SEOUL, South Korea – Aug. 19, 2026 – Which general-purpose AI is actually better at discovering promising drug candidates — OpenAI’s ChatGPT, Anthropic’s Claude, Google Gemini, or DeepSeek?
That question is now being tested in a real-world scientific competition through the Open Discovery Challenge, an open AI drug-discovery project developed by VIDRAFT and launched publicly on Hugging Face.
More than 2,000 candidate molecules were submitted within just three days of launch, turning the challenge into a rapidly growing public experiment in whether general-purpose AI can move beyond answering questions and contribute to actual scientific discovery.
Unlike conventional AI benchmarks, the Open Discovery Challenge does not ask models to solve a fixed set of questions. Participants use OpenAI, Claude, Gemini, DeepSeek, Qwen, KIMI, or their own AI systems to explore chemical space, design new molecular structures, and submit candidate compounds for evaluation.
Each molecule is automatically assessed across key drug-development dimensions, including predicted efficacy, toxicity, target binding affinity, ADME properties, and simulated preclinical and clinical potential. The resulting scores and rankings are published openly.
In other words, the challenge is not about asking which AI is the smartest, but rather which AI can actually help discover a better drug candidate. Early results are already revealing meaningful differences among the models.
In the malaria season, submissions generated with Claude-family models achieved the highest median score at 43.7, compared with 31.7 for OpenAI-family models, a gap of roughly 12 points.
The same pattern appeared again in the tuberculosis season, where Claude-family submissions recorded a median score of 39.9, compared with 30.9 for OpenAI. Seeing the same ordering across two different disease targets makes the trend particularly noteworthy, although larger datasets will be needed before drawing definitive conclusions about model superiority.
Another notable result is the performance of Chinese open and openly accessible AI models, including DeepSeek, Qwen and KIMI. Their combined median score reached 37.7, exceeding the OpenAI group in the current sample. Across both seasons, their results were also not statistically distinguishable from OpenAI, suggesting that freely accessible or open models may already be competitive with leading commercial systems in AI-assisted molecular discovery.
Gemini-family submissions currently show a median score of 17.1, the lowest among the major groups. However, only 12 Gemini-based submissions were available in the current sample, making it too early to draw strong conclusions.
Perhaps the most important finding is that the same AI model can produce dramatically different outcomes depending on how it is used.
Scores from submissions using identical or closely related models ranged from single digits to above 78 points. This suggests that model selection alone may matter less than the full research workflow surrounding it — including prompt design, chemical tools, search strategies, external databases, reasoning methods and agent architecture.
The real competition, therefore, may not simply be Claude vs. OpenAI vs. DeepSeek, but rather a combination of the AI model, prompt design, tools and scientific strategy.
Both Season 1: Malaria and Season 2: Tuberculosis are currently open for participation. VIDRAFT provides a practical AI drug-discovery guide so that even participants without a traditional pharmaceutical research background can use modern AI systems to design and submit candidate molecules.
The choice of malaria and tuberculosis reflects a broader public-interest mission.
The World Health Organization has long highlighted a structural problem in global health: diseases concentrated among low-income and vulnerable populations can carry enormous medical burdens while attracting comparatively weaker commercial incentives for pharmaceutical R&D.
In simple terms, the medicines humanity needs most are not always the medicines the market rewards most.
That mismatch creates a persistent gap between public-health need and commercial drug-development investment.
VIDRAFT designed the Open Discovery Challenge as an open-science response to that gap.
By lowering the barrier to entry and enabling anyone with access to modern AI to participate in molecular discovery, the project aims to mobilize the collective intelligence of researchers, developers, students, independent scientists and ordinary AI users toward diseases that may otherwise receive less attention from traditional commercial pipelines.
At the same time, the challenge represents a broader shift in how AI itself may be evaluated.
Until now, systems such as OpenAI, Claude, Gemini and DeepSeek have largely been compared through benchmarks in reasoning, coding, mathematics and knowledge.
The Open Discovery Challenge pushes that competition into a different arena:
Can AI help discover something new, useful and scientifically valuable in the real world?
VIDRAFT CEO MINSIK KIM said: “The next phase of AI competition will not be about answering a few more benchmark questions correctly. It will be about whether AI can discover new molecules, materials and scientific solutions that humanity actually needs. Our goal is to develop the Open Discovery Challenge into a global open-science platform where anyone can bring their own AI and participate in real scientific discovery.”
About VIDRAFT
VIDRAFT Inc. is a Seoul-based deep-tech company founded in March 2024, developing Pre-AGI artificial intelligence models and quantum computing technology in-house. Its Darwin foundation models, the open-source AETHER, the on-device POCKET family and the AX-RAY safety diagnostics system are published openly on Hugging Face, and the company is registered in Korea for research and development in physics, natural sciences and life sciences. VIDRAFT works on the principle that research comes first and products follow from its results, aiming to make artificial intelligence a working instrument for real problems in drug discovery, new materials and fundamental science.
Open Discovery Challenge
https://huggingface.co/spaces/FINAL-Bench/open-discovery-challenge
VIDRAFT
https://vidraft.net
Media ContactCompany Name: VIDRAFTContact Person: Minsik KIMEmail: Send EmailCountry: South KoreaWebsite: http://www.vidraft.net
