Has AI cured cancer yet?
Last checked: today
Nudging technocrats towards the more important benchmarks, in a language they already understand.
Tracker result
No.
Status change: none
CB CancerBench
The frontier remains at zero.
CancerBench measures whether an AI system has cured a cancer. It is currently the most difficult benchmark in AI—harder than DeepSWE, Humanity’s Last Exam or SkateBench—with every frontier model scoring 0%. We have even open-sourced the dataset and are encouraging every frontier lab to benchmaxx it.
CancerBench score
Frontier models · pass@1
Frontier score over time
Best score by model availability
CancerBench score vs. model cost
CancerBench score · Artificial Analysis blended price
Benchmark questions Research objectives 80 questions
| ID | Question | Best score |
|---|---|---|
| Loading the open dataset… | ||
What is working
AI updates
New model beats every benchmark
Cancers cured: 0
AI discovers a drug candidate
Years of human trials begin
CEO declares medicine transformed
Clinical evidence pending
Another “breakthrough” press release
Patients treated: not stated
Researcher updates
A computer-designed drug candidate has reached an early human trial. Researchers developed it. Clinicians and trial participants now have to find out whether it is safe or works. The trial is recruiting and has no results.
ClinicalTrials.gov ↗AI-assisted tools may have informed successful treatment for individual patients. Clinicians treated those patients. Remission means the cancer is currently undetectable. A cure means it will never return.
Cure vs remission, NCI ↗Decades of human research have made some cancers routinely curable. Low-stage testicular cancer cure rates approach 100% with established treatments.
Testicular cancer treatment, NCI ↗The NCI sees promise for AI across cancer care. Researchers, clinicians and patients must still establish any clinical benefit through randomized trials.
AI and cancer, NCI ↗Researchers turn candidate molecules into medicine through laboratory work, clinical trials and patient care.
A model may supply a candidate to a human process in which researchers design the studies, clinicians deliver the care and patients take the risk, but the gigafactory gets the headline.