A study evaluated five large language models (LLMs)—including Gemini 2.5 Pro, DeepSeek R1, and ChatGPT 5—for adherence to clinical guidelines for anaplastic thyroid cancer (ATC), a rare and aggressive malignancy. Seventy clinical questions spanning three complexity levels (simple, moderate, complex) were developed from three international guideline sets: ATA, NCCN, and ESMO. The study addressed three gaps: first evaluation of 2024–2025 models, first multi-guideline assessment using ATA, NCCN, and ESMO concurrently, and first LLM evaluation for a rare, aggressive malignancy where guideline adherence determines survival. The models were compared against these guidelines to clarify whether they can reliably support expert judgment, with intended use as supportive decision tools for healthcare professionals, particularly outside high volume centers, and not as autonomous agents. The study followed TRIPOD-LLM guidelines.