The Korean AI startup VIDRAFT has launched the "Open Discovery Challenge" on Hugging Face, a generative benchmark designed to move beyond static prediction tasks by tasking participants with proposing novel drug candidate molecules. Unlike traditional datasets, this competition allows developers to use any frontier large language model (LLM)—including Claude, GPT-4, Gemini, and custom proprietary systems—to explore chemical space for specific disease targets. Participants submit candidate molecules which are evaluated using a composite metric encompassing simulated target-binding affinity, drug efficacy, toxicity, and ADME (Absorption, Distribution, Metabolism, Excretion) properties. Early data from the first three days of the competition, which saw over 2,000 submissions, indicates a performance advantage for Anthropic's Claude models over OpenAI's models across two disease tracks. In the Season 1 Malaria track, Claude-family submissions achieved a median score of 43.7 compared to 31.7 for the OpenAI family. This trend persisted in the Season 2 Tuberculosis track, where Claude-family models scored 39.9 against OpenAI's 30.9. While these figures suggest a directional signal regarding model capabilities in molecular reasoning, VIDRAFT notes they are early-stage results with small sample sizes and undisclosed specific model versions or exact scoring weights.