Log inSign Up
V. LazukaCTO
April 23, 2026
There is a test that Stanford built specifically to check whether medical AI can actually be trusted with a patient's chart. Not a trivia quiz. Not a USMLE-style exam. The real thing. Look up a patient's lab results. Decide if a number is dangerous. Order the right follow-up, at the right dose, without breaking anything else the patient is on.
Stanford released the first version in 2025. Then, because the best AI labs in the world were getting it wrong too often, they released a harder second version in 2026.
Before us, no one had scored 100% on either.
Anthropic's Claude Opus 4.5, the model powering Claude for Healthcare, got 92.3% on the first version. OpenAI's GPT-4.1 reached 98.0% on version one with memory and hand-tuned examples, and 88.7% on version two without memory (the highest published v2 result). Google's Gemini and other frontier models, all below perfect. Always a few percent of the time, they got something wrong.
We are the first system, to our knowledge, to reach 100% on either version. And we got it on both with the same engine, no retuning.
If an AI gets 92% of medical tasks right, that sounds fine on a slide. In a hospital it is not fine. Eight times out of a hundred, the AI misses a lab value, confuses one patient's chart with another's, or invents a dose that sounds plausible. And here is the part that makes it dangerous: the wrong answer looks identical to the right ones. There is no flashing light. The nurse, the doctor, the pharmacist, none of them can tell which response happens to be the one that got it wrong.
That is what every medical AI on the market has been asking hospitals to accept. "Usually right. Good luck catching the times we aren't."
Hospitals started using this kind of AI in 2025. In January 2026, both OpenAI and Anthropic launched entire healthcare product lines. Each of them, at deployment, was wrong a few times out of every hundred. Which means somewhere in the world, right now, an AI assistant is reading a patient's blood work and quietly making a mistake that no human is going to catch.
That is the gap we closed.
Most medical AI fails for the same reason. It asks one AI model to do everything at once. Read the chart. Find the right number. Interpret it. Apply the rule. Write the order back to the record. Every single step is a place the AI can hallucinate, miscount, or confuse one value for another. Stack enough of those steps and you get 8% errors.
We broke the job apart. The AI runs the workflow, but the parts that have to be exactly right, pulling a specific lab value, applying a threshold, writing to the medical record, go through a pipeline that cannot hallucinate by design. BloodGPT never invents a number. It never guesses a threshold. If something is missing from the record, it says so instead of making something up.
There is a side effect to building it this way. We do not need the most expensive flagship model to get 100%. We used two lightweight models, Gemini 3 Flash and Claude Haiku 4.5, both costing roughly one-tenth per call of what competitors are using on the same benchmark. Both models scored identically: 100% on every run.
Same test. Higher score. Cheaper.
If you are a patient who has ever uploaded lab results into a tool that claims to use AI, you have something concrete to ask now. Which clinical benchmark did it pass? What score? Was it reproduced?
Because the difference between 92% and 100% in medicine is the difference between "a nurse has to re-check everything it said" and "the nurse can focus on the patients."
For our subscribers, this is the engine that reads your labs. Same architecture. Same guarantees. When you upload a PDF of your blood work, the pipeline that pulls each value, checks it against reference ranges, and tracks it across your history is the exact one that just scored 100% at Stanford. Not a lighter version. The real thing.
For clinics, labs, and pharma partners already in pilot with us: your integration runs on this engine. The 100% score is not a new product. It is a validation of the engine you are already using.
Around the same time, a different team at Stanford, Harvard, and several other institutions published something uncomfortable. They took 100 real patient cases and asked 31 different AI models to recommend what to do. Severe harm happened in up to 22% of cases, depending on the model.
The most striking finding: 77% of the serious harms were not wrong advice. They were missing advice. The AI forgot to order the test that would have caught the problem. It did not flag the medication interaction. It did not escalate the chest pain. And because a confident chatbot sounds the same whether it is complete or not, nobody noticed.
That study, called NOHARM, is the clearest evidence yet that a single general-purpose chatbot should not be the thing between you and a decision about your health. We wrote a longer piece about what it found, what it actually means for anyone uploading medical records into an AI, and why the architecture of the system matters more than the model you choose.
The short version: the one thing that reduced harm in that study was using multiple AI models from different companies, checking each other's work. That is how BloodGPT's clinical reasoning layer is built. GPT-5.1 and Gemini 3 Pro sit on top, reviewing and cross-checking. Not because any one of them is unsafe. Because the evidence now says no single model is safe enough alone.
Full technical write-up on how BloodGPT achieved 100% on MedAgentBench v1 and v2 - is available here.
BloodGPT's review on NOHARM - here.
Want to see it in action?
bloodgpt.com
For developers: docs.bloodgpt.com
Talk to us: [email protected]
About BloodGPT
BloodGPT is a B2B healthcare AI platform that turns unstructured laboratory data into standardised FHIR R4 records with clinical interpretation. We operate in Israel, Nigeria, Poland, the UK, and Greece, with enterprise deployments including white-label, API, and on-premise options.
References
Transform how you deliver lab results to patients while enhancing your practice's efficiency and value. Try BloodGPT or contact us for a personalized demo
Book a Demo