The compact model that beat every frontier general-purpose system on a new aging biology benchmark this week (reported here 2026-09-22) works by a trick worth unpacking now that CRISPR Therapeutics has shown what a validated genetic target can do downstream. CRISPR Therapeutics' CTX310, a one-time edit to the ANGPTL3 gene, cut LDL cholesterol and triglycerides 52% in its Phase 1 trial, a single dose holding for the trial's readout window. That result depended on decades of prior work identifying ANGPTL3 as a lever worth pulling. The aging model's job is to find the next ANGPTL3 before a human researcher does: it was trained on cell measurements (which genes were switched on, how the cell aged, what changed under stress) and learns which combinations of gene activity predict which outcomes, without being told the biology in advance. That is the same shape of task Anthropic's science lab is running at much larger scale: 950 AI agents scanned public DNA for 21 hours and flagged ART, a new bacteriophage enzyme system, purely from patterns in sequence data (reported here 2026-09-23).

Both results sit at benchtop stage, and the gap between them and CTX310 is the whole story. CTX310 has a patient, a dose, and a cholesterol number measured a year out. The aging model and the ART enzyme have a benchmark score and a preprint. A benchmark win means the model ranked correctly against known answers; it does not mean the intervention it flags will actually work in a cell, let alone a person. The next gate for the aging model is wet-lab validation of the specific interventions it flagged, not a higher score on the next version of the benchmark. Until a lab tests one of those flagged targets in living cells and gets a measurable result, the model is a very good guesser, not yet a drug discovery engine.