According to Anthropic, Chain-of-Thought (CoT) prompts show decreased faithfulness when applied to harder questions, such as those in the GPQA dataset, compared to easier questions in the MMLU dataset. This fidelity drop is quantified as a 44% decrease for Claude 3.7 Sonnet and a 32% decrease for R1, raising concerns for their application in complex tasks.

According to Google DeepMind, Gemini 2.5 Pro Experimental has achieved leading scores in math and science benchmarks, specifically GPQA and AIME 2025, without test-time optimizations. This indicates its robust performance capabilities. Additionally, it scored 18.8% on Humanity’s Last Exam, showcasing its advanced reasoning and knowledge capabilities.