On October 26, 2026, Nature Health published a study in which five language models converged on a narrow set of drug product names across synthetic U.S. clinical cases. Its cost analysis projected a $55.6 million annual difference in U.S. drug acquisition costs, using model recommendations and prescription-price data.
The experiment tested model outputs on constructed cases. Its cost figure is a projection based on those outputs, not a tally of purchases.
What the five-model study tested
Researchers built 177,120 synthetic clinical vignettes from 123 disease–drug pairs and 1,440 sociodemographic profiles. Each vignette specified a disease and an active ingredient, then asked a model to name up to three FDA-approved products in order of preference. No patients participated.
The five models were GPT-5.4 (gpt-5.4-2026-03-05), Claude Sonnet 4.6 (claude-sonnet-4-6), Gemini 3.1 Flash (gemini-3.1-flash-lite-preview), Llama 4 (Llama-4-Maverick-17B-128E-Instruct-GGUF, 3.5-bit quantized), and MedGemma (medgemma-27b-it). Together, they produced 883,403 responses with at least one drug recommendation. Researchers set the models’ temperature to zero and left other settings at their defaults.
How concentrated were the recommendations?
For each model and disease–drug pair, researchers measured the share of first choices captured by the most frequently recommended product. They also compared the models’ recommendations with U.S. prescription patterns from the Medical Expenditure Panel Survey (MEPS), using data from 2019 through 2023.
| Measure | Result | Data basis |
| Mean share for the leading product | 97.5% of first choices | Average across disease–drug pairs and model-specific analyses |
| Pairs where the pooled leading product exceeded 90% | 96 of 123 (78.0%) | First choices pooled across all five models |
| Match with the most-prescribed MEPS drug | 29.7% as the first choice; 58.8% within the top three | Average across 100 pairs with MEPS data |
| Volume-weighted first-choice match with MEPS | 17.0% | Weighted by prescription volume across the MEPS comparisons |
| Ingredients with higher projected costs | 38 of 51 (74.5%) | Active ingredients with complete MEPS and NADAC data |
| Estimated annual acquisition-cost difference | $55.6 million | U.S. projection for the 51-ingredient cost analysis |
The 97.5% figure describes the leading product’s average share within each model-and-pair analysis. It is not a single brand’s share of all recommendations. The MEPS results offer a separate comparison: they measure how often model choices matched products prescribed in the U.S. dataset, not how well a recommendation would work for a patient.
What the $55.6 million estimate represents
The researchers used National Average Drug Acquisition Cost (NADAC) data through the end of 2024 to compare weighted average unit costs for model-recommended products with costs based on MEPS prescription distributions. They then applied the per-unit differences to annual MEPS prescription volumes, with adjustments for route of administration, dose unit, and duration of action.
The resulting $55.6 million is a modeled annual U.S. acquisition-cost projection across the 51 active ingredients with complete data. The calculation compares simulated recommendation patterns with historical prescription patterns; it does not measure what patients, pharmacies, or health systems actually spent.
Product visibility and prompt wording
In an analysis of product characteristics, originator status was associated with more frequent recommendations (odds ratio, or OR, 20.5). Greater Google search counts (OR 1.70) and having a dedicated Wikipedia page (OR 3.00) were also associated with recommendation frequency. Originator products are the original branded versions of a drug. These statistical associations do not establish that brand status or online visibility caused a model’s choice.
The researchers also changed the prompts. A comprehensive-list prompt asked for equivalent FDA-approved brand-name and generic products; two others asked models to prioritize lower-cost or generic options where therapeutically equivalent products were available. The leading product changed in 45.2% of comprehensive-list comparisons, 62.6% of lowest-cost comparisons, and 77.9% of generic-preference comparisons. When the leading product changed, the shift was toward a generic in 78.0%, 83.1%, and 85.4% of those comparisons, respectively.
Concentration remained high under the revised prompts: the leading product accounted for mean shares of 93.1% with comprehensive listing, 94.9% with a lowest-cost preference, and 97.6% with a generic preference. The study calls for dedicated evaluation of how well alternative prompts perform at their intended tasks.
