Periodic has built an AI model called Neon that's significantly better at analyzing X-ray diffraction (XRD) data than existing frontier models while costing less to run. On their internal benchmark of 134 difficult samples (FrontierXRD), Neon achieved a 55.3% success rate compared to 2.7% from the base model they started with—a 20x jump. The model outperforms GPT-6 Astra and Claude Fable 5.1 on this task, which matters because XRD analysis is central to their materials discovery work. They trained Neon using only 1,300 H200 GPUs, vastly fewer than the 100,000+ Blackwell GPUs reportedly used for Astra, yet still got better performance on their specific problem. The challenge with XRD is that it's genuinely hard, not just computationally. When scientists synthesize new materials, the result is often a mixture of intended phases, unreacted precursors, and unexpected byproducts. X-rays scatter off crystalline structures in characteristic patterns, but overlapping patterns from multiple phases make identification difficult. Experts spend hours reasoning through which phases are chemically plausible given synthesis conditions, temperature, atmosphere, prior experiments, and thermodynamic data. They're not just pattern-matching; they're applying domain knowledge to rule out nonsensical interpretations. Periodic's evaluation focused specifically on these hard multiphase cases where samples contained five phases on average. Neon's edge came from three things working together: training on their own lab data through reinforcement learning, a scientific harness that integrates internal databases and materials simulation tools (which alone gave 3.8x better performance than Claude Code with open-source tools), and an LLM-judge system trained on expert scientist feedback to evaluate whether proposed phases actually make sense. The model generalized well beyond its training data—it outperformed frontier models on held-out chemical systems it had never seen, suggesting it learned transferable reasoning rather than memorizing. Periodic's point is clear: they're scaling autonomous labs and the AI systems that learn from them in tandem, and with more training compute at levels comparable to frontier models, they expect even greater scientific capabilities.