Google’s latest AI model, currently in testing on its AI Studio platform, shows remarkable advancements in automated handwriting recognition and reasoning capabilities. Users are experiencing a new A/B testing feature where they select the better outcome between two results. Speculation suggests this could be the forthcoming Gemini-3 model. Initial tests on handwritten documents have produced near-perfect transcriptions, comparable to expert human performance. The model also exhibits unexpected reasoning skills, completing tasks that seem to require a deeper understanding of context. The author, a historian, emphasizes the challenges of transcribing historical handwriting, which goes beyond mere visual recognition. Understanding historical texts requires contextual knowledge of language, measurement systems, and cultural nuances from different time periods. This insight is critical since many historical documents contain information that is difficult to interpret without a strong grasp of the era's context. The interplay of vision and reasoning makes handwriting recognition a complex task for AI, raising questions about whether models can achieve human-level reasoning or if they'll plateau at a lower performance level. The article highlights the limitations of current AI models, particularly in their predictive nature. They excel in predicting common sequences but struggle with less frequent, unpredictable elements like obscure names or specific dates. The author points out that while recent models have improved accuracy rates, the final 10% remains elusive. This gap in accuracy is essential for historians needing precise information. To evaluate these models, the author and a colleague created a benchmark of 50 historical documents—totaling about 10,000 words—representing various writing styles and conditions. Their findings could inform broader discussions about the potential of AI in historical research and other fields requiring a combination of visual precision and reasoning.