OpenAI releases MedChemBench to test AI in drug discovery
The new benchmark reveals that even top-tier AI models struggle to pass half of the tasks required for medicinal chemistry.
The short version
- MedChemBench measures tasks like predicting toxicity, interpreting chemical structures, and mapping synthesis routes.
- OpenAI's GPT-6 Astra currently holds the top leaderboard spot with a score of 49.7%.
- The benchmark is primarily used as an internal yardstick for OpenAI to refine its own models.
- The research highlights a significant 'ceiling' where models struggle with consistent multi-step chemical reasoning.
Why it matters
This benchmark provides a realistic look at the limitations of AI in high-stakes scientific fields. It shows that while AI can assist in chemistry, it still lacks the complex reasoning needed to replace human experts in drug development.
A new yardstick for chemistry
OpenAI has introduced MedChemBench, a specialized tool designed to measure how effectively artificial intelligence can perform medicinal chemistry. This field involves the specific process of designing and refining molecules that have the potential to become new medicines. By creating this benchmark, OpenAI aims to quantify the capabilities of current models in a laboratory context.
The benchmark is structured to mirror the actual daily tasks that a human chemist performs in a professional setting. It does not just look at simple facts but evaluates the model's ability to handle complex workflows. This includes interpreting various chemical structures and predicting how potent or toxic a specific compound might be during the development phase.
As reported by Crypto Briefing, the benchmark serves as a way to see where AI succeeds and where it fails in the drug discovery pipeline. It provides a standardized way to compare different models as they are developed to handle more sophisticated scientific data. The goal is to identify the specific hurdles that prevent AI from becoming fully autonomous in chemistry.
Current performance limits
The initial results from the MedChemBench leaderboard show that AI models are still finding the work quite difficult. Even the highest-performing models are currently landing below the 50% mark on the benchmark. This suggests that there is still a significant gap between current AI capabilities and the requirements of professional medicinal chemistry.
OpenAI's own models currently occupy the top spots on the leaderboard. GPT-6 Astra achieved a score of 49.7%, while the GPT-5.6 Sol model followed closely with a score of 48.3%. These numbers indicate that while the models are improving, they are not yet capable of handling the majority of tasks accurately.
The research identifies a specific reason for these low scores: the difficulty of consistent multi-step chemical reasoning. Chemistry often requires a sequence of logical steps where each result depends on the previous one. AI models currently struggle to maintain this chain of logic over long, complex sequences without making errors.
Internal tools and broader context
MedChemBench is described by OpenAI as an internal benchmark. This means that its primary purpose is to serve as a yardstick for the company to evaluate and refine its own models over time. While the leaderboard is visible, the primary use case is for internal development and iterative testing of their proprietary technology.
This benchmark is part of a larger family of evaluation tools created by OpenAI to test AI in the life sciences. For example, LifeSciBench was launched around June 2026 and contains 750 expert-authored tasks across seven different workflows. Additionally, the company has developed GeneBench to extend its evaluation into other areas of biological science.
These tools allow OpenAI to see how specific models perform on specialized tasks compared to their base versions. For instance, a model called GPT-Rosalind showed notable improvements on drug-discovery metrics when tested against base GPT-5.5 models after being evaluated on the MedChemBench framework.
Real-world chemistry applications
While the benchmark highlights current limitations, there is evidence of AI being used in practical chemistry experiments. In June 2026, a collaboration with Molecule.one demonstrated how these models can be applied to real-world problems. In that specific project, the GPT-5.4 model was deployed to assist in running 10,080 different reactions.
The goal of that specific deployment was to improve yields in a Chan-Lam coupling reaction. This type of work shows that AI can be a useful assistant for high-volume, repetitive tasks in a lab environment. It helps researchers process large amounts of data to find the most efficient path for a chemical reaction.
However, the difference between assisting with a single reaction and mastering the entire medicinal chemistry workflow is significant. The MedChemBench results suggest that while AI can help with specific parts of the process, it is not yet ready to lead the entire drug discovery journey independently. The focus remains on how these models can best support human scientists.
The path toward 50%
The fact that the top models are hovering just below 50% creates a clear target for future AI development. For the industry to see a major shift in how drugs are discovered, models will need to break through this reasoning ceiling. This will require improvements in how AI handles long-form logic and complex, multi-step instructions in a scientific context.
Outside researchers are likely to take a close interest in how these tasks are constructed and scored. Because OpenAI's own models are currently the leaders, the community will want to understand the specific criteria used to judge chemical competence. This transparency will be important for other labs trying to develop their own chemistry-focused AI tools.
Whether future models can meaningfully pass the 50% mark will be a key indicator of AI's role in the future of medicine. For now, MedChemBench serves as a sobering reminder of the complexity of chemistry. It highlights that while AI is a powerful tool, the nuances of molecular design still require a high level of sophisticated reasoning.
What to watch
- Whether future models can break past the 50% mark on the MedChemBench leaderboard.
- How OpenAI plans to improve multi-step chemical reasoning in its next generation of models.
- How other companies will develop their own benchmarks to compete with OpenAI's internal standards.
Try it
Look up the names of the three benchmarks mentioned (MedChemBench, LifeSciBench, and GeneBench) to see how OpenAI is categorizing different types of scientific AI testing.
Sources and links
Read the original at Crypto Briefing
First reported by Crypto Briefing. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.