A new AI system can read a breast cancer patient's tumor slide and clinical chart and return a recurrence-risk score in hours, not the two to six weeks patients currently wait on genomic tests like Oncotype DX. That is the headline out of a study published this year in Nature Communications by researchers at NYU's Center for Data Science, and it has been circulating alongside a specific number: 3,500. Whether that number describes a price tag or a proof point has already blurred in public conversation, and the distinction matters more than it looks.
What the Test Actually Reads
The tool does not sequence anything. It reads two things clinicians already have on file: a digitized hematoxylin-and-eosin pathology slide of the tumor, and routine clinical variables, including tumor stage, patient age, and hormone-receptor and HER2 status. From that combination, the model outputs a recurrence-risk score across hormone-receptor-positive, triple-negative, and HER2-positive breast cancers, using two metrics oncologists already rely on to judge prognostic tools: the C-index, which measures how well a model ranks patients from lowest to highest risk, and the hazard ratio, which compares recurrence rates between risk groups over time.
Krzysztof J. Geras, the NYU researcher who co-authored the study and serves as chief scientific officer at Ataraxis AI, described the result directly: the test matched or outperformed a widely used genomic test, while sparing tumor tissue that a gene-expression assay would otherwise consume, tissue that can matter later if a patient needs additional testing or enters a clinical trial.
"Our AI test matched or outperformed a widely used genomic test... and could deliver answers in hours instead of weeks, at lower cost, while sparing tissue for future testing."
What Is Confirmed, and What Is Not
The validation cohort is real and substantial: more than 3,500 patients across 15 populations in seven countries, published in a peer-reviewed journal. That is a legitimate, multi-country retrospective validation, not a single-hospital pilot. But the study's own authors are explicit that this remains a research-phase finding. Their paper calls for evaluation in completed randomized clinical trials before the model changes practice. That caveat is standard, and important: retrospective validation on banked samples shows that a model correlates with outcomes that already happened. It does not yet show that using the score prospectively changes what oncologists do, or improves what happens to a patient next. Nothing in the reporting on this study describes FDA clearance for this recurrence test specifically. Separately, other AI-driven histopathology tools, including a chemotherapy-benefit tool from a different developer, have moved through FDA clearance pathways in the past year. That shows the regulatory route exists. It has not yet been walked by this test.
The Price Question, and Who Actually Gets Tested
Here is where the $3,500 figure gets slippery. The number attached to this story traces most directly back to the size of the validation cohort, not a disclosed per-test price. Ataraxis AI, the company tied to this research, has said publicly that it intends to price its tests below the roughly $3,000-and-up list price of Oncotype DX, the incumbent genomic test this AI tool is being measured against. But no fixed, company-disclosed sticker price for clinical use appears in the public record as of this writing. That distinction is worth stating outright rather than repeating a number as settled fact.
That ambiguity is a governance problem, not a reporting footnote. Genomic recurrence tests like Oncotype DX took years to win broad insurance coverage, and coverage still varies by payer, plan, and state Medicaid program. A test that is faster and potentially cheaper to run does not automatically inherit that coverage. It has to earn its own billing code, its own payer contracts, and its own evidence base for medical necessity, a process that historically lags the underlying science by years, not months. A test priced at "lower cost" than a $3,000 to $4,000 incumbent could still land anywhere from a few hundred dollars to a few thousand, and where it lands determines who can pay out of pocket while insurers catch up.
The open questions worth tracking as this technology moves from journal page to hospital lab:
- Will Ataraxis AI, or a competing developer, publish an actual list price before or after seeking regulatory clearance, and will that price meaningfully undercut genomic testing for an uninsured or underinsured patient?
- Which payers will require prospective, randomized-trial data before covering an AI recurrence score, and how long will that evidence bar take to clear?
- Will hospital systems serving lower-income and rural populations get access to the digital pathology infrastructure the test requires, including scanners, cloud pipelines, and trained staff, or will "hours instead of weeks" remain a benefit concentrated at academic medical centers that already have the fastest turnaround?
None of this argues against the science. A retrospective validation across 3,500 patients in seven countries, published in a major journal, is a genuine step. But the distance between "this model performs well on banked slides" and "your insurer will pay for this test at your community hospital next year" is where equity gets decided, and it is decided by payer contracts and infrastructure budgets, not by the accuracy of the underlying algorithm. Patients considering any new recurrence test should raise it directly with their oncologist, who can say whether it is available, covered, and appropriate for their specific diagnosis.
