An AI agent called ATHENA-R1 can reason through treatment decisions across every FDA-approved drug since 1939 — and it does so more accurately than current frontier models.
Researchers introduced ATHENA-R1 as a reinforcement-learning-trained agent that works iteratively: it identifies what information it's missing, selects from a library of 212 biomedical tools, runs them, and incorporates the results before reaching a conclusion. The team built it without human-annotated training traces, instead using a two-level self-learning framework where multi-agent systems generated the tasks and reasoning paths, then reinforcement learning rewarded evidence quality and logical consistency. Across five benchmarks covering 3,168 drug reasoning tasks and 456 patient cases, ATHENA-R1 hit 94.7% accuracy on open-ended drug reasoning and 82.9% on treatment reasoning — 17.8 and 10.7 points above GPT-5, respectively. Physicians rated it favorably on complex cardiovascular and infectious-disease cases, and experts from 28 rare disease organizations preferred it over reference models on every criterion.
Treatment reasoning is one of the harder problems in clinical AI because it requires weighing contraindications, comorbidities, and evolving evidence before committing to an answer — exactly the kind of multi-step, verifiable logic that standard language models struggle with. The adverse-event hypotheses ATHENA-R1 generated were validated in electronic health records from 5.4 million patients, producing adjusted odds ratios of 1.48 to 1.84 with no signal in negative controls, which is a meaningful real-world check that most benchmark papers skip.
The 17.8-point gap over GPT-5 is striking, though it reflects a purpose-built tool-use system against a general-purpose model — a comparison that flatters ATHENA-R1 more than a head-to-head with a clinically fine-tuned rival would.