I thought Gemini was supposed to be the long context king?

A viral benchmark result is challenging the narrative around long-context AI models. Google's Gemini 1.5 Pro and the newer 3.1 Pro have been heavily marketed for their massive, industry-leading 1 million token context windows, promising the ability to process entire codebases or lengthy documents. However, the recently released Multi-Context Reasoning Check (MRCR v2) benchmark reveals a dramatic performance cliff. When tested at its full 1M token capacity, Gemini 3.1 Pro's accuracy on complex, multi-step reasoning tasks plummets to 25.9%, down from a robust 71.9% at a 128K token context. This represents a staggering 64% drop in performance, suggesting the model struggles to maintain coherence and retrieve relevant information across its entire advertised context.

In direct comparison, Anthropic's Claude 3 Opus model demonstrates significantly more robust long-context performance. On the same MRCR v2 benchmark, Claude Opus maintained a high accuracy of 78.3% regardless of context length, showing no such degradation. This stark contrast underscores a crucial technical distinction in the AI field: simply having a large context window is not the same as having effective "needle-in-a-haystack" retrieval capabilities or robust in-context learning. For developers and enterprises, this means a model's advertised maximum token count is a poor indicator of its practical utility for real-world, document-intensive tasks like legal review or code analysis, where consistent accuracy is paramount.