliver care. Even if patient and clinician both arrive briefed by AI, he expects the visit to be about empathy, bedside manner, physical examination and history-taking. Robotics and procedural care are not a next-few-years story. The nearer lever is prevention: flag the rising A1C early and avoid the expensive procedure later.

Above the floor sits the ceiling, and here his language turns speculative. His team wants a "Move 37" for health, a prediction no human would make that looks brilliant in hindsight. He concedes it is very early. The candidate path is models reading a 10-year longitudinal history no human can hold in their head, and agents mining records and claims data that sit unused. Over 5 to 10 years he expects the impact to come from AI-guided generation of targets and molecules, and from AI-guided prioritization and validation. Candidate drugs already outnumber trial slots. Validation, he predicts, becomes the bottleneck rather than ideas.

The obvious objection is the one the host raised. If every provider and payer deploys billing and paperwork bots, spending can rise. A Blue Cross Blue Shield Association analysis, as the host described it, found those bots added about a billion dollars. Singhal calls it short-term pain and a distraction. His ideal is a handshake. Ambient transcription and AI-reviewed notes give both sides a shared ground truth of what happened in a visit, an AI adjudicates, and a limited appeals process remains. That is a design, not a forecast. It assumes payers and providers will accept one source of truth, and he offers no mechanism for how that agreement happens. There is a second gap. The cost evidence comes from OpenAI's own benchmark, and Singhal says the models could still use improvement on it. And the claim that patients and clinicians will both feel safer with AI in the room is a prediction he says is not fully there yet.

His advice to builders is to bet on the models getting better, since many in healthcare are betting against them or staying neutral. He says OpenAI does not want to hoard health intelligence and wants it commoditized. If that works, what stays scarce is integration, workflows and the last 100 meters. Those are the parts he says drive savings, and the parts OpenAI says it will mostly leave to partners. Whether that adds up to a lower bill is still open. Singhal says the compression has not happened yet.