On the afternoon of September 21, XtalPi announced a major upgrade to its AI-driven scientific agent platform, XtalPi Science, which now features over 80 proprietary scientific skills and tools, alongside an expanded invitation-based testing phase. In internal evaluations, the platform demonstrated superior performance to Claude across multiple drug discovery tasks, and has been integrated with autonomous laboratories, enabling it to direct robotic systems in executing experiments, collecting data, and advancing subsequent workflows. Since the start of 2026, scientific research has emerged as a key frontier for AI agent deployment, following the earlier boom in coding—Anthropic has launched Claude Science, DeepMind is advancing its AI co-scientist initiative, and both are exploring robotic execution of wet-lab experiments, underscoring that XtalPi's combined AI and robotics strategy aligns with broader industry momentum.
The newly introduced suite of 80-plus capabilities encapsulates proprietary data, algorithms, and methodologies into deployable tools, scientific skills, and reusable workflows. These cover critical stages including molecular generation and property prediction, molecular docking, retrosynthetic route planning, structure-activity relationship (SAR) analysis, and LC-MS spectral interpretation, with users able to trigger any feature by simply typing a slash command. In internal benchmarking, blind expert reviews of the hit-to-lead DMTA workflow showed that Claude achieved a recommendation validity rate of 70%, whereas XtalPi Science reached 90%. Similarly, in assessments of molecule quality, approximately 20% of Claude's recommendations were rated excellent, compared to roughly 60% for XtalPi Science; the platform also produced superior results in synthetic route planning tests.
These capabilities are already being applied to internal molecular library synthesis projects, operating in tandem with autonomous labs. This integration has yielded a more than 40-fold improvement in data collection efficiency and an 80% reduction in expert analysis and decision-making workload. With end-to-end management, a 15-person team now delivers over 20,000 reaction data points and tangible molecular outputs each month.
Comments