OpenAI's Healthcare Push: GPT-5.5 Instant Outperforms Physicians in HealthBench, Consumer Tool Faces Scrutiny
核心洞察
OpenAI (搜索)'s latest free model, GPT-5.5 Instant, scored better than physician-written answers on HealthBench, a benchmark developed with over 260 physicians to evaluate AI health capabilities.
More than 230 million people globally use ChatGPT for health and wellness queries weekly, making healthcare one of OpenAI (搜索)'s most important verticals according to health strategy lead Nate Gross.
A Nature Medicine study of ChatGPT Health (搜索) found it failed to recommend hospital visits when medically necessary in over half of cases, though OpenAI (搜索) disputed the methodology.
OpenAI (搜索) is making an aggressive and multifaceted push into healthcare, betting that its latest AI models can transform how consumers seek medical advice, how clinicians practice, and how health systems operate. The effort, which company leadership describes as one of its most important verticals, comes as more than 230 million people globally turn to ChatGPT for health and wellness queries each week.
The company's latest free model, GPT-5.5 Instant, has achieved a notable milestone: it scored better than physician-written answers on HealthBench, OpenAI (搜索)'s proprietary benchmark developed in collaboration with a network of more than 260 physicians. The benchmark evaluates model performance across 48,000 criteria, including escalating emergencies and expressing uncertainty when required.
"Once you know how to evaluate it, it becomes a lot easier to improve it," said Karan Singhal, OpenAI (搜索)'s head of health AI, in an interview with Business Insider. Singhal, who previously helped develop Google's Med-PaLM models, joined OpenAI in mid-2024 and has since built a dedicated health research team.
A Top Priority for OpenAI (搜索)
Healthcare represents approximately 18% of the entire U.S. economy, and OpenAI (搜索) has moved swiftly to capture a share. In January, the company announced that eight major health systems, including Cedars-Sinai Medical Center and HCA Healthcare, are now customers of its enterprise-grade healthcare tools. The company has rolled out three new healthcare-focused products in the past six months alone: a new version of ChatGPT for clinicians, "ChatGPT Health (搜索)" — a tab within the main app allowing consumers to securely connect medical records and wellness apps — and tools powering other health companies' clinical note generation and lab result interpretation.
"It is one of our most important verticals at OpenAI (搜索)," said Nate Gross, who leads OpenAI's healthcare strategy. Gross, an MD-MBA who previously co-founded Doximity, added: "Everyone experiences health issues and this is an opportunity to help everyone."
OpenAI (搜索) has also assembled a network of more than 260 physicians to improve ChatGPT's health advice. These doctors rate responses, discuss specific cases, and train models to recognize when they lack sufficient information. The team has reviewed more than 700,000 example responses to improve accuracy for both consumers and clinicians.
"When I started, I did not think that it was very good," said Rebecca Soskin Hicks, a Stanford-trained pediatrician who now leads OpenAI (搜索)'s physician network. "Frankly, it wasn't that hard to get it to mess up."
Clinical Impact and Consumer Stories
For some users, ChatGPT has proven transformative. Lauren Bannon, a 42-year-old marketing agency founder in North Carolina, credits the tool with saving her life. After being diagnosed with a form of rheumatoid arthritis (搜索) that wouldn't appear on blood tests, she turned to ChatGPT, which suggested Hashimoto's disease (搜索). An ultrasound confirmed the condition and revealed two small lumps — Bannon had aggressive thyroid cancer (搜索). Surgeons removed her thyroid gland and two affected lymph nodes in January 2025.
"I honestly do believe, and I don't just say this flippantly, ChatGPT literally saved my life," Bannon said.
On the enterprise side, Florida-based Advent Health, which operates more than 50 hospitals, reported 80% time savings using OpenAI (搜索)'s models for administrative tasks. Memorial Sloan Kettering Cancer Center has rolled out a test pilot to 1,000 people. "The early signals are great," said Ophelia Chiu, MSK's VP of strategic innovation, while cautioning that "healthcare organizations are so early on in their learning curve despite all the marketing hype over the past 36 months."
Concerns and Limitations
The technology's rapid adoption has not been without controversy. A study published in Nature Medicine in February examining ChatGPT Health (搜索) found that the tool did not recommend a hospital visit when medically necessary in more than half of cases. It also missed patients who were suicidal if they identified a means of self-harm. For non-urgent cases, 65% were sent for care they didn't need.
OpenAI (搜索) published a detailed rebuttal, with Gross stating that "there were some pretty flawed elements to the methodology and it was using an older version of the model."
Clinicians report mixed experiences. Dr. Jinsy Andrews, a neurologist at NYU Langone specializing in ALS (搜索), noted that while ChatGPT sometimes helps patients seek earlier diagnosis, it can also create complications. One patient feared having ALS because ChatGPT suggested it as a cause for muscle twitching — it wasn't. Another requested experimental therapies that ChatGPT referenced from fabricated research.
"Trying to undo or untangle the information from these searches takes away from the time to delve into the actual problem and talk about a treatment plan," Andrews said.
OpenAI (搜索) has reportedly received subpoenas from several states regarding its handling of health data and safety. A Canadian woman sued the company over her daughter's death by suicide, alleging ChatGPT encouraged it. OpenAI has said it takes such concerns seriously and will respond constructively.
Partnerships and the Path Forward
In May, Labcorp (搜索), the $23 billion diagnostics giant, partnered with OpenAI (搜索) to launch an AI-powered mobile app helping consumers understand lab results and track trends. "We ask physicians who have tried to break it, and the accuracy is extremely high," said Bola Oyegunwa, Labcorp's chief information and technology officer.
Startups like Abridge, which uses AI to generate clinical notes from patient-doctor conversations, rely on OpenAI (搜索)'s models alongside those from Anthropic. Davis Liang, who leads Abridge's machine learning team, noted that OpenAI's models excel at notetaking but struggle with high-specificity tasks like generating precise diagnostic codes. "OpenAI will say this patient has diabetes mellitus (搜索). Actually the patient has diabetes mellitus with myopic retinopathy," he said.
OpenAI (搜索)'s GPT-5 model family is the first to be trained specifically at every stage of development to improve health advice quality. Comparing billions of anonymized health messages, the company reported a 71% drop over two months in responses flagged for inaccuracy.
"There's no magic bullet to 20% of GDP," Gross said. "We just have to make sure our models can be helpful to people."
The push comes at a critical juncture for OpenAI (搜索). The company brought in $13 billion in revenue in 2025 but saw net losses balloon to $39 billion. It filed confidentially to go public in June, though reports suggest the IPO may be delayed until next year. Healthcare is clearly central to the company's future — and its ability to convince both consumers and enterprises of its models' reliability will determine whether that bet pays off.